We think that semiconductor layer is going to capture a very, very significant amount of the economic rent over this next 10 years. —Andrew Homan

I’m convinced that 50% of your inference… will be on device. —Josh Wolfe

On October 15, Robin Hood, New York City’s largest local poverty-fighting philanthropy, hosted its 13th annual J.P. Morgan / Robin Hood Investors Conference. I had the good fortune of working with a couple of the founders of Robin Hood at the hedge fund I was at, one for quite a long time. It’s hard to overstate the positive impact the organization has had on those in need. And this hasn’t been limited to New York City. Their style of philanthropy, the principles they espouse, and the best practices they’ve developed have helped shape other organizations and efforts globally. They are the best of us.

Learn more about Robin Hood

Of note this year was the AI investor panel exploring the convergence of artificial intelligence and investment opportunity, featuring Brad Gerstner (Altimeter), Andrew Homan (Maverick Silicon), and Josh Wolfe (Lux Capital), and moderated by Chase Coleman (Tiger Global).

In a roughly thirty-minute session, these esteemed investors shared their insights on how they are approaching AI as an asset class. The conversation covered the capital cycle driving AI’s buildout, the layers of value creation across chips, models, and applications, and the emerging winners as compute, memory, and data converge into the next great investment opportunities.

To date, AInvestor has focused largely on the process of investing. With this note, we begin to explore where top investors look to capture the value created from building, running, supporting, and using AI. This is what I refer to as “AI as an asset class.”

The rest of the note is structured as follows:

• Learnings

• Transcript

• Frequently Asked Questions (FAQs)

• Mind Map

Read or skim at your pleasure. Of course, I hope you’ll take in everything below, but I also encourage you to click through and watch the session in full. You’re in for a treat.

Robinhood Investor Roundtable

Learnings

This has been a fun note to work on. Though I don’t personally know any of the folks on the panel, I’ve come to trust my ability to recognize those who know their craft. There is only so much that can be fit into thirty minutes of banter and Q&A, but there was enough here to get my juices flowing.

The core insights that jump out are:

AI Layer Cake

I’ve had the good fortune to work with some iconic investors, and others that I’d say are equally skilled. While assets and approaches have differed wildly, each has had a framing that brought structure and definition to the otherwise infinite set of options facing investors.

In introducing the AI Layer Cake, Andrew Homan acknowledges different investors have their own definitions and recipes. He thinks about it here as: chips → models → applications and tools.

Adding a couple of layers, my cake looks like:

• Infrastructure: This includes everything that goes into making AI come alive, from power to chips to everything else that makes data centers go hum.

• Models: I consider these all things that take inputs and process them into outputs. They range from the biggest foundational models targeting general intelligence to the smallest, task-specific.

• Applications and Tools: These are the products and services that manifest technology and data into outputs and outcomes. In the best of circumstances, this is where value is created.

• Beneficiaries: These are companies that capture the value AI helps create. In the early days, much of this appears to be coming in the form of cost savings and efficiency; increasingly, we should see AI play more of a role in expanding revenue opportunities. Scott Galloway, of Prof G fame, presents Amazon as Exhibit 1A as the company further automates its fulfillment and distribution efforts. You can read Scott’s arguments here.

Benefits of Scale Overcome Innovator’s Dilemma

Common wisdom holds that incumbents inevitably fail to adapt for two reasons. First, because innovating would cannibalize existing profit pools they have so diligently worked to build and protect. Second, success leads to size, which leads to bureaucracy, which in turn suffocates the ability to react.

As Brad Gerstner put it:

“We learned in business school that elephants can’t dance—that technology companies get big and get disintermediated. What’s happened over the last 15 years is the exact opposite. Scale has led to bigger advantages.”

The modern AI stack, however, rewards scale rather than punishes it. And beyond access to data, compute, talent, and distribution, an often-overlooked superpower is the ability of Google, Microsoft, Amazon, and Meta to be their own first-and-best customers. The immediacy and intensity of the internal feedback accelerates development while overpowering the traditional sources of friction that usually come with size.

User Interfaces Get Personal

The pathways and mechanisms by which we engage with AI are going to optimize for removing friction that impedes getting the necessary data and insights into and out of these systems.

In addition to the fancy Ray-Bans, Josh Wolfe highlights Meta’s push into neural interfaces, what he calls “life-cording, where every part of your life is going to be recorded 24/7.”

I’ll admit my mind went straight to a Black Mirror episode (c. 2011, if you can believe it), which had a dystopian gloss to it, but I think he is on point. Increasingly, the limiting factor to AI being useful is not just more data, but the right data, where “right” means that which is most relevant to you. In other words, your data.

Compute Gets Personal

This follows naturally from the ability to collect massive quantities of personal (and sensitive) data. Perhaps not always true, but generally, it is more expensive and less secure to cycle data from client to server and back again.

Inference will increasingly be driven to the supercomputers commonly known as our laptops and cellphones. Josh Wolfe puts the number at roughly half of our AI use being on-device.

Memory Is Personal

I’m taking editorial prerogative in slipping this in here as a bridge to some of the tangible investment ideas offered by the panel. If data is to be collected and processed on-device, it should be stored there, too. And when I say data, I include the questions, context, and output generated. That will be a lot of data.

Memory Gets Rerated

Few things make an investor salivate more than redefining what a company does or how it should be valued. Unlike compute, memory is a fairly commoditized product with commensurately low margins. Historically, companies in this sector ride the tides of common economic cycles, interrupted by spasms of boom and busts triggered by major hardware or platform transitions (see: PCs, cell phones, and the cloud).

As outlined, AI fits the bill for kicking off a new boom cycle for memory. The question becomes whether or not this represents a new paradigm for valuing memory companies. Both Homan and Wolfe would suggest so.

Andrew Homan,

“Memory is the one piece of the bill of materials that continues to grow each generation.”

Josh Wolfe,

[…] I do believe that the memory players have that same dynamic that, 11 years ago, people were like, ‘Oh, GPUs are just tightly coupled to the gaming PS5 console wars, Xbox, etc.’ And these are commodity memory chips, but I think they’re going to be ascendant, and I think it’s going to crack the narrative that we need to invest these tens of billions of dollars into data centers.

Investment read-through

Bonus advice

It’s too early for quantum.

J.P. Morgan Robinhood Investors Conference

Wednesday, October 15, 2025

This conversation has been edited for clarity and length.

Chase Coleman: Thank you. It’s a pleasure to be here with all of you today, supporting Robinhood and the incredible work they do to combat poverty in New York City. I’m looking forward to a good discussion between three incredibly thoughtful investors on a topic we all care about: AI.

To frame it, AI capital expenditures are contributing roughly 1% to U.S. GDP growth—about half of total GDP growth. ChatGPT is the fastest-growing consumer application of all time, expanding three times faster than Meta, Google, or AWS at similar scale. Anthropic may have added more than half of all new annual recurring revenue in the SaaS industry during the first half of this year.

We’re seeing multi-tens and hundreds of billions in infrastructure investments almost weekly. We’re at the outset of reasoning models, video models like Sora and Google’s Nano Banana and VO3, and intelligent agents that should drive sustained token growth. As we like to say in this industry—it’s on.

Brad, let’s start with you. This is Q4 2025, and everyone’s talking about AI infrastructure now. Help us extend the horizon: what applications and buildout trends do you expect over the next several years?

The Scale of AI Investment

Brad Gerstner: It’s great to be here—and for such a good cause. As I said to Andrew Ross Sorkin, who was hyperventilating about bubbles this morning on CNBC, rarely do you see bubbles when everyone is already talking about them and when the hit book of the day is about the 1929 crash. NVIDIA’s trading at 27 times earnings—it’s not frothy.

Jensen Huang told me last week that we’ll see $3–4 trillion in compute buildout over the next four to five years. That’s roughly ten times the Manhattan Project—$4 billion then, about $400 billion GDP-adjusted today. We’re doing $4 trillion, all privately funded. That’s a massive economic tailwind.

Why such scale? It’s not just generative AI. Every business is moving from general-purpose compute to accelerated compute. Without that shift, there’s no TikTok, Reels, or instant Google answers. I asked Jensen about the risk of a glut. He said, “No chance in the next two to three years.” The hyperscalers are driving this. Sam Altman’s flurry of deals? They’re frameworks for the next decade, not obligations.

We’re in inning one or two of a decade-plus buildout that will transform computing, business, and daily life. There’ll be volatility—NVIDIA’s had two 25% drawdowns this year—but the direction is clear.

The AI Layer Cake

Chase Coleman: Andrew, you’ve talked about the “AI layer cake.” Walk us through that.

Andrew Homan: The AI layer cake has three layers. Bottom: semiconductors. Middle: large language models like OpenAI and Anthropic. Top: applications and tools.

We believe the semiconductor layer will capture a very significant share of the economic rent over the next decade. That’s a major shift from the 2010–2021 era when software investing rode two tailwinds—the move to cloud and the rise of SaaS. Delivering solutions now requires far more compute. That’s why we’re most excited about chips.

Five-Year Psychological Bias

Chase Coleman: Josh, you’ve written about the “five-year psychological bias.” Explain what that means and how it applies to AI.

Josh Wolfe: First off, Brad’s the best-dressed guy on stage. Normally he’s in a black T-shirt like me—so his suit is a contra indicator. If he’s in a tux next year, we’ve hit the top.

The five-year bias means everyone wants to be invested today where they should have been five years ago. Our job as venture capitalists is to anticipate where people will want to be in three to five years—betting on people and technology.

I’ll give you a parallel. In 2015, I invested in Zoox, a self-driving company later acquired by Amazon. Early on, engineers were running thousands of simulations on chips not yet released—NVIDIA chips. Back then, NVIDIA was worth $15 billion, Intel $150 billion. I called it the pair trade of the century. Today NVIDIA’s worth $4.5 trillion.

I feel the same way now about memory players. Consensus says data centers will scale forever to benefit NVIDIA and AMD. I think 50% of inference will be on-device. A paper a year ago showed you can run large language models on-device using Flash, NAND, and other memory technologies.

The winners: SK Hynix, Samsung, and Micron. SK Hynix is the most nimble, with 60% high-bandwidth memory market share. Samsung’s slower; Micron faces U.S. export limits. Memory chips are going to be ascendant—challenging the assumption that we must spend tens of billions on new data centers.

CapEx booms drive innovation, but they can also create overbuild risk. Two-dimensional AI—voice, video, text, code—is saturated. The next wave is three-dimensional AI: biology and robotics. Those fields lack dense data repositories, so their data is scarce and valuable.

Lastly, all the big model players are overspending. Each wants 100% of user share but has about 20%, spending 500% on CapEx.

Lightning Round

Chase Coleman: Let’s do a lightning round. OpenAI or Anthropic?

Brad Gerstner: OpenAI.

Andrew Homan: Same.

Josh Wolfe: OpenAI. It’s worth $500 billion now and could hit $2 trillion. Anthropic likely gets acquired by Amazon.

Chase Coleman: Google or Meta?

Brad Gerstner: Meta. Google’s profit pool—those ten blue links—is fundamentally at odds with AI. None of our kids use that anymore. I still own Google, but over five to seven years, Meta’s founder-led risk-taking gives it the edge.

Andrew Homan: Agreed—Meta.

Josh Wolfe: Google will win in AI; Meta will win on devices. We sold Control Labs to Meta—it made neural wristbands that read muscle signals to replace keyboards and remotes. Zuck wants to reinvent human-computer interaction to route around iOS and Android.

I call it “life-cording”—recording your life 24/7. Younger generations will accept it; older ones will be horrified. But it will unlock enormous AI insights. Meta wins that race.

Chase Coleman: NVIDIA or Broadcom?

Brad Gerstner: NVIDIA.

Andrew Homan: NVIDIA.

Josh Wolfe: Broadcom—for TPUs and Google.

Chase Coleman: Microsoft, Oracle, or Amazon?

Brad Gerstner: Microsoft.

Andrew Homan: Microsoft.

Josh Wolfe: Amazon.

Chase Coleman: Tesla or Waymo?

Brad Gerstner: Tesla.

Andrew Homan: Tesla.

Josh Wolfe: Waymo.

Favorite AI Beneficiaries

Chase Coleman: Name a favorite AI beneficiary—public or private.

Andrew Homan: SK Hynix. Every new NVIDIA or Google chip generation uses more memory. SK Hynix dominates high-bandwidth memory and faces long-term supply limits, keeping pricing strong. Historically, memory cycles were brutal, but this one looks different. Book value’s growing fast—up 50% by 2026—so downside risk is shrinking.

Brad Gerstner: Everyone hunts for the undiscovered idea, but sometimes the best one is the obvious one. In 2006 it was Google; now it’s NVIDIA and OpenAI. The Magnificent Seven are up 10× while the S&P ex-Mag Seven is up 1.5×. Scale has become an advantage, especially in AI where compute and data scale compound power.

Josh Wolfe: I’m long Google—it’ll “Microsoft” the foundation model market by dropping prices and giving access away. I’m also long SK Hynix and Anduril, an AI defense company likely to IPO at $50 billion.

Short? The quantum and modular nuclear reactor baskets. Quantum firms trade at absurd valuations with minimal revenue. Nuclear’s future is large-scale fission, not modular startups.

Andrew Homan: I’m less bearish on quantum. Retail loves those stocks—about $100 billion in aggregate market cap. NVIDIA’s taking quantum more seriously, so it’s worth watching.

Winners and Losers in the AI Shift

Brad Gerstner: I’m with Josh—it’s tough to short meme stocks when rates are falling. But look at how value creation shifted. Google was the toll keeper of the internet, sending users elsewhere and collecting a fee. ChatGPT’s model is the opposite: keep users inside and give them answers directly.

That’s a major problem for intermediaries—Expedia, Booking, anyone collecting a 15–20% toll. Agent commerce, like Walmart’s new assistant, will push consumers toward direct brand relationships. Unless legacy platforms make themselves the default assistant channel, their traffic will fall.

Andrew Homan: Intelligence is moving to the edge. Devices are getting smarter. Qualcomm’s okay but may need acquisitions to keep pace.

Even in advanced tech firms, something as simple as getting a presentation on screen can take ten minutes. That’ll become instantaneous as intelligence moves into every device—phones, PCs, robots, autonomous systems. We’re just in the first inning.

Applications and Agents

Chase Coleman: Brad, what excites you most on the application layer?

Brad Gerstner: Everything’s getting rebuilt. The biggest app in history already exists—ChatGPT. Every internet company that used to do website optimization now asks, “How do I build an agent that interacts with the super-agent?” We’re entering a world of agent networks acting on our behalf.

Capitalism and Inclusion in the AI Age

Chase Coleman: You’ve also worked on expanding access to capitalism through Invest America. Tell us about that.

Brad Gerstner: Capitalism fails if 60% of people don’t have compounding assets. The Invest America Act, which passed this year, creates “Trump accounts”—prosperity accounts for every child at birth, seeded with $1,000 in the S&P 500.

Sixty-five million children qualify now. Starting in 2026, every child born in the U.S. will receive one. Companies and parents will contribute too. Sign-ups begin in December. It’s how we get everyone on Team America, especially as AI disruption accelerates. If people feel ownership in the system, capitalism will endure.

Chase Coleman: We could keep going, but our time is up. Thanks to our panelists and to Robin Hood for bringing us together.

Robinhood Investor Roundtable

FAQs

The following summarizes the discussions and insights regarding the AI investment landscape, drawing from the J.P. Morgan Robinhood Investors Conference Transcript featuring Brad Gerstner (Altimeter Capital), Andrew Homan (Maverick Silicon), and Josh Wolfe (Lux Capital), moderated by Chase Coleman (Tiger Global). Note that I have used Google’s NotebookLM to generate this.

Market Growth and Sentiment

Q1: What is the current scale and growth rate of the AI industry?

The growth metrics are immense:

• AI CapEx is contributing approximately 1% to US GDP growth, which represents half of US GDP growth.

• ChatGPT is the fastest-growing consumer application of all time, growing three times the rate of Meta, Google, and AWS when they were at a similar scale.

• Anthropic added more than half of the net new Annual Recurring Revenue (ARR) of the SaaS industry in the first half of this year.

• The industry is currently seeing multi-tens and hundreds of billions of dollar infrastructure investments on an almost weekly basis.

Q2: Are we currently experiencing an AI bubble?

Panelists expressed skepticism that the market is currently in a bubble. Brad Gerstner noted that rarely do bubbles occur when everyone is talking about them, or when the hit book of the day is about the stock market crash of ‘29. He also pointed out that NVIDIA is trading at 27 times earnings.

Q3: What is the long-term outlook for the necessary compute buildout?

The expected scale of the compute buildout is massive and long-term:

• Jen-Sen Huang projects $3 to $4 trillion of compute buildout over the next four to five years.

• To contextualize this, this buildout is roughly 10 times the Manhattan Project (which was $4 billion, inflation adjusted to $40 billion, or GDP adjusted to about $400 billion). This $4 trillion is being funded privately and is a massive tailwind for the economy.

• This buildout is required because all businesses in America are moving from general purpose compute to accelerated compute.

• The industry is believed to be in “inning one or two of a decade-plus long buildout“.

Q4: Is there a risk of a “glut” in compute capacity soon?

Jensen Huang indicated that there is “no chance in the next two to three years” of a glut. This is because in the near term, only the hyper-scalers are paying for this massive investment.

Investment Strategy and Value Capture

Q5: How is the AI industry structured, and which part captures the most value?

Andrew Homan describes the industry using the AI layer cake analogy, which consists of three layers:

1. Bottom layer: Chips (semiconductors).

2. Middle layer: Large Language Models (LLMs), such as OpenAI and Anthropic.

3. Top layer: Applications and Tools.

Panelists believe the semiconductor layer is going to capture a very, very significant amount of the economic rent over the next 10 years. This differs from the playbook that worked from the Global Financial Crisis through 2021, which favored software and the SaaS business model.

Q6: What is the “five-year psychological bias”?

The bias is that “everybody wants to be invested today where they should have been five years ago”. The role of venture capitalists is to anticipate where people will want to be invested in the next three, four, or five years.

Q7: Where is AI inference expected to be conducted, and what companies benefit from this view?

The consensus view is that massive data centers and H100 clusters will handle inference, benefiting hyper-scalers and chip players like NVIDIA and AMD. However, Josh Wolfe is convinced that 50% of inference will be on device.

Companies expected to benefit from this shift to on-device inference are the memory players:

• SK Hynix: Considered the most nimble on the Korean side, they have 60% of the NAND Flash market share. They are benefiting from attach rates on high bandwidth memory (HBM) for GPUs. SK Hynix dominates HBM, which is critical for AI models.

• Samsung: Considered the “Intel of the space,” but potentially more bureaucratic and slow-moving.

• Micron: Based in the US, but may face restrictions due to US export and geopolitical issues.

Q8: Which sectors of AI are the next investment waves?

The panel suggests that any investment in 2-dimensional AI (text, voice, video, code) is considered “done” due to the high entrance rate of capital. The next wave of investment is in 3-dimensional AI, specifically biology and robotics. These domains are valuable because they lack a dense repository of data to train upon, making data scarce.

Q9: What established scale players still hold the greatest advantage?

Panelists strongly recommend staying invested in scale players, arguing that the idea that “elephants can’t dance” has been disproven over the last 15 years.

• NVIDIA and OpenAI are highlighted for having massive scale advantages.

• If forced to invest in only two stocks, Brad Gerstner would choose NVIDIA and OpenAI, believing ChatGPT is the winner in consumer AI, which will be the biggest market in the next 5 to 10 years.

Quick Picks and Wrong Side of Change

Q10: What companies are favored in the lightning round matchups?

Q11: Which assets are considered on the “wrong side of change”?

Josh Wolfe identifies two baskets that he would personally be short:

1. Quantum companies: Described as “quantum flibbidy-doodled”. These companies often have low revenue ($10M to $50M) but high market caps (around $5B). All promised capabilities (like molecular modeling and unbreakable encryption) can currently be achieved with modern compute and GPUs.

2. Modular nuclear reactor companies: While nuclear power is seen as the future (large-scale fission, rebranded as elemental energy), these modular companies are expected to be “disasters”.

Q12: What business models are threatened by AI agents?

The business model of existing internet companies is under pressure because AI’s objective is the opposite of Google’s original objective.

• Google’s objective was to get the user to ask a question and then charge a fee to send them to sites (like Booking.com or Amazon).

• ChatGPT’s objective is to give the user the answer directly.

• Intermediaries like Expedia and Booking.com that collect a 15% or 20% toll are threatened when consumers rely on a personal assistant agent (”super agent”) to book things without specifying the site. Winners in the existing app ecosystem are suspect.

AI and Societal Impact

Q13: How is AI displacement and wealth disparity being addressed?

Brad Gerstner noted that in the age of AI, there will undoubtedly be dislocation and displacement. To overcome these challenges and prevent the country from moving toward socialism, the goal is to double down on capitalism by getting everyone into the system.

The Invest America Act was passed into law this year to address this:

• Starting in 2026, every child born in America will receive a prosperity account at birth.

• This is a 401k from birth, initially seeded with $1,000 in the S&P 500.

• This aims to ensure that 60% of people who currently lack compounding assets feel like they are “on Team America”.

Analogy: The current AI investment phase is like a massive gold rush, but instead of focusing on the surface-level gold (the applications), the most confident investors are betting heavily on the companies that sell the specialized mining equipment (the semiconductors and memory) that everyone needs, regardless of which application wins the race.

Robinhood Investor Roundtable

Mind Map

I. AI Investor Panel: Overview and Context

A. Central Theme

• Exploration of how artificial intelligence reshapes the investment landscape.

• Focus on infrastructure, scale, value capture, and emerging frontiers like memory, robotics, and human–machine interfaces.

II. Macro Context: Scale and Capital Intensity

A. AI CapEx as Economic Engine

• AI capital spending adds roughly 1% to U.S. GDP growth.

• Private sector driving a $3–4 trillion compute buildout over 4–5 years—about 10× the Manhattan Project.

B. Inversion of the Innovator’s Dilemma

• Scale is now an advantage, not a liability.

• “Elephants can dance”: large incumbents like NVIDIA, Microsoft, and Meta leverage capital, data, and feedback loops to accelerate innovation.

• Immediate feedback from massive user bases turns corporate size into a compounding edge.

III. The AI Layer Cake Framework

A. Layer 1 – Semiconductors (Infrastructure)

• Foundation of value creation; hardware dominates economic rent.

• Companies: NVIDIA, AMD, Broadcom.

B. Layer 2 – Models (Foundation Models)

• LLMs such as OpenAI, Anthropic, Gemini form the middle tier.

• Competition split: OpenAI (consumer) vs. Anthropic (enterprise).

C. Layer 3 – Applications and Agents

• Transition from static apps to dynamic agent ecosystems.

• “Everything will be rebuilt” around interaction with super-agents (Brad Gerstner).

D. Layer 4 – Value Capture and Beneficiaries

• Need to map technical bottlenecks (compute, memory, data) to business models.

• Structural advantage shifts toward those controlling scarce inputs or distribution.

IV. The Memory Frontier

A. Structural Rerating Argument

• Memory’s share of system cost rises each generation.

• High-bandwidth memory (HBM) supply constrained; pricing power increasing.

• “Your downside is actually going down as we move into the future.” — Andrew Homan

B. On-Device Inference Thesis

• “I’m convinced that 50% of your inference will be on device.” — Josh Wolfe

• On-device AI reduces dependence on data centers; flash and NAND become strategic.

• Key beneficiaries: SK Hynix, Samsung, Micron.

C. Investment Implication

• Memory transitions from cyclical to structural profit pool.

• Parallel demand curves: cloud training + edge inference.

V. 3D AI and Robotics

A. Next Frontier Beyond Text and Images

• “Anything in two-dimensional AI is done… The next wave is three-dimensional AI: biology and robotics.” — Josh Wolfe

• Data scarcity in physical domains (movement, manipulation, sensing) makes these markets defensible.

B. Early-Stage Opportunity

• Robotics, bioengineering, and synthetic data platforms will define the next investment cycle.

VI. Human–Machine Interfaces and Meta’s Bet

A. From Screens to Sensing

• “Typing is so yesterday.” Interaction shifting from keyboards to neural inputs and gestures.

• Focus on removing friction in how data enters and exits AI systems.

B. Wolfe on Meta’s Vision

• “Zuck doesn’t want to be remembered for reinventing advertising… but for reinventing hardware and the human-computer interaction.”

• Meta’s Control Labs acquisition → wrist-worn neural bands enabling “life-cording”—continuous, passive data capture.

• Goal: bypass iOS and Android; create new device platforms for personalized AI.

VII. Agent Commerce and Platform Disruption

A. The End of Search Toll Collectors

• Brad Gerstner: “Google’s objective was to send you away. ChatGPT’s objective is to keep you inside and give you the answer.”

• Transition from search-driven traffic to agent-executed transactions.

B. Consequences

• Pressure on intermediaries (Expedia, Booking).

• Direct brand-to-agent commerce reshapes digital distribution.

VIII. Energy, Quantum, and Overhyped Narratives

A. Energy as Bottleneck

• AI data center expansion constrained by energy availability.

B. Wolfe’s Skepticism on Quantum and Modular Nuclear

• “I’d be short the basket of all the quantum companies… They make absolutely no sense.”

• Prefers large-scale fission over speculative modular startups.

IX. Defense and Autonomy

A. Anduril as Archetype

• “AI and autonomy is the future of defense… a single operator controlling thousands of systems.” — Josh Wolfe

• Example of AI-native industrial transformation with scalable, software-defined hardware.

X. Closing Themes

A. Common Investment Threads

• Structural over cyclical plays.

• Scale, feedback, and control of scarce inputs define enduring edge.

B. Shared Long Positions

• NVIDIA, OpenAI, SK Hynix, Meta, Microsoft, Google, Anduril.

C. Meta-Insight

• AI is no longer just a software story; it is an infrastructure revolution spanning compute, memory, energy, and human interfaces—the new architecture of value creation.

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis

Help Wanted

A job posting from a large asset manager landed in my inbox last week.

[The firm] is looking for an AI Product Leader to spearhead [our] enterprise-wide AI product strategy. This critical role will own, manage, and prioritize all AI initiatives at the intersection of our business and technology, working across Operations, Finance, Investments, Asset Management, and Other teams. You will be instrumental in streamlining our technological capabilities, building AI roadmaps, driving prioritization, and managing successful rollouts of impactful AI solutions.

This is brilliant. It shows a realization that regardless of the quality of your data and technology, it’s all for naught if your people won’t use it.

I have some experience playing at the intersection noted, and if I have learned anything, it’s that you need cooperation and a common understanding if you want the adoption needed to achieve the outcomes you are shooting for. The challenge is that each side has their own north star, measures of success, and language through which they operate. And this is before you start navigating vertically from the C-suite through management down to the end users. Each of those layers has their own priorities and considerations. It gets complicated.

What makes this role compelling is its capacity to take on a broad mandate and focus 24/7 on design, communication and execution. That capacity to focus is critical. Everyone else has their day jobs—the ones they were hired for; the ones they have been doing successfully, and, highly likely, spending more than 40 hours a week at before being asked to take on this AI thing.

That focus becomes even more critical when you consider what this role actually requires: the ability to see across competing and complementary needs simultaneously, identifying the common abstractions that create leverage and do so without losing sight of the details that make or break adoption. It takes a combination of strategic thinking, tactical execution and a fluency across domains that demands both breadth and depth. Reach for the stars.

Responsibilities & Qualifications

There is a lot to the role, so there is, correspondingly, a lot laid out in these two sections. Eighteen bullet points, in fact. To their credit, they offer a coherent listing of activities and capacities needed to manifest them, providing a roadmap that runs through data science, engineering, behavioral science, and the many facets of the business itself.

A summary from the specs:

As the AI Product Leader, you will be the driving force behind shaping our technical AI roadmap in close partnership with engineering. Your mission will be to ensure the development and deployment of impactful, scalable AI solutions that accelerate productivity and innovation across [the firm]. The ideal candidate will possess a strong foundation in product management and/or engineering, with a proven ability to translate complex business needs into cutting-edge technical solutions. You’ll navigate [the firm’s] diverse business units, identifying opportunities for AI to deliver significant value and championing a user-centric, agile approach to product development.

In layman’s terms, this role will lead the AI roadmap, working hand-in-hand with engineering to turn ideas into real products. The goal is simple: build and roll out AI tools that boost productivity and spark innovation across the firm. To do that, leadership needs both product and technical know-how, plus the ability to turn messy business problems into clear outcomes. They’ll work across all of the firm’s businesses to spot where AI can add the most value and make sure the tools are built with users in mind and improved quickly through feedback.

The core elements that jump out are:

• Impactful: Perhaps this is obvious, but I still give credit for beginning with the end in mind. Understanding markets and harvesting returns have always been my main thing—technology, including AI, is a means to a business end.

• Scalable: GenAI makes prototyping and building functional proof of concepts demonstrably easier. It is easy to be seduced into thinking they can replace existing tools, or be rolled out more broadly. But at this point they are fragile, and when they break the signs can be easy to miss. Or not. Hidden within scale is the issue of security. There is a whole post to be written on this.

• Innovate: The I-word can be a cliche—we had an Innovation Day at my old shop twenty years ago. But what I read here is an understanding that like ideas, solutions need to be pursued with conviction, but held loosely. The competencies of the technology are simply evolving too fast to build otherwise.

• Translate: This may be the secret sauce. As noted, there is a rat’s nest of north stars, objectives and languages that must be navigated and reconciled. People frequently see differences when there is mostly similarity; just as often, there are stark differences masked by common misunderstandings. Being effectively multi-lingual is table stakes.

• Engineering: For all the advancements in data science, enterprise scale AI sits on a foundation of sound engineering fundamentals. Security, transparency, memory, search, agents and tools all need to be deliberately designed and integrated. Engineers do this.

• Product Management: This recognizes the softer skills that raise the likelihood that the people whose problems you are solving will participate in the building, and ultimately, use the AI solutions provided. Product oriented leadership helps overcome the barriers of adoption—inertia, fear, misunderstanding, and the lack of time.

Product Mindset

The tell that this unnamed institution “gets” the critical importance of adoption is their explicit framing of the role as Head of AI Product. Including Product in the job title emphasizes the human dimension. It’s worth noting they are also hiring someone to lead their AI Platform development, where the expectations and requirements are appropriately focused on the technical. This separation shows they understand the difference between building capabilities and driving adoption, and that both require dedicated expertise.

I spent over two decades in the front office of a large hedge fund building and innovating with technology in support of my day job. If you had asked me I would have said I was creating capabilities. The product was our returns. It wasn’t until I was readying myself to run product at a fintech that I came to appreciate that much of the process behind creating those capabilities was rooted in a particular mindset. Namely, working backwards from the outcomes wanted, identifying where we could get the most leverage, ease of use, and collaboration.

Coming at it from the direction of thinking “we can build this and you should want it” doesn’t work. I’ve had to step over too many corpses—some of my own making—of projects that failed to launch, reach escape velocity, or land where hoped. Most were well-intentioned, some were vanity projects, none were adopted.

Getting Started

Some public breadcrumbs are available, but I don’t really know the current state of play at this company. That said, the new AI Product Leader will undoubtedly inherit a complex landscape, but there’s a logical sequence to navigating it successfully.

Understand the Third Rails
In a highly regulated industry such as they’re in, the first job isn’t identifying opportunities but staying out of trouble. Age teaches you that. What simply can’t be done? What data can’t leave? What’s allowed in? Which processes require human sign-off? Where does compliance draw hard lines? What are acceptable tradeoffs? These are design parameters that will shape every solution. Better to know upfront so they can be incorporated in the components with which everything else will be built. Solving the security and regulatory requirements at the beginning will save time, effort and legal fees down the road.

Define and Prioritize Outcomes
Start with the business outcomes that matter most: faster decision-making, reduced operational risk, improved communication, whatever moves the needle for this firm. Identify the common abstractions that when addressed, offer the greatest operational and technical leverage. Then work backwards to identify which AI interventions could realistically impact those outcomes. This isn’t about what’s technically possible, per se, it’s about taking what’s possible to accelerate productivity and innovation across all business units to generate higher returns.

Provide Requirements, Resources, and Cover
Once you’ve identified the highest-leverage opportunities, the job becomes enablement. Internal stakeholders need to provide clear requirements and constraints around their particular use cases. Developer resources need to be secured (data access, compute power, but also time and attention from subject matter experts). Most importantly, provide air cover from organizational politics or when experiments don’t work out. Innovation requires permission to fail fast and learn quickly.

Experiment, Implement, Innovate
While there is experimentation and innovation taking place at every step, I’m referring here to a more organizational and sequential context. Large, successful organizations do not have the luxury of revolutionizing how they operate overnight. Better to start by learning what can and cannot be done on representative but clear use cases. At this point, you have many siloed efforts—many unknown even to management.

Next comes the implementation phase, which is where this company is entering. At this stage, you bring in common infrastructure, standardized tooling, governance frameworks, and user feedback systems needed to scale safely and predictably. Executed well, management and users will be delighted, but still able to recognize the organization they work for.

With time, the company will gain deep competencies and a clearer understanding of where AI can best add and capture value. Most importantly, it will become part of the company’s culture. Add to this the ever-growing capabilities of the technology, and there will be opportunities to reimagine how business is done. Just don’t start with that.

Execute the Work
The day-to-day comes down to signal extraction and communication. Identify what’s working (and what isn’t) from user behavior, not just user feedback. Then communicate those insights up, down, and across the organization in language each audience understands. The C-suite cares about business impact; engineers care about technical feasibility; end users care about their daily workflow. Same data, different stories.

Success in this role won’t be measured by the sophistication of their AI, but rather by how much the organization’s behavior changes for the better and the outcomes that result.

Wrapping Up

If you are reading this it’s safe to assume you already believe that we are currently in an extraordinary time. Just from when I entered the workforce, the technical gains starting with the PC have compounded to a level that is hard to comprehend. And now we have GenAI. It can feel like the growth curve is pointing straight up.

With this progression comes hyperbole. It is easy to get swept up in the emotion, or conversely, become overly skeptical. Both reactions are understandable but neither is productive, certainly in the extreme. Balance needs to be struck across risks, rewards and human nature. Part of that balance comes from recognizing the scope of the technology needed paired with an understanding of the needs and experiences of those who will be using it. This is where the worlds of engineering and product meet. The most progressive organizations recognize this duality, and are acting accordingly.

When wanting to understand what a company is actually doing, rather than just saying, I have found job postings invaluable. They show where the company is spending their money; they reveal where management sees their future, and who and what is needed to get there. I hope you have found this particular walk through revealing. Please reply and comment with your thoughts.

I’ve purposefully avoided including the name of the asset manager to avoid the distraction that comes with iconic names. I can tell you, though, it’s not Tudor. That would be too easy.

Click through for FAQs and Mind Map summaries of the key concepts from article

Custom GPT for Light Edits

Click above to access a custom GPT we spun up to help clean up our own writing. It’s designed to apply the lightest of touches. If you’re a paying ChatGPT customer, it’s free to use (we don’t get a cut). Feel free to play around with it as much as you’d like.

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis.

AI’s real challenge inside large financial institutions isn’t data, models, or even governance—it’s adoption. A recent job posting from a major asset manager reflects this truth perfectly: success now depends on how well firms translate technical capability into products people actually use. The questions below unpack what this shift means for asset managers, and why product thinking—not just AI expertise—is fast becoming the differentiator.

Q1. Why is this job posting worth paying attention to?

Because it signals a major shift in how large financial institutions are thinking about AI. By naming the role Head of AI Product Management, the firm is acknowledging that success depends not only on models or infrastructure, but crticially, on how people adopt, trust, and use the technology.

Q2. What makes “Product” the right framing for AI in this context?

“Product” introduces accountability for usability and outcomes. It bridges engineering and business, ensuring that what’s built solves real problems and fits into how investors, analysts, and operators actually work. It makes adoption a first-class goal rather than an afterthought.

Q3. What’s the main challenge a role like this will face?

Coordination. Different groups—Investments, Operations, Finance, and Other Businesses—each have their own objectives, systems, and languages. Aligning them requires translation, patience, and credibility. The hardest part isn’t building models; it’s building shared understanding.

Q4. How does the article define successful AI adoption?

Adoption isn’t measured by the number of models deployed—it’s measured by behavior change. Do teams make faster, better decisions? Do they rely on the tools without being forced to? Are outcomes measurably improved? Those are the real KPIs.

Q5. Why is governance mentioned so early in the roadmap?

Because in financial services, constraints define design. Knowing the “third rails” of regulation, data privacy, and security upfront ensures that solutions can scale safely and sustainably. Governance done early is an enabler, not a blocker.

Q6. What is meant by “adoption muscle”?

Adoption muscle is the organizational discipline to connect outcomes, leverage points, user experience, and collaboration. It’s a repeatable way of turning new capabilities into trusted, widely used products—the real differentiator between firms that experiment with AI and those that transform with it.

Q7. Where should an AI leader start inside a large, complex firm?

By defining desired outcomes first, not technologies. Then identify the leverage points—shared abstractions or bottlenecks—where AI can make the biggest difference. Build from there, iterating with users and respecting the regulatory frame from day one.

Q8. How should success be evaluated?

Not by sophistication, but by scale of impact. The winning metric will be how much the organization’s habits, workflows, and results improve as a result of AI adoption.

Q9. What’s the broader lesson for other firms?

That AI adoption is a human problem disguised as a technical one. Institutions that treat it as product work—anchored in outcomes, leverage, user experience, and collaboration—will move faster and compound advantage over time.

One of our motivations for starting AInvestor was to create a reason to actively engage with AI in an operational setting—learning by doing. I maintain active editorial oversight of instruction, model, and platform choices, but much of the above summary was written by AI. In the context of what we’re doing, I see this as a feature, not a bug. By experiencing the highs, and yes, the lows, we can better understand both the possibilities and the limitations of this new generation of AI.

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis

AI’s real challenge inside large financial institutions isn’t data, models, or even governance—it’s adoption. A recent job posting from a major asset manager reflects this truth perfectly: success now depends on how well firms translate technical capability into products people actually use. The questions below unpack what this shift means for asset managers, and why product thinking—not just AI expertise—is fast becoming the differentiator.

I. There Is No AI Without Adoption: One Large Asset Manager Seems to Understand

A. Core Idea

• True success in enterprise AI isn’t just about technology—it’s about adoption.

• A major asset manager’s Head of AI Product role signals that the firm understands this.

• “Product” is the key word: the bridge between engineering and business outcomes.

II. The Role and Why It Matters

A. Head of AI Product Management

• Tasked with leading an enterprise-wide AI strategy across Operations, Finance, Investments, Asset Management, and Insurance.

• Focused on building roadmaps, driving adoption, and ensuring impact, not just deploying technology.
  Represents a cultural shift: from “AI experiments” to AI as product.

B. Why Framing It as “Product” Matters

• Product thinking forces empathy for users and accountability for outcomes.

• Encourages iteration, prioritization, and translation between technical and non-technical teams.

• Centers adoption as a design objective, not a happy accident.

III. The Core Challenge: Coordination and Translation

A. Horizontal & Vertical Complexity

• Different departments have distinct objectives, languages, and priorities.

• Aligning these requires fluency across business, data, and engineering.

B. The Translator Role

• The AI Product Lead must bridge C-suite vision, engineering feasibility, and user reality.

• Success hinges on creating shared understanding, not just shared tools.

IV. Product Mindset: The Missing Ingredient

A. From Capability to Product

• At Tudor, tech innovation served investment outcomes—the “product” was performance.

• Product management reframes that process: start with the outcome, define leverage, optimize user experience, and build collaboratively.

B. Lessons Learned

• Building something great doesn’t ensure adoption.

• Failed launches often result from misaligned incentives, unclear users, or missing collaboration.

• Adoption must be intentional, designed in from the start.

V. The AI Adoption Flow

A. Sequence of Work

1. Desired Outcomes → Define what success looks like.

2. Identify Leverage → Find abstractions that create cross-functional value.

3. Optimize User Experience → Design tools that fit real workflows.

4. Collaborate on Build → Ensure buy-in and iteration across teams.

B. Core Principle

• Adoption is a process, not an event—it evolves with user behavior, not just user feedback.

VI. The Roadmap for a New AI Product Leader

A. Understand the Third Rails

• In asset management, governance, compliance, and data rules are design constraints.

• Knowing them early enables faster, safer scaling.

B. Define and Prioritize Outcomes

• Start with business impact: faster decisions, lower risk, better communication.
  Align AI interventions with measurable outcomes.

C. Provide Requirements, Resources, and Cover

• Enable collaboration with data, compute, and domain experts.

• Create organizational “air cover” so innovation can fail fast and learn faster.

D. Experiment → Implement → Innovate

• Large organizations can’t transform overnight; start with clear, representative use cases to learn what’s possible and what isn’t.

• Implementation is about developing core infrastructure, governance, usage, and impact at scale.

• Over time, implementation builds competence and trust until AI becomes part of the firm’s culture, laying the groundwork to reimagine how business is done.

E. Execute the Work

• Measure adoption by observing behavior change, not dashboards.
  Tailor communication for each audience: executives, engineers, and end users.

VII. Defining Success

A. Real Measure of Impact

• Success = organizational behavior change + improved outcomes.

• Adoption > sophistication.

B. The Duality of Engineering and Product

• Engineering builds capability.
  Product builds trust and adoption.

• Both are essential for durable impact.

VIII. Broader Reflections

A. Context and Perspective

• From the PC to GenAI, technology’s compounding effect has been extraordinary.

• Amid hype and skepticism, balance comes from focusing on users and outcomes.

B. The Big Takeaway

• Job postings reveal strategy: where firms spend, what they prioritize, and who they trust to lead.

• This one shows a firm that understands the real challenge of AI isn’t building it—it’s embedding it.

One of our motivations for starting AInvestor was to create a reason to actively engage with AI in an operational setting—learning by doing. I maintain active editorial oversight of instruction, model, and platform choices, but much of the above summary was written by AI. In the context of what we’re doing, I see this as a feature, not a bug. By experiencing the highs, and yes, the lows, we can better understand both the possibilities and the limitations of this new generation of AI.

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis

Introduction

I recently came across a paper that speaks to much of what my career has been about—using computers, code, data, and ideas to generate returns: The Disappearing Edge: AI, Machine Learning, and the Future of the Discretionary Portfolio Manager by Frank J. Fabozzi, Andrew Chin, Igor Yelnik, and Jim Liew. Their practical approach and understanding of the continuum between systematic and discretionary is what caught my attention. Human judgment isn’t going away, though, but when, where, and how it’s applied will evolve. I’d be inclined to call it the changing edge rather than disappearing.

Go to paper

Abstract

The discretionary portfolio manager’s role is evolving as artificial intelligence and machine learning increasingly supplement or replace traditional investment insight. This article explores how advances in large language models and deep learning are narrowing the discretionary edge once defined by judgment and narrative skill. A new model is emerging in which the portfolio manager acts as an allocator and model steward, rather than a sole decision-maker. We examine the implications for governance, performance, and risk and argue that firms that retool talent, workflows, and oversight may be best positioned to harness the promise—and manage the limits—of AI-driven asset management.

What follows are three styles of summaries we use to highlight and profile work we believe is especially relevant to investors today. Each offers a different angle on the material. Any shortcomings or errors are ours.

One of our motivations for starting AInvestor was to create a reason to actively engage with AI in an operational setting—learning by doing. I maintain active editorial oversight of instruction, model, and platform choices, but almost everything in the summaries below was written by AI. In the context of what we’re doing, I see this as a feature, not a bug. By experiencing the highs, and yes, the lows, we can better understand both the possibilities and the limitations of this new generation of AI.

Custom GPT for Light Edits

Click above to access a custom GPT we spun up to help clean up our own writing. It’s designed to apply the lightest of touches. If you’re a paying ChatGPT customer, it’s free to use (we don’t get a cut). Feel free to play around with it as much as you’d like.

Learnings & Takeaways

1. The discretionary edge is being redefined

• Author’s words: “The edge once associated with discretionary judgment may no longer reside in the decision itself, but in how PMs frame, interpret, and adapt machine-generated insights.”

• Asset management context: Core skills like interpreting earnings calls or reading sentiment are increasingly automated. The PM’s value shifts to model calibration, governance, and contextual judgment.

• Our add: A new source of edge will come from knowing when to accept it’s better to step back and let the computer do the work, and when it is appropriate to override.

2. Discretionary vs. systematic is a spectrum

• Author’s words: “No investment strategy is entirely devoid of either discretion or structure.”

• Asset management context: LLMs allow discretionary PMs to scale breadth and systematic managers to add qualitative nuance. The line between styles is now blurred.

• Our add: This spectrum has always been so, at least in the forty years we’ve been involved. What’s changing is variability in decisions–which is the essence of discretion–which the new AI-powered tools will increasingly be made a choice, not an artifact of circumstances.

3. Hybrid PM models are often superficial

• Author’s words: “Hybrid frameworks only succeed when discretionary and quantitative teams collaborate closely, share outputs, and have mutual accountability.”

• Asset management context: Many PMs use quant tools only when they confirm existing views, ignoring conflicting signals. True integration requires cultural, governance redesign.

• Our add: We’ve been involved in projects attempting to achieve synergies between the discretionary and systematic–it is hard. In addition to the considerations covered in the paper, the requirements around data and other inputs differ across the two dimensions. There is a whole article for us to write on these differences, why and how they matter.

4. AI is eroding the last bastions of human intuition

• Author’s words: “AI models may soon rival discretionary PMs in their ability to anticipate discontinuities.”

• Asset management context: Regime shifts, once the hallmark of human intuition, are now detectable through models trained on diverse, unstructured signals.

• Our add: Regime identification and correctly anticipating change is the holy grail. We’ll take whatever help we can get.

5. Complexity creates credibility risks

• Author’s words: “Model opacity can hinder the ability to perform robust performance attribution and risk oversight.”

• Asset management context: Complex ML may outperform, but without interpretability, fiduciaries struggle to defend exposures to boards, consultants, or regulators.

• Our add: Few things are less transparent than the real reasons behind a discretionary trader’s decisions. We supported one of the best for two decades; stylized explanations are the best you should typically expect. And it’s not like they are trying to hide anything. Optimistically, AI will facilitate creating credible audit trails (we address this with Michael Mauboussin in our discussion here).

6. Institutional readiness determines success

• Author’s words: “Realizing these benefits depends on institutional readiness, not just technological capability.”

• Asset management context: Dashboards and overlays alone don’t change outcomes. Firms must retool workflows, governance, and talent development for AI to be transformative.

• Our add: Knowing what your “good” looks like determines success. Good process, good data, good IP, good governance, good incentives, and more.

Go to paper

FAQs

Q1: How is the role of the discretionary portfolio manager (PM) changing in the age of AI?

From Decision-Maker to Model Steward
The article makes clear that the PM’s traditional role—making conviction calls based on qualitative judgment—is shrinking. AI now handles tasks like parsing earnings calls, summarizing filings, and mapping sentiment across assets. That leaves the human with oversight functions: interpreting machine output, calibrating signals to liquidity and risk budgets, and deciding when to override.

Context for Asset Managers
This isn’t a minor shift. In practice, a PM’s daily work increasingly resembles editing, curating, and translating model results rather than originating every idea. Firms that cling to the “star manager intuition” model will underperform firms that redefine PMs as model stewards.

Q2: Is the distinction between discretionary and systematic strategies still valid?

A Spectrum, Not a Split
The authors argue the binary is outdated. Every “discretionary” manager now uses quantitative tools, and every “systematic” manager makes discretionary choices in model design and parameter setting.

Practical Implications
For allocators and consultants, the relevant question isn’t “Is this discretionary or systematic?” but “Where does it sit on the spectrum, and how are human and machine inputs combined?” Discretionary PMs are now using LLMs to scale breadth across universes, while systematic managers use AI to add qualitative nuance. The edge comes from the integration design, not the label.

Q3: What is the reality of the “hybrid PM” model?

Superficial Integration Is the Norm
Many firms talk up “hybrid” approaches, but the paper shows that most implementations are shallow. Quant screens or sentiment dashboards are bolted on, and PMs selectively use them when they align with their priors. This is confirmation filtering, not integration.

Why Most Firms Get Stuck
The sticking point is governance and incentives. If accountability remains tied to the discretionary PM alone, then quant inputs will always be treated as optional. True hybridization requires reassigning decision rights and creating joint accountability between discretionary and quantitative staff. Without that, the “hybrid PM” is more myth than practice.

Q4: Can AI fully replace human intuition, especially during regime shifts?

AI Closing the Gap
The article notes that regime detection—once a human edge—is being eroded. Models trained on unstructured data (policy announcements, commodity flows, social sentiment) now flag early signs of discontinuity. These signals often surface before humans recognize them.

Remaining Human Role
Still, machines are not yet fully reliable in low signal-to-noise environments or under unprecedented shocks. Human oversight matters in spotting spurious correlations, contextualizing geopolitical nuance, and deciding when to discard model output. The edge is no longer “intuition versus machine” but “intuition applied to supervising the machine.”

Q5: What risks do complex machine learning models create?

Opacity as a Fiduciary Problem
High-dimensional ML models can improve out-of-sample prediction and Sharpe ratios. But they are often black boxes. When exposures can’t be explained, fiduciaries cannot defend them to boards, consultants, or regulators.

Governance and Accountability
Opacity also clouds accountability. If a model misfires, is the fault with the PM who accepted the output or the quant team who built it? Without interpretability tools and audit trails, performance attribution and risk oversight break down. The paper argues that complexity is acceptable only if paired with transparency frameworks like feature attribution and robust documentation.

Q6: What skills must tomorrow’s PMs develop?

Model Literacy as a Core Competency
The paper emphasizes that future PMs don’t need to code neural nets but must understand how models fail. This includes recognizing overfitting, model drift, and constraint violations.

Cross-Disciplinary Integration
The winning PM profile blends market expertise with technical fluency. They must translate model outputs into investment decisions that respect liquidity, regulation, and client mandates. Firms that treat PMs as “market-only experts” risk falling behind.

Q7: What should asset management firms do to adapt?

Beyond Dashboards and Overlays
Superficial adoption—dashboards, sentiment screens, or GenAI summaries—does not change outcomes. The authors stress that genuine adaptation requires embedding models in rebalancing, signal weighting, and transaction cost management.

Organizational Redesign
That means iterative feedback loops between quants and PMs, shared attribution of performance, and protocols for model ownership and override. Talent development must include prompt engineering and fine-tuning, with PMs contributing labeled data from their judgment process. Firms that fail to retool incentives and governance will be left with symbolic AI, not competitive AI.

Go to paper

Mind Map

I. The Disappearing Edge: AI, Machine Learning, and the Future of the Discretionary Portfolio Manager

Purpose: Explore how AI and ML are reshaping discretionary portfolio management.
Key Theme: The role of the PM is shifting from decision-maker to curator, interpreter, and steward of models.

II. Discretionary vs. Systematic: The Borderline

A. Traditional Definitions

• Discretionary: Relies on human judgment, experience, qualitative/quantitative inputs.

• Systematic: Driven by models with limited human intervention.

B. Continuum, Not a Dichotomy

• All strategies contain elements of both.

• Boundaries are blurring with AI/LLM integration.

C. Practical Differences

• Systematic: backtesting, attribution, structured risk controls.

• Discretionary: adaptable but inconsistent, harder to scale.

D. Grinold Fundamental Law

• Discretionary = depth (high IC, low breadth).

• Systematic = breadth.

• AI increases breadth for discretionary, depth for systematic.

III. From Edge to Erosion

A. Historical Edge

• Human ability: tone, sentiment, subtle shifts, intuition.

B. AI’s Capabilities

• Parse transcripts, detect contradictions, infer policy directions.

• Faster, scalable, cost-effective.

C. Human Relevance

• Still strong in low signal-to-noise and regime shifts.

• Edge shifting to framing and interpreting machine insights.

IV. The Myth of the Hybrid PM

A. Concept

• Human + machine = best of both worlds.

B. Reality

• Often superficial integration.

• Confirmation filtering: use models when aligned, ignore otherwise.

C. Structural Barriers

• Legacy incentives and accountability.

D. Stronger Models

• True hybrid = redesign roles, workflows, and accountability.

• Example: “Iron-person” PM curating signals and narratives.

V. Where the Edge Used to Be and Where It Is Going

A. Past Sources

• Interpreting qualitative signals, spotting regime changes.

B. AI Encroachment

• LLMs summarizing, DL uncovering nonlinearities.

• AI starting to handle regime detection.

C. New Human Value

• Oversight, interpretation, refining assumptions, contextualizing outputs.

VI. Lessons from the Machine Learning Literature

A. Early Findings

• Nonlinear ML > linear models in prediction.

B. Practical Limitations

• Predictive accuracy ≠ investability.

• Need governance, explainability, transaction cost integration.

C. Dimensionality & Complexity

• Reduce overfitting via dimensionality reduction.

• Balance complexity with interpretability.

D. NLP & GenAI

• Extend ML into unstructured data (earnings calls, filings, news).

E. Implications for PMs

• Interpretation and application > raw generation.

• Fluency with models critical.

VII. The GenAI Shockwave

A. Unique Capabilities

• Generate text, simulate scenarios, interpret qualitative data.

B. Applications

• Draft research, translate macro, test judgments.

C. Implications

• Challenges PM narrative role.

• Enhances efficiency, assumption testing, documentation.

VIII. Redefining the Edge from Discretion

A. Past Role

• Synthesizer of info, conviction builder.

B. New Role

• Model calibration to constraints.

• Oversight, integration, validation.

C. Talent Shift

• PMs need model fluency, cross-disciplinary integration.

IX. The Virtue of Complexity? Or a Crisis of Credibility?

A. Complex Model Promise

• Nonlinear, high-dimensional patterns = performance gains.

B. Caveats

• Dependent on methodology, diminishing returns in noisy settings.

C. Governance Challenges

• Complexity = opacity.

• Problems of accountability, interpretability, client communication.

D. Balanced View

• Complexity useful if responsibly managed with explainability.

X. Adapting or Disappearing: What Firms Need to Do Now

A. Superficial vs. Deep Integration

• Dashboards/overlays vs. core process redesign.

B. Organizational Changes

• Shared attribution, iterative feedback loops.

• Talent: model literacy, prompt engineering, fine-tuning.

C. Governance

• Protocols for ownership, overrides, escalation, auditability.

D. Bottom Line

• Institutional readiness > tech capability.

• Superficial adopters risk irrelevance.

XI. Conclusion: When the Machine Becomes the PM

A. Redefined Human Role

• Editors, curators, translators of model output.

B. Firm-Level Practices

• Embed ML into core, foster cross-functional collaboration.

C. Success Factors

• Interpretability, constraint calibration, governance alignment.

D. Future of Discretion

• Traditional edge eroded.

• New edge = model supervision, validation, stewardship.

Go to paper

Notes on contributors

Frank J. Fabozzi

Frank J. Fabozzi is a Professor of Practice in Finance at Carey Business School, Johns Hopkins University, Baltimore, Maryland.

Andrew Chin

Andrew Chin is the Chief AI Officer and the previous Head of Investment Solutions and Sciences, Chief Risk Officer, and Head of Quantitative Research at AllianceBernstein LP in New York, New York.

Igor Yelnik

Igor Yelnik is the Founder, Chief Executive Officer, and Chief Investment Officer at Alphidence Capital Ltd., London, United Kingdom.

Jim Liew

Jim Liew is an Associate Professor of Practice in Finance at Carey Business School, Johns Hopkins University, Baltimore, Maryland.

References

• Bartram, Söhnke M., Jürgen Branke, and Mehrshad Motahari. 2020. Artificial Intelligence in Asset Management. CFA Institute Research Foundation.

• Blitz, David, Thijs Hoogteijling, Harald Lohre, and Patrick Messow. 2023. “How Can Machine Learning Advance Quantitative Asset Management?” The Journal of Portfolio Management 49 (9): 78–95.

• Buncic, Daniel. 2025. “Simplified: A Closer Look at the Virtue of Complexity in Return Prediction.” SSRN Working Paper

• Bybee, Alexander. 2024. “Ghost in the Machine: AI Misalignment and the Limits of Model Control.” Ethics and AI Quarterly, forthcoming

• Cartea, Álvaro, Qianjin Jin, and Yichen Shi. 2025. “The Limited Virtue of Complexity in a Noisy World.” SSRN Working Paper

• Chen, Liyun, Markus Pelger, and Juanyi Zhu. 2024. “Deep Learning in Asset Pricing.” Management Science 70 (2): 714–750.

• Chin, Andrew. 2025. “Leveling the Divide Between Discretionary and Systematic Investing: How AI Enables Breadth and Depth.” The Journal of Portfolio Management 51 (6): jpm.2025.1.730.

• Grinold, Richard C. 1989. “The Fundamental Law of Active Management.” The Journal of Portfolio Management 15 (3): 30–37.

• Gu, Shihao, Bryan Kelly, and Dacheng Xiu. 2020. “Empirical Asset Pricing via Machine Learning.” The Review of Financial Studies 33 (5): 2223–2273.

• Kelly, Bryan, Semyon Malamud, and Kien Zhou. 2024. “The Virtue of Complexity in Return Prediction.” Journal of Finance 79 (1): 459–503.

• Kirtac, Kemal, and Guido Germano. 2024. “Sentiment Trading with Large Language Models.” Finance Research Letters 62: 105227.

• Kozak, Serhiy, Stefan Nagel, and Shrihari Santosh. 2020. “Shrinking the Cross-Section.” Journal of Financial Economics 135 (2): 271–292.

• Lopez-Lira, Ariel, and Yuehua Tang. 2023. “Can ChatGPT Forecast Stock Price Movements?” SSRN Working Paper.

• Rudin, Alexander, Igor Yelnik, Juan Antolin–Diaz, Frank J. Fabozzi, and Suhail Shaikh. 2025. “From Economics to AI: Integrating Discretionary and Quantitative Approaches in Asset Management.” The Journal of Portfolio Management 51 (1): jpm.2025.1.737.

• Yelnik, Igor. 2016. “On the Style Edge: Discretionary vs. Systematic.” The Hedge Fund Journal 116.

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis.

Introduction

It makes my world bigger.

That’s probably the best answer I can give to the question of why I’m doing AInvestor. The author of the note we are profiling here is a good example. I’m meeting new people and exploring their ideas—sometimes directly, sometimes from afar. I’ve never met or talked with Ben Lorica, but I’ve found his work to be a steady stream of signal-rich content since I started reading him.

Gradient Flow isn’t about investing per se, but it’s very much about the process of making present-day AI useful. In that sense, it has a lot to offer those of us looking to harness AI in the service of investment returns. Where I think I can play a role is in curating, translating, framing, relating, and making his insights practical for professional investors and the people who support them.

What follows are three styles of summaries we use to highlight and profile work we believe is especially relevant to investors today. Each offers a different angle on the material. Any shortcomings or errors are ours.

Go to article

One of our motivations for starting AInvestor was to create a reason to actively engage with AI in an operational setting—learning by doing. I maintain active editorial oversight of instruction, model, and platform choices, but almost everything in the summaries below was written by AI. In the context of what we’re doing, I see this as a feature, not a bug. By experiencing the highs, and yes, the lows, we can better understand both the possibilities and the limitations of this new generation of AI.

Custom GPT for Light Edits

Click above to access a custom GPT we spun up to help clean up our own writing. It’s designed to apply the lightest of touches. If you’re a paying ChatGPT customer, it’s free to use (we don’t get a cut). Feel free to play around with it as much as you’d like.

Learnings & Takeaways

1. Data is the foundation

• Author’s words: “The core issue in enterprise search is the nature of the data itself… garbage in, garbage out.”

• Asset management context: Research notes, models, policy documents, and compliance memos must have clear ownership, versions, and effective dates. Drafts and duplicates poison retrieval. Your “gold copies” are the only ones that matter.

• Our add: I would underscore the temporal element. Your data and research content will evolve over time; you don’t want to only consider the most recent but you do want to prioritize it.

2. Relevance is contextual

• Author’s words: “Relevance is deeply contextual and ambiguous.”

• Asset management context: A PM looking for “exposure” wants sector/theme/positioning, while compliance might mean a policy definition. Engines must understand user role, scope and perspective, not just keywords.

• Our add: The ability to filter for relevance is critical. [add here… your IP/view of the world… ]

3. RAG 2.0 is required

RAG colloquially refers to retrieval-augmented generation — a technique that improves LLM responses by grounding them in external data sources.

• Author’s words: “RAG is not a magic bullet… the reliable pattern is ‘RAG 2.0.’”

• Asset management context: Answers about GAAP vs. non-GAAP earnings, fund exposure, or risk limits must come from document intelligence → hybrid retrieval → reranker → model that cites sources and abstains if unclear.

• Our add: Think process, infrastructure, tools and orchestration design.

4. Curated answer engines work

• Author’s words: “The practical approach is ‘curated answer engines’ for specific domains.”

• Asset management context: Stand up narrow engines for earnings/models, risk/exposure, compliance policies, and investor relations (RFPs/DDQs). PMs and IR teams don’t trust a monolithic “Google for the firm.”

• Our add: Pulling on the orchestra metaphor, you need a deliberately curated team of players and instruments. Some parts of the AI stack reward broad capabilities, your role and task specific tools do not.

5. Service, not product

• Author’s words: “Enterprise search is a service, not a product.”

• Asset management context: Expect ongoing integration and tuning. Linking OMS, RMS, compliance, and SharePoint, for example, requires professional services. Budget for stewardship, not just licenses.

• Our add: Getting your data and IP to work effectively with AI isn’t a one-time decision or purchase. Budget and plan for the journey — you’ll be rewarded.

6. Reliability beats benchmarks

• Author’s words: “Enterprise search is never going to be turnkey… reliability comes from explainability, clear citations, and internal test sets.”

• Asset management context: Forget public LLM benchmarks. What matters is whether the system reliably points your PM to the right model tab or policy clause, every time. Or when it breaks, it breaks in a predictable manner.

• Our add: This is the AI equivalent of both sides of an accounting ledger footing out. For software developers — think unit tests. Moreover, include both relevant examples and questions that are intentionally unanswerable.

7. Agents are the future

• Author’s words: “The emerging third paradigm is the agent: ‘do this task for me.’”

• Asset management context: Beyond retrieval, agents will run exposure reports, pull comp tables from filings, or draft DDQ responses by orchestrating multiple steps across systems.

• Our add: This is where most of the investing “front office” will see AI in action and the attendant benefits. Massive efficiency gains can be plowed back into all the inefficient activities—talking, debating, thinking, long-form reading—that are the true alpha in the investing process.

8. It’s a systems problem, not an AI problem

• Author’s words: “Enterprise search is fundamentally a systems engineering and data governance challenge that happens to use AI.”

• Asset management context: Winning teams will focus less on model hype and more on curating data, encoding house rules, and governing workflows.

• Our add: Good engineering force multiplies good data science.

9. Pragmatic steps now

• Author’s words:

  • “Start with a Data Census, Not a Model Evaluation.”

  • “Ship Hybrid Retrieval with Reranking.”

  • “Stand Up One Curated Answer Engine to Pilot.”

  • “Evaluate Privately and Continuously.”

  • “Think in Workflows, Not Just Answers.”

  • “Budget for Integration and Stewardship.”

• Asset management context: Begin by cleaning research/model repositories, building one reliable compliance or earnings answer engine, and standing up private evaluation before expanding.

• Our add: Develop early on a clear picture of what good looks like. And get started.

Go to article

FAQs

Q1: Why is enterprise search still considered broken inside investment firms?

Messy Data as the Core Problem
The article argues that the real obstacle isn’t the model—it’s the data. Enterprise information often lacks clear ownership, governance, and version control. Stale, duplicative, and “shadow” documents creep into repositories, polluting the knowledge base and eroding trust. For investment firms, this translates to outdated models, draft decks, and policy versions that surface at the wrong time.

Contextual Ambiguity
Unlike the public web, the enterprise lacks universal authority signals. In practice, this leads to problems of timing and authority, among others. A query for “China exposure” might surface an analyst’s draft, a PM’s offhand remark, or the official risk report. Which one matters depends on recency and hierarchy.

Q2: What role does Retrieval-Augmented Generation (RAG) play—and why isn’t it enough?

RAG’s Dependency on Retrieval
The article stresses that retrieval-augmented generation (RAG) is “not a magic bullet but a component in a larger system.” If the wrong documents are retrieved, the generated answer is not just useless but potentially worse than a hallucination.

The Reliable Pattern: RAG 2.0
AI models are trained on data, but they are not databases. The fundamental implication is that the AI needs to be augmented with relevant and credible information. Initial efforts at this have been largely simplistic. A stronger architecture including document intelligence, a mixture of retrievers, and a reranker trained on business rules is needed. This ensures that when a PM asks about the impacts of tariffs, for example, the system points to the the latest facts, opinions and decisions within the firm.

Q3: Why are curated answer engines more effective than a single “Google for the firm”?

Domain-Specific Reliability
Employees want direct answers, not a dump of links. The article argues for narrow, curated “answer engines” with well-defined scope so users don’t have to hunt further. This makes responses more consistent, predictable, and trustworthy.

Investment Firm Application
In practice, that means separate engines for earnings/models, risk/exposure, compliance policies, and investor relations materials (Investment Memos/DDQs). PMs and IR teams won’t adopt a one-size-fits-all search box—but they will use tools that reliably serve their domain.

Q4: Why is enterprise search a service, not a product?

Fragmented IT Reality
The article highlights that data is spread across dozens of SaaS platforms, file shares, and legacy systems. A turnkey solution rarely survives contact with this complexity.

Platform Plus Services
The article stresses that search is not plug-and-play. A flexible platform must be paired with professional services—teams that handle integration with firm systems, encode business rules, curate domain-specific engines, and continuously clean, tune, and govern the data. For investment managers, this means budgeting not just for licenses but for the ongoing staffing required to keep answers reliable and trusted.

Q5: How should reliability in enterprise search be measured?

Limits of Public Benchmarks
Open-domain leaderboards are misleading because they cannot capture the private, messy, and contextual nature of enterprise data.

Internal Evaluation Suites
The article emphasizes building firm-specific test sets that reflect real questions—like, “How has the narrative on why to own this stock changed with the latest news?”—alongside unanswerables to test the model’s ability to say “I don’t know.”

Reliability is proven on your own data and use cases, with consistent, explainable answers and clear citations, not by chasing external accuracy scores.

Q6: What are the biggest pitfalls firms should avoid when building enterprise search?

Over-Reliance on Larger Models
A common mistake is believing that simply increasing context window or model size will fix retrieval issues. The article warns this raises costs and latency without addressing the trustworthiness of the output.

Monolithic Search Boxes
Broad, one-size tools are prone to unpredictability and low adoption. Domain-specific curated engines earn trust because users know what they cover and what they don’t. They reduce cognitive overload.

Q7: What pragmatic steps should asset managers take right now?

Start with a Data Census
The first step is auditing what data you have, where it lives, who owns it, and which version is authoritative. Without this foundation, no retrieval system will be reliable. Second, define what “good” looks like for your use cases.

Build in Small, Testable Increments
The article advises shipping hybrid retrieval with reranking, standing up one curated answer engine, and testing internally before scaling. For a hedge fund, that might mean piloting in the back office, compliance or covering earnings releases first, proving adoption, then expanding to other domains and use cases.

Q8: What’s next after search and chat interfaces?

The Rise of Agents
The article describes a third paradigm shift: from search boxes (“find me a document”) to chatbots with RAG (“answer my question”) to agents (“do this task for me”).

Agentic Workflows in Asset Management
For funds, this could mean agents that not only retrieve documents but also run exposure reports, extract data from websites, compare and contrast internal and external research, stress-test decisions, or monitor for events that might impact the portfolio—executing multi-step workflows with human checkpoints.

Editor’s note: The point of using agents to optimize efficiency is to maximize the time one can be “inefficient”—spending time with others or in reflective work that prepares you for the differentiated decisions which are the true source of alpha.

Go to article

Mind Map

I. Enterprise Search Reality Check

Summary: Despite advances in AI and foundation models, enterprise search remains difficult because the real obstacles are not models, but data and context .

• Disconnect between model capability (e.g., explaining quantum mechanics) and inability to answer basic enterprise questions.

• Obstacles lie in messy, unstructured, and duplicative enterprise data.

II. Foundational Rot: Data Quality Problem

A. Nature of Data

• Enterprise information lacks clear ownership, governance, and structure .

• Staleness, duplication, and “shadow documents” create ambiguity.

B. Solutions

• Appoint knowledge managers.

• Establish governance and data hygiene.

• Implement knowledge graphs to generate reliable signals.

III. Signal Problem: Ranking Fails

A. Web vs. Enterprise

• Web search thrives on authority signals (PageRank, clicks).

• Enterprise lacks clear authority; relevance is contextual.

B. Hybrid Retrieval

• Use BM25 (exact matches), dense embeddings (concepts), graph traversal (authority).

• Add an instructable reranker to encode business rules.

• Employ hard-negative mining and enterprise-tuned embeddings .

IV. Architectural Shift: Curated Answer Engines

A. From Links to Answers

• Employees expect direct answers, not link lists .

• Incorrect answers carry liability in critical functions.

B. Strategic Split

• Build curated “answer engines” for high-value domains.

• Blend internal docs, pre-written expert answers, and gated external enrichment.

V. Implementation Reality: Service, Not Product

A. Complexity of IT

• Data fragmented across many SaaS and legacy systems .

B. Deployment Model

• Platform plus services: flexible software + professional integration and tuning.

• Budget for engineering effort beyond licenses.

VI. Measurement Mandate: Prove Reliability

A. Limits of Public Benchmarks

• External leaderboards irrelevant; cannot capture enterprise context.

B. Internal Evaluation Suites

• Build gold-standard test sets from internal knowledge bases.

• Include unanswerable questions and multi-step queries.

• Prioritize explainability, citations, and predictable failure.

VII. Next Frontier: Agentic Workflows

A. Paradigm Shifts

• First: Search box (“find me a document”).

• Second: Chatbot with RAG (“answer my question”).

• Third: Agents (“do this task for me”).

B. Multi-Step Reasoning

• Agents plan and execute workflows: query, parse, cross-reference, synthesize.

• Workflows encoded as graphs (DAGs) with human-in-the-loop checkpoints.

VIII. What AI Teams Should Internalize

A. True Nature of Challenge

• Enterprise search is a systems engineering and data governance challenge, not just AI.

• Reliability > leaderboard scores.

B. Traits of Successful Teams

• Accept messy data reality.

• Build systems for predictability, trust, and auditability.

IX. Pragmatic Steps to Take Now

A. Action Items

• Start with a Data Census, not model evaluation.

• Ship hybrid retrieval with reranking.

• Stand up one curated answer engine.

• Evaluate privately and continuously.

• Think in workflows, not just answers.

• Budget for integration and stewardship

Go to article

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis.

Introduction

Coming across this interview Will England did with Dan Shipper pushed me over the edge. I couldn’t shake the idea for AInvestor all spring, but after listening to their conversation I was finally moved to action. Dan proves to be an excellent interviewer, and in Will, he has a guest who speaks with clarity and conviction. Hopefully, we’ve done them justice. What follows are the three styles of summaries we use, each offering a different angle on the material. Any mistakes herein are ours.

Will England - Managing Partner, CEO and CIO of Walleye Capital

Will sets the strategic direction at Walleye, overseeing strategy allocation, risk, and talent. He began his career as a quantitative researcher at Man Group’s AHL division and later worked as an analyst at Valor Equity Partners. A graduate of Princeton (BSE in Operations Research and Financial Engineering, Academic All-Ivy) and Oxford (Master’s in Mathematical and Computational Finance), Will also rowed internationally for the U.S. National Team and won the Oxford-Cambridge Boat Race.

I have never met Will, but was duly impressed when I first listened to his interview with Patrick O’Shaughnessy a couple of years ago on the Invest Like The Best podcast. They cover the world of multi-strategy hedge funds and get more into Will’s backstory.

Dan Shipper - co-founder and CEO of Every

Dan is the co-founder and CEO of Every, a media and product company focused on business and AI. He runs an AI-first operation, writes the weekly Chain of Thought column, and hosts the AI & I podcast. Before Every, he founded the enterprise software startup Firefly, which he sold to Pegasystems. Today he’s best known for practical playbooks on how to use AI to think, write, and build.

He sounds like our kind of guy.

One of our motivations for starting AInvestor was to create a reason to actively engage with AI in an operational setting—learning by doing. I maintain active editorial oversight of instruction, model, and platform choices, but almost everything in the summaries below was written by AI. In the context of what we’re doing, I see this as a feature, not a bug. By experiencing the highs, and yes, the lows, we can better understand both the possibilities and the limitations of this new generation of AI.

Custom GPT for Light Edits

Click above to access a custom GPT we spun up to help clean up our own writing. It’s designed to apply the lightest of touches. If you’re a paying ChatGPT customer, it’s free to use (we don’t get a cut). Feel free to play around with it as much as you’d like.

Learnings & Takeaways

In this interview, you’ll learn:

1. Why Walleye Capital made AI usage mandatory across all departments

2. How Will England leverages large language models (LLMs) to sharpen thinking and communication

3. What Walleye’s internal AI platform “Current” does for analysts during earnings season

4. Why England sees AI adoption as a responsibility to employees and investors alike

5. How cultural rituals—like AI leaderboards and weekly meetups—accelerate adoption

6. Why he records nearly all firm communications and envisions a “Borg-like” collective data lake

7. What lessons history—railroads, barbed wire, and robber barons—offers about technological disruption

8. How incentives and intellectual honesty guide England’s decision-making framework

9. Why journaling with AI helps him track family, work, and health in harmony

10. How he frames leadership responsibility in an AI-driven world

Some takeaways:

1. AI adoption is not optional at Walleye. England mandated AI fluency across all 400 employees—whether in trading, compliance, or accounting—arguing that refusing AI is like refusing the internet in 1995.

2. Cultural design drives uptake. Walleye runs AI meetups, tool leaderboards, and incentive systems to normalize experimentation, making adoption social rather than top-down.

3. Internal tools deliver measurable edge. The “Current” platform synthesizes analyst notes, broker PDFs, and transcripts in real time—now indispensable for stock pickers during earnings.

4. AI reshapes leadership communication. England drafts memos in bullet points and uses LLMs to generate polished prose in minutes, freeing time for higher-order thinking.

5. Recording everything creates future leverage. By capturing calls, meetings, and transcripts, Walleye is building a “collective memory” that LLMs can process for insights and risk management.

6. History shows disruption accelerates. England draws parallels between AI and 19th-century railroads, barbed wire, and industrialization—technologies that rapidly obsoleted old skills.

7. First principles matter in decision-making. Incentives and intellectual honesty are England’s core anchors; he measures his own performance through journaling and tracked habits.

8. Responsibility is multi-layered. England views AI leadership as a duty: to investors (maximize returns responsibly), to employees (prepare them for change), and to his family (model adaptability).

9. AI journaling reduces friction. Capturing daily reflections across family, work, and health takes 30 seconds with AI, making the practice sustainable and insightful over time.

10. Human intuition and machine patterning complement each other. England argues that quant models and neural networks mirror intuition, and the future edge comes from combining them.

Go to interview on YouTube

FAQs

Q1: Why did Walleye Capital make AI usage mandatory?

Framing AI Like the Internet

At Walleye, AI isn’t framed as a nice-to-have experiment; it’s considered a baseline competency on par with adopting the internet in the mid-1990s. Will England’s view is that ignoring AI today is like refusing to use email or Excel when they first appeared—an untenable position for anyone who expects to remain competitive in financial markets. That conviction led him to issue a firm-wide mandate: every employee, whether in trading, research, compliance, finance, or accounting, must learn to use AI tools fluently in their day-to-day work.

Sending a Cultural Signal

England made this directive explicit in a memo where he opened by saying, “I used ChatGPT to write this email. You should be using it too and be proud of it.” The message was not only symbolic but also practical. He wanted to strip away the lingering hesitation employees felt about AI—whether it was the fear of “cheating” or anxiety about job security—and replace it with a clear cultural norm: using AI is part of the job.

Competitiveness and Responsibility

In his eyes, it would be irresponsible for leadership to allow Walleye to fall behind when competitors are already moving quickly. By making adoption mandatory and leading by example, England aimed to give the firm an organizational edge, ensuring that every team member, regardless of role or technical background, benefits from the productivity lift AI offers.

Q2: What are the top challenges Walleye has had with adoption or implementation?

Employee Fear of “Cheating”

Many employees initially felt that using ChatGPT to draft memos or emails was a form of cheating. Some even went so far as to edit or “dust up” AI-written drafts to hide the fact that they had used the tool. England confronted this stigma head-on by sending a firm-wide email that began, “I used ChatGPT to write this email. You should be using it too and be proud of it.” His intent was to normalize AI use and make it clear that adoption was not just acceptable, but encouraged as part of Walleye’s culture.

Anxiety About Job Security

There was also real concern among staff that AI might eventually replace their roles. England reframed this anxiety by emphasizing that AI doesn’t eliminate jobs, it changes them. He compared the shift to the arrival of spreadsheets: those who failed to learn Excel became obsolete, but those who did became more valuable. To reinforce this message, he mandated that every employee—whether in analysis, compliance, accounting, or another function—develop a baseline level of AI proficiency to stay competitive.

Imperfect Tools and User Frustration

Another hurdle has been the imperfection of the tools themselves. AI systems often make mistakes, and England recalled one demo where the “demo gods were against me” and everything failed in real time. Employees, frustrated by these errors, tended to give up on tools too quickly. England worked to shift the culture, encouraging staff to embrace imperfection, iterate, and keep experimenting. His message: the direction of travel is positive, even if the tools aren’t flawless today.

Discovery and Diffusion of Best Practices

England acknowledged that it’s hard for employees to discover the best AI use cases on their own. To address this, Walleye created structures that make sharing and discovery easier. Weekly AI meetups allow staff to exchange prompts and examples. Firm-wide leaderboards highlight heavy users and celebrate adoption. Incentives were also introduced so that employees who suggest tools that later roll out across the firm receive recognition. These steps transformed AI adoption from a solitary exercise into a social, collaborative effort.

Governance and Data Capture

Finally, governance around data capture remains a challenge. Walleye records nearly every call and meeting to build what England calls a “Borg”-like collective memory for the firm. While he views this as essential for long-term advantage, it raises questions about what should or shouldn’t be recorded and how to handle sensitive material. England recognizes that governance structures need to keep pace with this level of data capture, ensuring that trust and compliance aren’t compromised as the firm pushes forward.

Q3: What principles guide decision-making at the firm?

Power of Incentives

England emphasized that most human behavior in organizations can be explained through incentives. In investments, aligning incentives sheds light on why companies, management teams, or counterparties act the way they do. Inside the firm, structuring incentives ensures employees are moving in the same direction as both the firm and its investors. His litmus test is simple: “Are the incentive vectors all pointing in the same dimension?” If the answer is yes, decisions are more likely to succeed.

Intellectual Honesty

A second anchor is intellectual honesty. England has a strong distaste for fluff and empty rhetoric; the real question is whether an idea, analysis, or thesis holds up under scrutiny. Leaders at Walleye are expected to be candid about mistakes as well as successes. This kind of honesty acts as a safeguard against bias—especially relevant when using AI tools, which can produce outputs that look polished but may be misleading if not interrogated carefully.

Measurement and Feedback Loops

Another principle is the discipline of measurement. As England puts it, “you can’t manage what you can’t measure.” He applies this personally by journaling daily—now assisted by AI—and tracking his health and workouts as time series. The same philosophy applies at the firm level: track operational and investment data rigorously, revisit it over time, and use it to identify patterns and lessons that can refine decision-making.

Harmony Across Domains

Finally, England sees decision-making as spanning multiple domains. Instead of striving for strict “balance” between family, work, and health, he pursues “harmony.” Trade-offs are not judged only by their financial outcomes but also by how well they align with personal responsibilities and long-term sustainability. This integrated view ensures that decisions serve the whole, not just one isolated area of life or business.

Q4: What role does data capture play in Walleye’s AI strategy?

Creating a Collective Memory (“The Borg”)

Walleye records nearly every Zoom, phone call, and meeting with the explicit aim of building what Will England calls a “Borg”-like collective system—a firm-wide memory where nothing important is lost. The vision is to connect text data like emails, notes, and transcripts with numerical data such as market prices, accounting, and internal performance metrics, so that the firm can operate on a unified information base.

Risk Oversight and Decision Auditing

Daily risk calls are captured in full and then processed by language models. This allows the risk team, and leadership more broadly, to revisit not just what decisions were made but why they were made in context. By preserving the reasoning behind actions, the firm can spot blind spots, biases, or inconsistencies in decision-making over time, creating a feedback loop that strengthens risk oversight.

Enhancing Analyst and Portfolio Manager Productivity

Notes, broker PDFs, and earnings transcripts flow into Walleye’s internal tool, Current. This platform enables analysts and portfolio managers to instantly synthesize vast amounts of recorded information, rather than manually sorting through it. During earnings season in particular, the ability to pull together and process information in real time has become indispensable, freeing humans to focus on higher-order judgment while AI handles the heavy lifting.

Organizational Learning and Knowledge Retention

One of the less obvious but equally valuable benefits is the preservation of institutional knowledge. In most firms, when employees leave, their thought processes and reasoning walk out with them. At Walleye, recordings and transcripts capture those thought patterns, ensuring continuity for the organization and creating a training resource for new hires. Weekly meetups and demos also draw on this recorded material, turning data capture into an ongoing learning loop.

Future Vision: Full Data Lake Integration

England sees today’s recording practices as the foundation for a broader strategy. The long-term goal is to unify all data streams—structured and unstructured—into a single collective system. Once text and numerical data are connected in one place, AI can go beyond summarization to surface higher-order insights, such as linking the content of risk conversations to subsequent portfolio outcomes. In his view, this kind of predictive intelligence will define the firms that turn AI into a real operating advantage.

Q5: How does Walleye think about humans and machines working together?

Complementary Processing

At Walleye, the relationship between humans and machines isn’t viewed as a zero-sum competition but as a complementary partnership. Will England draws a parallel between human intuition and neural networks: both operate in high-dimensional spaces, processing complexity in ways that are difficult—sometimes impossible—to fully explain. Just as a seasoned investor may “feel” that a trade is right based on years of pattern recognition, a machine learning model can surface relationships across thousands of variables that no human could parse line by line. The two modes of reasoning mirror one another, and when combined, they open up possibilities that neither could achieve alone.

Extending Intuition

For England, the real opportunity lies in using machines to extend human intuition. Neural networks can highlight patterns or correlations that may feel intuitive once revealed but would have been nearly impossible for an analyst to uncover unaided. That doesn’t mean outsourcing judgment entirely. He stresses that tools are like engines—they need a plane to carry them somewhere.

Human Judgment Still Central

Humans still design the “plane”: setting objectives, framing problems, and ultimately deciding what to do with the machine’s output. In this way, AI becomes less about replacing decision-makers and more about expanding their cognitive reach, giving portfolio managers and risk teams access to insights that would otherwise stay hidden in the noise of markets and information flow.

Q6: What historical parallels inform England’s view of AI?

Railroads and Industrialization

England often looks back to the late 19th century for perspective. He points to the railroads, barbed wire, and the wave of industrialization as examples of technologies that reshaped the economy almost overnight. Railroads connected markets that had once been isolated, barbed wire transformed how land could be used and defended, and industrialization moved production from manual craft to mechanized scale. Each of these breakthroughs made entire categories of skills obsolete within a single generation. For him, AI represents the same kind of structural shift, but one that is unfolding in markets and information processing rather than physical infrastructure.

Frontier vs. Civilization

He also uses the metaphor of the American frontier to describe how technology adoption works. Startups are like cowboys—pushing into uncharted territory, experimenting, and breaking rules. Larger institutions, by contrast, act as the builders of civilization: they bring order, scale, and governance once the frontier is established. England sees AI as today’s frontier. The first movers push boundaries with speed and experimentation, while established firms like Walleye have to balance that frontier spirit with the resources and discipline needed to operate at scale.

Acceleration of Change

The most striking difference, in England’s view, is the pace. Railroads and industrial infrastructure took decades to build out. Entire lifetimes could pass with people adapting gradually to new technologies. AI is moving much faster. Shifts that once unfolded over a generation are now being compressed into a few years. For a hedge fund competing in real time, that means adaptation is no longer optional—it’s urgent. Firms that hesitate risk being left behind before they even realize the ground has shifted.

Q7: How is AI changing Will England’s personal workflow?

Bullet Points to Polished Prose

England explained that he thinks best when he writes out his thoughts. In practice, he now begins important memos or emails by jotting down bullet points—what he’s thinking, why it matters, and the context around it. He then feeds that outline into a large language model, often along with prior material he’s written on the same subject, and asks the system to draft the communication in his own voice. This workflow means he can focus on ideas while the model handles the mechanics of stitching sentences together.

Time Savings

This shift has cut hours off his process. Lengthy memos that once might have taken four or five hours can now be finished in 15 minutes. For someone in a leadership role where communication is constant—whether with employees, investors, or counterparties—that time savings adds up quickly and allows him to spend more of the day on higher-order decisions rather than wordsmithing.

Sharper Thinking

England is clear that using AI doesn’t dilute his thinking. Quite the opposite—he finds it clarifies it. By separating conceptual work from linguistic polish, he avoids getting bogged down in sentence structure or stylistic flourishes. He pointed out that much of writing is really “the tying your shoes part,” and letting a model handle that frees him to sharpen the concepts he wants to convey. The result, in his view, is clearer and more effective communication.

Go to interview on YouTube

Mind Map

I. Walleye Capital’s AI Mandate

A. Organizational Shift

• AI usage made mandatory for all 400 employees.

• Applies to all departments—investment, compliance, accounting, legal.

• Framed as equivalent to refusing the internet in 1995.

B. Leadership Role

• Will England leads by example, positioning himself as “chief AI evangelist.”

• Responsibility extends to employees, investors, and broader ecosystem.

II. Cultural Adoption Strategies

A. Social Structures

• Weekly AI meetups for sharing use cases and prompts.

• Firm-wide AI usage leaderboards to encourage competition.

• Incentives for suggesting tools adopted across the firm.

B. Overcoming Fear and Resistance

• Messaging that AI use is not “cheating.”

• Normalization of imperfect tools—public demos even when flawed.

• Emphasis on experimentation and accessibility.

III. Tools and Applications

A. Internal Platform: “Current”

• Synthesizes analyst notes, broker reports, transcripts.

• Critical during earnings periods; provides real-time analysis.

• Widely viewed as indispensable by portfolio managers.

B. Productivity Enhancements

• Large language models (LLMs) used for communication, writing memos, and emails.

• Drafting shifts from hours to minutes, freeing time for higher-level tasks.

C. Quantitative Strategies

• Longstanding use of advanced statistics and sentiment analysis.

• Integration of unstructured data with LLMs.

IV. Data and Risk Infrastructure

A. Recording and Collective Memory

• Nearly all calls and meetings recorded.

• Built toward a “Borg”-like collective data lake for firm-wide memory.

• Enables LLMs to provide retrospective insights and risk monitoring.

B. Governance and Risk Management

• Recorded data supports oversight and predictive risk processes.

• AI embedded into daily risk calls and control center activities.

V. Historical Analogies and Lessons

A. Technological Disruption

• Comparisons to railroads, barbed wire, and industrialization.

• Past innovations rapidly made old skills obsolete.

B. Frontier vs. Civilization

• Cowboy era and frontier metaphor: individual exploration vs. structured systems.

• Parallel to startups vs. institutionalized firms in technology adoption.

VI. Human and Machine Decision-Making

A. Human Nature and Patterns

• Human nature seen as timeless, shaping markets across centuries.

• Investing blends repeating patterns and human-driven intuition.

B. Intuition and Neural Networks

• Neural networks and human intuition both operate in high-dimensional spaces.

• Machines complement intuition by surfacing unexpected but intuitive patterns.

VII. Leadership Principles and Responsibility

A. Decision-Making Anchors

• Core principles: incentives and intellectual honesty.

• Journaling and tracking personal metrics reinforce self-evaluation.

B. Journaling with AI

• Daily reflections across family, work, and health categories.

• Reduced friction with AI: 30 seconds to capture and process.

C. Responsibility Layers

• To investors: maximize returns responsibly.

• To employees: prepare them for disruption and change.

• To family: model adaptability and stewardship.

Go to interview on YouTube

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis.

Check out AI Street where we found this

Introduction

After three direct one-on-one conversations with me in the interviewer seat, this note marks our first step into the Community Wisdom part of the program. This is where we curate the best and most practical material we find on how top investors and builders are incorporating AI into their processes and investment decisions.

I’m happy to share a recent interview I came across with Mark Steed, CIO of the Arizona Public Safety Personnel Retirement System. It took place on The Institutional Edge podcast, hosted by Angelo Calvello. What struck me almost immediately was Mark’s practical, level-headed approach. The absence of hyperbole is refreshing. Hopefully, we’ve done justice to the conversation between Angelo and Mark. What follows are the three styles of summaries we use, each offering a different angle on the material. Any mistakes herein are ours.

Lastly, the value we provide, hopefully, will come not only from the content and how it is presented, but also from introducing new, quality sources to your attention. In that spirit, I encourage you to check out The Institutional Edge, hosted by Angelo, and AI Street, published by former Bloomberg reporter Matt Robinson. It was in Matt’s weekly that I found this conversation. I find his newsletter a great top-of-funnel resource.

One of our motivations for starting AInvestor was to create a reason to actively engage with AI in an operational setting—learning by doing. I maintain active editorial oversight of instruction, model, and platform choices, but almost everything in the summaries below was written by AI. In the context of what we’re doing, I see this as a feature, not a bug. By experiencing the highs, and yes, the lows, we can better understand both the possibilities and the limitations of this new generation of AI.

Custom GPT for Light Edits

Click above to access a custom GPT we spun up to help clean up our own writing. It’s designed to apply the lightest of touches. If you’re a paying ChatGPT customer, it’s free to use (we don’t get a cut). Feel free to play around with it as much as you’d like.

Learnings & Takeaways

In this interview, you’ll learn:

1. Why PSPRS focuses its AI program on two pillars: operational efficiency and decision-making.

2. How local, offline LLMs (e.g., Llama, Gemma) plus simple RPA handle secure document intake and field extraction.

3. Which unstructured-to-structured fields matter in PE diligence (partners, fund size/vintage, compliance officer, carry terms, portfolio metrics).

4. How ML improves predictions on small, non-normal, dependent datasets compared with classic linear models.

5. Where deep learning fits—pattern discovery in unstructured diligence—alongside human review.

6. What PSPRS’s governance looks like: decision logging, confidence calibration, human-in-the-loop, and parallel model vs. human runs.

7. How screening evolves toward a manager web portal once factor importance stabilizes with enough observations.

8. Why PSPRS runs on-prem first (security, cost) and what compute is required for large local models.

9. How team design (investing + data science) reduces bias and accelerates build-out.

10. Why PSPRS is skeptical of sentiment scraping as a trading edge amid flows, constraints, and misinformation.

Some takeaways:

1. Automate high-volume, low-value work to free humans for judgment. PSPRS targets doc retrieval, PDF parsing, field extraction, and first-draft memo generation so PMs spend time debating meaning, not chasing files.

2. Local LLMs solve near-term security and cost constraints. By running Llama/Gemma entirely offline and pointing models at staged repositories, PSPRS avoids sharing sensitive GP materials while proving value with quick R&D loops.

3. Structured outputs from PDFs are the backbone of the stack. Consistent extraction of partners, vintages, carry terms (with synonyms), compliance officers, and operating metrics feeds a queryable database that supports screening and attribution.

4. Use ML where statistical assumptions break. Small samples, non-normality, and dependence make linear regressions brittle; PSPRS applies ML to rank features and predict outcomes on structured data.

5. Treat deep learning as a black box—govern it accordingly. For unstructured packets, DL surfaces patterns, but PSPRS pairs it with calibration, clear decision scope, and parallel runs before granting decision authority.

6. Governance is a growth enabler, not a brake. Decision logging with explicit success criteria and confidence bands builds trust internally and with the board while de-biasing post-mortems.

7. Screening will shift to a portal with fewer, higher-signal fields. After enough observations, PSPRS will publish the handful of inputs that matter most, letting managers self-submit and enabling faster triage.

8. Right-size compute to the documents you actually read. A ~70B-parameter local model with ~96 GB RAM supports long filings and large context windows—enough for 10-Qs and data rooms without cloud exposure.

9. Blend backgrounds to reduce institutional bias. Pairing investors who learned data science with scientists learning investing keeps feature selection honest and problem framing grounded.

10. Be critical of sentiment feeds as alpha. Regulatory flows, rebalancing mechanics, and misinformation weaken any causal path from social sentiment to realized trades; PSPRS prioritizes verifiable, higher-signal data.

    Go to interview on YouTube

FAQs

Q1: What are PSPRS’s two primary AI use cases?

Operational efficiency and decision quality. Efficiency covers automated document intake (data rooms, PDFs), structured-field extraction, and first-draft memo generation. Decision quality covers ML on structured data and deep-learning pattern discovery in unstructured diligence materials, with calibration against stated confidence.

Q2: What specific efficiency wins are in scope right now?

• Robotic access to proprietary portals (including two-factor workflows)

• Bulk download of Private Placement Memorandums (PPMs), Due Diligence Questionnaires (DDQs), spreadsheets, filings

• PDF/OCR parsing to extract fields (partners, portfolio metrics, fund size/vintage, compliance officer, carry terms)

• Normalization into a queryable database

• Auto-draft investment memos for human review

Q3: How is PSPRS handling data security?

Local-first. Large language models (LLMs) run on PSPRS machines, pointed at on-prem documents. No internet connection is required for extraction, fine-tuning, or memo drafts. Enterprise cloud will be evaluated after proof-of-concept success.

Q4: Which models and tools are being used?

Open-source LLMs such as Llama and Gemma for local use; Robotic Process Automation (RPA) scripts for retrieval; a house database for normalized fields; attribution tools (e.g., LIME/SHAP-style methods, activation maps) where interpretability helps.

Q5: What compute is required for the local LLM approach?

A high-memory workstation. For example, a ~70B-parameter model requires on the order of ~96 GB RAM to comfortably process long filings within a large context window.

Q6: What does the manager-selection workflow look like today?

Front-loaded quantitative requests to GPs, followed by model-assisted analysis of returned spreadsheets and data-room documents. Screening filters will tighten after enough observations establish which factors truly matter.

Q7: How will the external-facing data collection evolve?

A manager web portal is planned. Once factor importance stabilizes, PSPRS will publish a concise set of required fields so managers can self-submit the most predictive information up front.

Q8: How is explainability addressed?

Every material recommendation—human or model-assisted—includes success criteria and a confidence level. PSPRS tracks calibration (e.g., “80% confident” predictions should land near 80% realized accuracy). Interpretable models and attribution tools are used where practical.

Q9: What is the governance model?

• Decision logging with explicit definitions of success

• Calibration tracking across confidence bands

• Human-in-the-loop for scope and overrides

• Board education on a recurring cadence (semiannual planned)

• Parallel runs for deep-learning decisions until performance is proven

Q10: How does PSPRS handle small samples and messy data?

By design. ML methods handle nonlinearity, non-normality, and dependence better than classic linear models. For alternatives, the pipeline converts unstructured PDFs into consistent rows, then builds evidence over time via calibration and parallel testing.

Q11: What is PSPRS’s stance on market sentiment analysis?

Skeptical. Flows, regulatory constraints, and misinformation weaken the link between surface sentiment and actual trades. Priority is on verifiable, higher-signal data tied to manager quality and portfolio fundamentals.

Q12: What KPIs matter for the efficiency program?

Cycle-time reduction (from data-room access to memo), throughput (docs/fields processed per day), staff hours saved, and first-draft memo quality (edit distance vs final).

Q13: What’s on the medium-term roadmap?

A multi-agent system:

The goal is to remove low-value, high-volume work from the team and focus human time on judgment.

Q14: How are bias and overfitting mitigated?

Team design blends investment and data-science backgrounds to challenge assumptions. Feature sets are constrained to auditable, repeatable fields. Models are judged by out-of-sample calibration and tracked against explicit confidence bins.

Q15: What must an asset owner have in place to do this well?

Talent, data, compute, governance.

• Embedded data scientists working hand-in-hand with PMs

• A pipeline that turns PDFs into structured, versioned records

• Sufficient local compute for private LLM workflows

• A written governance program with calibration, HIL, and board education

Q16: How do humans and models share responsibility for decisions?

Models prepare and score. Humans decide. Deep-learning recommendations run in parallel with investment-team decisions until the evidence base is large and stable. Overrides are documented like any other decision.

Q17: What does “good enough to ship” look like for productionization?

• Stable extraction accuracy on key fields (measured against human labels)

• Demonstrated calibration at chosen confidence thresholds

• Documented decision scope and escalation paths

• Board-level fluency and sign-off on policy

Q18: What is the ultimate target state for the memo process?

End-to-end automation for first drafts: models fetch, extract, analyze, and assemble a complete memo with sources, tables, and attribution. PMs then edit for judgment, risk, and fit—without spending time on retrieval, reformatting, or basic calculations.

Go to interview on YouTube

Mind Map

Please share if you know of a tool capable of creating clean, professional looking mind map graphics. We lightly played around with Mermaid and that didn’t cut it.

I. Central Theme

A. One‑line Summary

• AI at PSPRS centers on two pillars: Operational Efficiency and Decision Quality, under a clear governance program.

B. Why It Matters

• Free investment staff from high‑volume, low‑value work.

• Improve screening and underwriting rigor with calibrated models.

• Build trust with stakeholders via documentation and education.

II. Operational Efficiency (RPA → Extraction → Drafts)

A. Document Ingress

• Navigate proprietary portals and 2FA; bulk‑download data‑room materials (PPMs, DDQs, spreadsheets, filings).

B. Unstructured → Structured

• Parse PDFs/OCR to extract canonical fields: partners_count; portfolio‑company metrics; fund_size; vintage_year; track‑record stats; compliance_officer; carry/“performance bonus” synonyms.

C. Normalization & Storage

• Write extracted fields to a versioned, queryable database for analysis and reporting.

D. First‑Draft Investment Memos

• LLM‑generated drafts assembled from extracted tables and standard sections; PMs review and edit.

E. Verification Loop

• Human spot checks; precision/recall tracking on key fields; error triage back into prompts/parsers.

III. Decision‑Making (Analytics)

A. Machine Learning on Structured Data

• Feature discovery and prediction when linear assumptions fail (small N, non‑normality, dependence).

B. Deep Learning on Unstructured Diligence

• Pattern surfacing across documents/spreadsheets; acknowledge black‑box aspects.

C. Interpretability Aids

• Use feature attribution (LIME/SHAP‑style) and activation‑pattern inspection where helpful.

D. Calibration & Parallel Runs

• Track realized accuracy vs. stated confidence bins (e.g., 70%, 80%).

• Keep DL recommendations in parallel with human decisions until evidence is strong.

IV. Governance & Oversight

A. Decision Logging

• Every material recommendation records success definition and confidence; benchmark PMs and models over time.

B. Human‑in‑the‑Loop (HIL)

• Define decision scope and override paths; models assist, humans decide.

C. Board Education

• Build fluency with recurring briefings (semiannual planned); align vocabulary and risk controls.

D. Rollout Discipline

• Start with verifiable tasks (labeling, math checks), then expand to predictive uses as calibration evidence builds.

V. Data & Information Architecture

A. Sources & Constraints

• GP data rooms (PPMs, DDQs), filings (10‑Q), spreadsheets—heavy PDF bias; limited sample sizes in alternatives.

B. Field Dictionary & Schema

• Standardize key entities (team, economics, track record, compliance, operating metrics) for reliable extraction and analytics.

C. Manager Web Portal (Planned)

• After sufficient observations, publish the handful of factors that matter; enable self‑submission to speed screening.

VI. Tooling & Stack

A. Local LLMs

• Run Llama/Gemma offline on PSPRS machines for security; point models at staged repositories.

B. Automation & Parsers

• RPA for retrieval/staging; PDF/OCR parsers; templated memo assembler; internal database with provenance.

C. Multi‑Agent Vision

• Ingest → Clean/Normalize → Analyze/Attribute → Draft Memo → Compliance → Humans.

VII. Compute & Deployment

A. Hardware Profile

• ~70B‑parameter local model; ~96 GB RAM for long‑context processing (suitable for long filings and multi‑doc synthesis).

B. Security Posture

• Local‑first R&D; evaluate enterprise cloud solutions after proof of concept.

VIII. Screening & Workflow

A. Current Practice

• Front‑loaded quantitative request to GPs; model‑assisted analysis of returns plus data‑room docs.

B. Evolving Filters

• Identify the 5–6 most predictive fields once observations suffice; move to tighter pre‑filters.

C. Philosophy

• 80/20 filter—accept some false negatives; focus on quality of done deals.

IX. Stances & Opinions

A. Sentiment Analysis

• Skeptical as an alpha source (causality may run trade → sentiment; misinformation and regulatory/flow confounds).

B. Explainability

• Preferred when feasible; black‑box acceptable under calibration, clear scope, and HIL.

X. Talent & Organization

A. Embedded Data Science

• Two data scientists (investor→DS; DS→investing) to counter institutional bias and align with PM workflows.

B. Stakeholders & Enablement

• Board and executives with varied AI literacy—bring them along with education and transparent metrics.

C. External Expertise

• Former PM (computer‑science focus) consulting on LLM use cases.

XI. Roadmap & Status

A. Now

• Authorization to run local LLMs; R&D on extraction and memo drafting; verification‑first approach.

B. Next

• Semiannual board education; portal design; progressive expansion of model scope; continue parallel runs.

C. Later

• Evaluate enterprise cloud; scale multi‑agent workflow.

D. End‑State Goal

• Remove low‑value, high‑volume tasks; humans focus on interpretation and judgment.

XII. Metrics & KPIs

A. Efficiency

• Cycle time (access → first draft); docs/fields processed per day; staff hours saved.

B. Quality

• Extraction precision/recall on key fields; memo edit distance (model draft → final).

C. Model Performance

• Calibration accuracy by confidence bin; stability over time.

XIII. Illustrative Anecdote (Color)

A. Worst Pitch

• Formula 1 racetrack in Monterrey; glossy pitch books requested back—illustrates disciplined triage and scarcity realities.

Go to interview on YouTube

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis.

Armando Gonzalez is a tech entrepreneur at the forefront of the AI revolution. As the visionary CEO and Co-founder of RavenPack, he’s transforming how the world's top financial institutions harness the power of data analytics. With over 20 years of pioneering experience, Armando is a globally recognized authority in AI and systematic data analysis. In 2024, Armando took innovation to new heights by founding Bigdata.com, an AI research assistant that’s redefining how business and finance professionals access critical information.

I first met Armando when he reached out after I left Kensho Technologies, post-acquisition from S&P Global. That was six years ago. We instantly connected over a shared compulsion to supercharge the man-machine process for investing. Soon after, Armando invited me to join a panel on why natural language processing was going to be the crown jewel of AI. I represented the user. At the time, I felt clever comparing the state of NLP to that of black-and-white TV powered by vacuum tubes. Clearly, the state of play has moved forward considerably. Please enjoy our conversation.

Click here for FAQs and Mind Map summaries of the key concepts from our interview with Armando.

In this interview, you’ll learn:

1. How investors judge AI tools by their ability to generate alpha, manage risk, or reduce costs

2. Why renewal cycles are the real test of whether a data or AI product delivers value

3. How careful source curation and whitelisting build the foundation for trustworthy outputs

4. Why traceability and audit trails protect both credibility and compliance in financial settings

5. How knowledge graphs resolve entities across securities, making data AI-ready at scale

6. Why investors expect exponential ROI (5–10x) from data products to justify adoption

7. How subscription models provide sustainable economics for both data providers and users

8. Why compliance has become the key gatekeeper for scaling AI in regulated industries

9. How AI-related skills are emerging as a baseline competency alongside Excel and coding

10. Why smaller, more agile firms often gain an adoption edge over slower incumbents

Some takeaways:

1. Value is measured in hard outcomes. AI products in finance survive only if they generate returns, cut risk, or save costs—anything less fails at renewal.

2. The bar is set at exponential returns. Because adoption diverts scarce talent and budget, tools need to deliver multiples of value, not marginal gains.

3. Data quality comes first. Trusted sources and rigorous whitelisting determine whether AI outputs can be relied on in high-stakes investment contexts.

4. Audit trails preserve credibility. When outputs can be traced back to underlying sources, both clients and compliance teams gain confidence in the results.

5. Entity resolution unlocks scale. By connecting securities, subsidiaries, and products through a knowledge graph, Bigdata.com removes ambiguity and accelerates analysis.

6. Subscription economics align incentives. One or two good decisions can justify the fee, while recurring renewals signal durable value creation.

7. Compliance dictates the pace of adoption. A single violation can outweigh years of returns, making regulatory alignment central to every deployment.

8. Skills are shifting. Where Excel once differentiated careers, the ability to craft effective prompts is now becoming a core workplace requirement.

9. Adoption speed matters. Smaller firms can test, onboard, and integrate faster than larger incumbents bogged down by pilots and red tape.

10. Build vs. buy is no longer binary. The trend is to buy for speed and reliability, while building only where compliance, IP protection, or security demand it.

Introduction and Background

Rob Marsh: Welcome to AI Investor, where we focus on how AI can help investors generate better returns. I'm Rob Marsh.

It's my great privilege today to introduce Armando Gonzalez, founder and CEO of RavenPack and Bigdata.com. If there was a title for today's conversation, it might be data's role in providing an edge with a subtitle: Without data, there is no AI.

I first heard of Ravenpack probably eight or nine years ago when I was developing products at Kensho. For our listeners and readers, quick background: Kensho was a fintech company sponsored by major investment banks that was ultimately sold to S&P for its AI talent. Back then, I thought of you as a competitor. Not long after I left Kensho, we connected, and you kindly asked me to participate on a panel around why natural language processing was going to be the crown jewel of AI. I think you were very prescient in that, and even when we met you were already well into your journey. So I'm so excited to have you and have this conversation today.

Armando Gonzalez: It's a real honor, Rob, to have this conversation with you, especially since this conversation has been going on for many years between us. I've always considered you one of the real AI practitioners, which back then meant you understood how to extract value from technology.

Now it's so complex when we think about AI that the challenge is even bigger and exponential. It's difficult to have good conversations with folks in this space because of the amount of noise and exaggeration, as well as the expectations that have been put on what AI will do. Expectations are so high that it feels like we're at the top of the hype, and there is potentially a significant bubble forming, at risk of bursting.

Cutting Through the AI Hype

Rob Marsh: Absolutely. It's this weird paradox where there is so much hype. That's why there's value in focusing on process: What are people doing? What are the sources of the edge? What have they been doing throughout their careers as successful investors? And then looking at the core competencies of the technology—where they are now, where they're evolving—and focusing on the overlap.

I agree with you about the hyperbole. The challenge I have when talking with people or working with investors and companies is this expectation, almost like there's a magic eight ball that’s really easy to use. It takes some time to move people off that perception. But at the same time, I'll finish our conversation today asking you: If you had a magic wand, where would you wave it? Because the answer to that can be very practical.

It's weird—wanting to move people away from thinking it's magic, but still asking the question of where you would wave it to make things more practical and focused.

Armando Gonzalez: I would also say that what makes AI so interesting in financial services, where I've spent the last 25 years, is that there's always a way to measure it. It's either helping you generate alpha, reduce risk, or reduce headcount. More specifically for us, it's always been about signal and making our customers money.

If it doesn't do that, you get punished very quickly—within a subscription cycle of 12 months—if you can't deliver the value you may have demonstrated in trials. In our case, it’s harder than most trials because we have to demonstrate that historically you could backtest a strategy and actually make money, at least on paper. That's the decision point to buy a product.

In other cases, for consumer applications, you might play around with it, subscribe, and then realize after 12 months—or even 30 days—that it's not useful. For us, the bar has always been high. You have to add exponential value, not just a 1x return on investment, but ideally 5x or 10x, because of all the added costs associated with the firm using your product.

More importantly, there’s the opportunity cost of highly talented, expensive resources being dedicated to one data set or set of tools. This forces you, as an AI company, to ensure you’re connected to the value chain and know exactly where you fit to produce exponential value—not relying on marketing, a nice-looking brand, or a good-looking UI.

Data Quality as the Foundation

Rob Marsh: Coming from both the buy side and the consumer side for so many years, I’m very sensitive to opportunity costs before you even start actually using and consuming something. On the vendor side, one of my mantras is: Value is confirmed on the renewal, not the first contract. The first contract is a function of salesmanship and a feeling you can provide.

Armando Gonzalez: Exactly.

Rob Marsh: There’s also an opportunity, especially when we talk about Bigdata.com. As you said earlier, there is no AI without data—period. We’re learning quickly that there’s legitimate demand from AIs operated by our customers for high-quality, timely, trusted, time-sensitive information. This allows next-gen LLMs to make decisions and plans effectively.

We’ve seen how poorly they perform with low-quality training data. While they’re good at writing and generating concepts, they don’t fit in the real world of, say, writing a trading idea or creating a thesis for investing in one country over another. There are so many nuances in our space, where real money is involved, that we can’t rely on just running things through ChatGPT and being impressed by the output if the substance and data aren’t accurate.

It's cliché—garbage in, garbage out—but there’s also “gold in, gold out.” I’ve been doing the Big Moves letter, and you can see what can be done using publicly available information. But when I introduce behind-the-paywall data and structured data—things where someone has done materially important preprocessing, exactly what you do—you see a step-change improvement in the outputs.

Armando Gonzalez: I was running an exercise the other day using a very famous new AI search engine. There aren’t many famous ones, but there are lots of wrappers on top of other LLMs. It was a specific task: I wanted annualized market capitalizations for a series of cryptocurrencies.

In our world, we wouldn’t need an LLM; we’d go to a terminal and grab the data that should give the right values—market cap for Bitcoin in 2020, 2021, and so on. I ran it through this search engine, and it created a nice-looking report, explained the trends, gave me a table, even generated a chart. Then I did the same exercise on Bigdata.com, which gave me a straightforward table—no fluff, just the data.

When I compared the two tables, they looked very different. The other search engine’s output looked better—chart, explanation—but some values were completely wrong. For example, it had Bitcoin’s market cap highest three or four years ago, not recently. Nine out of ten values were materially wrong.

If I had started my investment research with those wrong assumptions about trends, every assumption moving forward would be flawed. That trade would go sideways. An audit trail is necessary in our space, and such mistakes are crucial—they could cost you your job.

Auditability and Source Trust

Rob Marsh: For most of our audience, the greater risk isn’t losing money on a trade; it’s losing credibility with clients or your boss. That’s a far greater risk.

Armando Gonzalez: Exactly. At the end of the day, job security and the ability to generate income are key human assets. AI should help preserve and grow income, not put it at risk.

I think we’ll see more AI brands—or, more specifically, data brands. You start with good data sources before you even think about models. Together, they’ll become brands of AI—“This AI only uses this class of data,” or “These AIs don’t use these sources.” I wrote an article pre–Gen AI about “tribal AI,” predicting more AIs that act like tribes, with moral codes, cultures, processes, and missions.

The AI’s output is a reflection of the data that goes in. It’s not necessarily that one model is better, but the data going into the analysis drives the difference in policy, actions, and recommendations.

Rob Marsh: You mentioned auditability. At the highest level, it’s about trust with the institution or model you’re dealing with. But there’s also the job-by-job traceability to the source. That’s important not just for trusting the output, but as a superpower—AI can, in some cases, leave a trail of what went in and how it was considered, which becomes data and an asset itself. How do Ravenpack and Bigdata.com support that trail and transparency?

Armando Gonzalez: It starts with a whitelist of sources you want to work with. We look at when a publisher was created, where it was created, whether it has a paywall, how many subscribers it has, whether it follows AP style, potential biases, and other factors. We rank sources—least biased, most read, best covered. These go into the whitelist.

Because we have to choose from billions of potential sources, and many automated news outlets pop up and disappear. There are also highly ranked sources that are pure clickbait. So we have a research process to determine which sources we track, rank, and feed with metadata. That way, the large language models using those sources can produce reports, summaries, or analysis with context about the sources and their origins.

We also build relationships with data providers whose content isn’t necessarily public—content behind a subscription or paywall. If customers are willing to pay for it, that shows trust. People don’t spend $50 or $100 a month on a service they don’t believe in. This gives us confidence that these sources will be trusted by our customers. As part of that relationship, we secure rights to use the content in AI models and ensure compliance for banks or hedge funds to be comfortable using it. Onboarding unique, high-quality, trusted sources before thinking about models is a massive and ongoing endeavor for us. This curation and AI-readiness is our business.

Client Control and Customization

Rob Marsh: As a user, I appreciate being able to rank providers differently and impose my own opinions, filtering through. It’s not just taking your word for it—I can incorporate my own IP. That’s mission critical.

Armando Gonzalez: Absolutely. Some customers have their own stack, models, and fine-tuning processes for certain tasks. They might have agents that call specific content types. For example, one customer prefers a certain earnings call provider over another. We have to support both, because clients may have different vendor relationships and compliance approvals.

The value of Bigdata.com is one API call with a single function change for the provider name, still getting standardized output designed for agents. This centralization works even if providers have already been vetted by compliance and are difficult to replace. We support “bring your own license”—if you already have rights to a provider, we can deliver that same content through our search, retrieval, and extraction technology, connected to our knowledge graph.

Our knowledge graph took over 20 years to build. We annotate and connect all publicly named securities—stocks, bonds, any tradable asset—with unique names and point-in-time identifiers, regardless of source. This entity resolution means your AI can ask about “Meta” and we resolve it to “Meta Platforms,” which owns WhatsApp, Facebook, subsidiaries, and products. Any data source covering any related element will be delivered in a consistently identified format, so AIs don’t need to disambiguate—they just request data, get relevant content, and use it in their stack, whatever model, delivery, or UI/UX they’ve built.

Monetization and Subscription Models

Rob Marsh: Pay once, use it [data and content] where I want to—that’s the direction it’s going.

Armando Gonzalez: Exactly. Rights themselves are critical. We’re educating the market on the value of data and safe monetization in the AI world. Models like paying per click or per article often fail—the economics are too limited, and AIs will optimize to minimize transactions. Subscription models are proven. One good trade can justify a subscription; two good trades can make it exponential.

This makes the case for a strong subscription model that builds provider confidence: there’s renewal potential as long as content stays relevant, well-covered, and produced with integrity, plus rights to use it in models. Most firms aren’t trying to fine-tune LLMs; they want retrieval augmented generation—feeding in the right text or reports when needed without training the model. Educating providers on this usage helps them join the ecosystem and grow the TAM without being exploited.

Compliance as a Core Stakeholder

Rob Marsh: As an investor, I need my counterparty to be successful too. If you’re valuable, I need you to be around to renew—otherwise, the game ends. One trade pays for a subscription; one violation can be fatal. Compliance in financial services can outweigh even the PM or dev team.

Armando Gonzalez: Yes. Compliance adoption is slower, seen as back-office or admin. Risk-focused teams value it, but PMs with P&L get faster value. We’ve historically delivered to quantitative PMs as one factor among many in multifactor strategies, retaining a 110% net retention rate by continuously adding valuable data.

Now the sell side is entering with Gen AI. Researchers, analysts, and sales traders need quick insights on complex, personalized themes or client portfolios. They need to ask quickly and get answers they can deliver to clients or use to make calls. These research copilots are emerging—every bank has its “own AI,” really a co-pilot. They can generate analysis, PowerPoints, summaries, and reports—but they still need good data delivered in a way that avoids hallucinations that would alarm compliance.

AI strategy now touches engineering teams, CTOs, and innovation officers. We work with dev teams tasked with delivering promised AI capabilities. It’s an investment in the future, and I think the sell side will be on the right side of this trade.

From Excel Skills to Prompting Skills

Rob Marsh: In a world with so much gray, it’s clear that’s the direction we’re going. I think back to my early days—in the mid-80s, I got a job because I could use a computer, after a FORTRAN class my freshman year. I was helping build tax partnership models on a bootleg Lotus 1-2-3, just two years after doing accounting on a calculator with pen and paper. Fast forward to now—I joke none of us would have the lives we’ve had without Excel.

Armando Gonzalez: Exactly. You could get a job because you knew Excel. Now we ask: What models do you use? Can you show your prompting skills? We’re sometimes not even asking if you can code—just if you can prompt AI to write code. Using these tools well will be key to getting jobs, more than competing with them.

Addressing IP Leakage

Rob Marsh: One concern in financial services is IP leakage. How do you address that?

Armando Gonzalez: There are degrees of concern. Protecting a firm’s own data starts with highly sensitive emails or IMs—no firm I know is ready to let that leave their systems, especially in banking. That’s the crown jewels—valuable but heavily untapped because of compliance.

Comfort comes with less sensitive categories: research reports, customer-facing content, structured CRM data (anonymized), and prompts (sanitized to avoid PII). You can run a query on an alternative data company without revealing the client wants to invest $20 million in it. Stacks are getting smarter about deciding which tasks stay within internal infrastructure versus which can be sent externally.

Our business has reached a point where firms confidently send us questions about financially relevant entities, knowing Bigdata has trusted sources. We return ranked, customizable results that feed into internal processes—decisions and analysis remain with the client, and we have no visibility into that.

Deploying Behind the Firewall

Rob Marsh: Is it on the roadmap for clients to spin up an instance of Bigdata.com locally, behind their firewall, leveraging the data and knowledge graph alongside their own unstructured gold mine?

Armando Gonzalez: Absolutely. We already have projects deploying and maintaining managed services within client VPCs. These instances process internal content alongside Bigdata.com’s content, allowing prompts, watchlists, and reports to be handled internally. This minimizes leak risk and meets technical/legal requirements. Some customers would outsource to us completely, but constraints keep it in-house.

Rob Marsh: You’ve got open-source data, paywall data, and behind-the-firewall data. Integrating all three is where the real gold comes from.

Armando Gonzalez: Many customers have unique relationships with information providers—sell side, research partners, investors—that others don’t. Analyzing that unique content with AI produces a different footprint than a new VC firm with six months of history.

Rob Marsh: And it’s not just the data—it’s the audit trail from raw data to decision. Having that inside the firewall changes the output dramatically.

Armando Gonzalez: Exactly. The process is the product. When investors put money in a fund, they’re investing in a process.

Rob Marsh: In banking, I didn’t need trade recommendations. You could tell me tomorrow’s Wall Street Journal headlines, but if it didn’t fit our process, it didn’t matter. That’s what I was paid for. Build versus buy isn’t binary for sophisticated clients—it’s about where in the stack you focus, especially with tech changing so fast.

Build vs. Buy in a Rapidly Changing Landscape

Armando Gonzalez: Post–ChatGPT, the trend was build—firms thought they needed to build the whole stack. Now, with costs and obsolescence moving so fast, the trend is buy. Build only for compliance, IP protection, and security; buy everything else, because you’ll fall behind if you try to build it all.

We specialized in NLP 20 years ago, proving alpha from text, and stuck with it. Clients told me to start a fund if it was so good, but I didn’t know how to run one. My clients are experts in that; they don’t want to be experts in NLP or big data infrastructure. If I keep delivering value in my specialty, I have a sustainable business.

Rob Marsh: That edge compounds over time. You have to start learning early to widen the advantage.

Armando Gonzalez: Indeed.

Lessons from Successful Adopters

Rob Marsh: What key lessons have you learned from clients successfully incorporating Gen AI into their investment process?

Armando Gonzalez: Speed of innovation is faster than you think, driven by developments outside financial services. Smaller, nimble organizations have an edge because they can adopt quickly without red tape, while larger firms risk losing share to them.

Rob Marsh: I see a barbell—large funds with capital and behind-the-firewall data, and nimble smaller firms. The speed of underlying tech and the innovation it enables changes how fast you can develop ideas, strategies, and infrastructure.

Armando Gonzalez: Exactly. Time to market is key—onboarding new data can take 6–12 months now, but if we can cut that to 6–12 days, you realize value (or lack of it) much faster. Many AI projects have been stuck in POCs for 2–3 years with no value yet.

Rob Marsh: The opportunity cost of not exploring other opportunities quickly can be even greater than the cost of onboarding.

Armando Gonzalez: Agreed.

If You Had a Magic Wand

Rob Marsh: Wrapping up—if you had a magic wand to wave at some part of the process, where would you wave it?

Armando Gonzalez: Compliance. We need to move faster while meeting requirements. That would foster faster adoption and innovation in capital markets, financial services, healthcare, and other regulated industries.

Rob Marsh: Great place to stop—I could keep you all day.

Armando Gonzalez: Thank you, Rob, for inviting me. It was fun.

Share

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis.

I. Introduction and Background

• Rob Marsh introduces AI Investor, focusing on how AI can help investors generate better returns.

• Guest: Armando Gonzalez, founder and CEO of Ravenpack and Bigdata.com.

• Theme: Data as the foundation of AI—"Without data, there is no AI."

• Shared history: Rob recalls Kensho and prior conversations with Armando on natural language processing as a key AI tool.

• Gonzalez notes today’s AI landscape is complex, noisy, and potentially in a hype bubble.

II. Cutting Through the AI Hype

A. Investor Expectations vs. Reality

• Misconception: AI as a “magic eight ball.”

• Need to focus on process, edge sources, and real competencies.

B. Financial Services Context

• AI must prove value through:

  • Generating alpha.

  • Reducing risk.

  • Lowering headcount.

• Trials are strict: subscription decisions hinge on demonstrable ROI (5x–10x, not just 1x).

• Exponential value required due to high opportunity cost of talent and resources.

III. Data Quality as the Foundation

• “No AI without data.”

• High-quality, timely, trusted information critical for decision-making.

• Poor training data leads to flawed assumptions and costly mistakes.

• Example: AI search engine report on crypto market caps produced attractive but factually wrong outputs.

• Audit trail and accuracy essential in financial use cases.

IV. Auditability and Source Trust

A. Risk Context

• Greater risk is credibility loss, not just financial loss.

• AI should preserve and enhance income/job security.

B. Source Evaluation

• Ravenpack/Bigdata.com uses a whitelist: evaluate publisher creation, paywalls, subscribers, AP style, biases, etc.

• Rank sources (least biased, most read, best covered).

• Partner with trusted subscription-based providers.

• Ensure compliance for banks/hedge funds.

V. Client Control and Customization

• Users can apply their own ranking and filtering.

• Support for multiple providers (e.g., earnings calls).

• One API with standardized outputs.

• Bring-your-own-license supported.

• Knowledge graph built over 20 years connects entities, ensuring accurate resolution (e.g., Meta Platforms vs. subsidiaries).

VI. Monetization and Subscription Models

• Subscription > per-click models.

• Proven economics: one or two good trades justify costs.

• Rights and usage compliance critical.

• Retrieval-augmented generation prioritized over model fine-tuning.

• Educating providers to join ecosystem sustainably.

VII. Compliance as a Core Stakeholder

• Compliance can outweigh PM/developer priorities.

• High retention rate by delivering quant value.

• Sell side entering with Gen AI research copilots.

• AI must deliver insights quickly while avoiding hallucinations.

• Broader adoption requires balancing compliance and innovation.

VIII. From Excel Skills to Prompting Skills

• Past: Excel skills secured jobs.

• Present/future: Prompting skills (including AI code prompting) becoming critical.

• Ability to use AI tools well defines job competitiveness.

IX. Addressing IP Leakage

• Sensitivity tiers:

  • Crown jewels (emails/IMs) kept internal.

  • Research reports, CRM data, sanitized prompts safer externally.

• Firms send entity-related queries to Bigdata; results return ranked and customizable.

• Analysis/decisions remain internal to preserve confidentiality.

X. Deploying Behind the Firewall

• Managed services within client VPCs.

• Integrate internal data with Bigdata.com sources.

• Supports prompts, watchlists, reports internally.

• Gold lies in combining open-source, paywall, and internal data.

• Process and audit trail become product itself.

• Build vs. buy decision depends on compliance/IP vs. broader functionality.

XI. Build vs. Buy in a Rapidly Changing Landscape

• Early post–ChatGPT trend: build entire stack.

• Now: buy for speed and relevance; build only for compliance/security.

• Ravenpack specialized in NLP, proving value without building funds.

• Sustainable edge comes from sticking to core specialty.

XII. Lessons from Successful Adopters

• Speed of innovation driven outside finance.

• Nimble firms adopt faster, gain edge.

• Larger firms risk losing ground despite resources.

• Time-to-market critical: onboarding new data in days vs. years.

• Opportunity cost of slow adoption can exceed direct costs.

XIII. If You Had a Magic Wand

• Armando Gonzalez: Wave it at compliance to accelerate adoption.

• Compliance improvements would unlock innovation across regulated industries.

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis

Q1: Why does data matter so much for AI in finance?

AI models are only as good as the data they consume. In finance, poor inputs can destroy credibility or lead to flawed investment assumptions. As Armando Gonzalez puts it: “There is no AI without data.” Trusted, curated, and auditable data sources are the foundation for making AI outputs reliable.

Q2: How do investment firms measure whether an AI tool is successful?

AI products survive only if they clearly generate alpha, reduce risk, or lower costs. Vendors must prove exponential ROI—often 5–10x—because firms incur opportunity costs when allocating scarce analyst and engineering resources. Renewal cycles, not initial contracts, are the ultimate test of value.

Q3: What risks come from using generic or poorly curated AI outputs?

Outputs that “look right” but rely on inaccurate data can derail investment decisions. For example, a polished AI-generated report on crypto valuations contained nine out of ten materially wrong figures. Without audit trails, firms risk not just money but their reputations with clients and compliance teams.

Q4: How do professionals ensure AI outputs are auditable and trustworthy?

Firms demand traceability—knowing exactly which sources fed an AI’s conclusion. This involves whitelisting publishers, ranking sources for bias and reliability, and maintaining an audit trail. For asset managers, this isn’t just about accuracy—it’s about credibility with clients and regulators.

Q5: What role does customization play in using AI for investments?

Asset managers want control. They may prefer one earnings call provider over another or want their own rankings of data sources. Platforms like Bigdata.com enable “bring your own license” integration and deliver standardized, AI-ready outputs across different providers.

Q6: What is the value of a knowledge graph in this context?

Entity resolution is essential. Knowledge graphs map securities, subsidiaries, and products into consistent identifiers, so when an analyst queries “Meta,” the system automatically connects to Meta Platforms and its business units. This removes ambiguity and accelerates investment research.

Q7: How do asset managers and vendors think about monetization?

Subscription models dominate. One or two good trades can justify the fee, making renewals the best indicator of lasting value. Pay-per-click or per-document models fail because AI systems are optimized to minimize those transactions.

Q8: Why is compliance such a critical factor in AI adoption?

One compliance violation can be more damaging than years of strong returns. As a result, adoption is often paced by compliance teams rather than PMs. Vendors that deliver AI in compliance-friendly formats—transparent, auditable, and source-controlled—earn greater trust and traction.

Q9: How are professional skills evolving with AI?

Just as Excel skills once opened doors in finance, prompting skills are now becoming a baseline expectation. Increasingly, it matters less whether someone can code, and more whether they can effectively direct an AI to produce usable outputs.

Q10: Should firms build or buy AI infrastructure?

The trend is shifting toward “buy for speed, build for compliance.” Post–ChatGPT, many firms initially tried building their own stacks, but the pace of obsolescence and rising costs have pushed them toward vendor solutions. Internal builds are now reserved for compliance, IP, and security-sensitive use cases.

Q11: Who gains adoption advantages—large or small firms?

Smaller, more agile firms can onboard new data and test strategies faster, often cutting cycles from years to weeks. Larger firms have resources and behind-the-firewall data, but they risk losing share to nimbler competitors if slowed by compliance and bureaucracy.

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis

AInvestor Dialogues, a newsletter and interview series where we sit down with top fund managers to explore how they are approaching AI—from the ground level of their processes to the high-stakes strategies shaping their portfolios.

Big Price Moves Explained, a newsletter where we engage with AI to concisely explain the biggest price moves in US equities each day. This is us living our (current) use case.

Community Wisdom, where we curate the best and most practical material we find on how the best investors and builders are incorporating AI into their processes and investment options.

Playbooks will be where we move beyond interviews to ‘how-to’ guides; and create detailed reviews or comparisons of AI tools relevant to financial professionals. Launch date for this series is still tbd. We’d love, though, to hear what you would be most interested in first.

Thanks for reading AInvestor! Subscribe for free to receive new posts and support my work.

Target audience

We are writing and curating content for professionals, whether they be heads of firms, portfolio managers, analysts, traders or technical in orientation and responsibilities. Our experience is that if we do this well, the material will be accessible and useful for the many others that are interested in learning how AI can be harnessed to generate better investment returns. If you don’t understand what’s presented—it’s our fault.

About the author

AInvestor is largely written and curated by me, Rob Marsh. I spent 26 years at a highly notable hedge fund, a few years running product for an AI fintech bought by S&P, and a fair bit of time in the C-suite of a non-profit.

What perspectives do I bring?

• Deep experience in the customer/user seats as an analyst, strategist and portfolio manager

• Extensive experience in developing technical solutions for finance and trading

• Relevant experience in the vendor seat at a startup

• Deep experience bridging front office and technical teams; bridging C-suite and staff

• In and out of AI initiatives since early '90s

Relevant Background

Client and Builder: Tudor Investment Corp, JUST Capital, Big Price Moves Explained: With nearly three decades of experience in asset management, including 26 years at Tudor as both a Portfolio Manager and Strategist, I developed a deep understanding of the processes and requirements demanded by large, top-performing investors in high-stakes environments. I actively built tech-enabled solutions, both personally and in collaboration. Use cases include analytics, databases, search, platform design, and extensive coding. I was also involved in a number of AI-related initiatives dating back to the early 1990s. For the past year and a half, I have been publishing a daily report explaining the largest price moves in U.S. equities leveraging AI to write much of the content.

Builder and Vendor: Kensho Technologies, Praxis Solutions, AInvestor.co: Served as Chief Product Officer at Kensho, a Wall Street–backed startup bought by S&P Global for $550mn in 2019. The business model was a talent arbitrage: exceptional engineers who wanted to develop cutting-edge technology for finance without working at a large institution. I was the first and primary employee with deep experience in the customer seat. At Praxis, I conceived and led the building of an AI supported platform that would use data and news to explain portfolio performance. At AInvestor.co, we are building a highly curated platform bringing together the best practices of notable investors and those that support them.

Startup Experience: Kensho Technologies (CPO), Intrinio (Board), JUST Capital (Chief Information Officer), Praxis Solutions (AI consulting to asset and wealth managers), various informal advisory relationships.

Posting schedule

AInvestor Dialogues: Shooting for once a week. Things might be a little uneven in the early days of our launch. Thank you for your patience.

Big Price Moves Explained: 5:59 am, every Tuesday through Friday + Sunday

Community Wisdom: Ad hoc as we come across voices and perspectives worth sharing.

Playbooks: Not launched yet—will keep you posted.

Our business model

Mission

Good, practical stuff; open for all

Target Market

Professionals +

Product & Services

Articles, web content, consulting & advising

Distribution

Website, email; podcast and YouTube on the roadmap

Economics

Free to read; sponsorship upon reaching a critical mass

Join the crew

Be part of a community of people who share your interests. Participate in the comments section, or support this work with a subscription.

People

Robert Marsh

@ainvestor

What matters most here are the 26 years I spent at Tudor Investment Corp, mostly working directly with Paul Tudor Jones, focusing on all things around why asset prices move and what to do about it.

Michael Mauboussin’s legendary career spans Wall Street, academia, and the science of complex systems. He was Chief Investment Strategist at Legg Mason Capital Management, partnering with the iconic Bill Miller during his record-breaking streak of beating the S&P 500. Previously, he served as Chief Investment Strategist and Head of Global Financial Strategies at Credit Suisse. Today, Michael leads Consilient Research at Morgan Stanley’s Counterpoint Global and has spent three decades as a beloved adjunct professor at Columbia Business School. Along the way, he’s chaired the Santa Fe Institute—the world’s hub for complexity theory—written investing classics like More Than You Know and The Success Equation, and reshaped how we think about behavioral finance, market expectations, and the role of luck versus skill.

I first met Michael thirty years ago, though most of our conversations have probably been at hockey rinks and coffee shops. Every time, my head spins. He’s that rare thinker who’s both brilliant and relentlessly practical. Which is why I’m thrilled he’s kicking off our investor series on how top minds are using AI to drive returns.

Click here for FAQs and a Mind Map summaries of the key concepts from our interview with Michael.

In this conversation, you’ll learn:

1. How to define and articulate a true investment edge using Michael Mauboussin’s BAIT framework

2. Why AI’s greatest potential may be bridging quantitative and discretionary investing approaches

3. How AI can simulate market views and audit investment decisions to mitigate human “noise”

4. Why documenting investment decisions is crucial and how AI provides honest feedback on decision processes

5. The significant learning challenge AI creates and how to structure analyst training around it

6. How AI can be instrumental in “sizing” to effectively monetize an investor’s edge

7. How to leverage AI for base rate analyses and premortems to improve decision quality

8. Why distinguishing between systematic and judgment-based tasks is crucial for optimization

Some takeaways:

1. An investor’s edge implies a belief different from market pricing that is expected to materialize with positive expected value. This means actively seeking situations where your assessment diverges from current market consensus, and you expect that divergence to resolve in your favor. Before investing, explicitly document why you think you have a behavioral, analytical, informational, or technical advantage.

2. AI can serve as the bridge between quants and discretionary investors, taking the “greatest hits” from both camps. These communities often operate separately, but AI can function as the “glue” by blending systematic rigor with fundamental insights, similar to how multi-strat center books systematically extract alpha from discretionary pods.

3. Noise and bias can be fought by simulating a “wisdom of crowds” with AI. Handing the same case to different human analysts yields wildly different answers—scaling this approach is inefficient. AI can create agent personas (Seth Klarman, Warren Buffett) that analyze ideas from multiple perspectives, quickly surfacing counterarguments and reducing randomness in assessments.

4. The model of an investor is often better than the investor themselves, and AI can audit the “slippage.” Investors frequently deviate from their own well-defined processes, costing 100 to 150 basis points of performance annually. AI can codify your ideal process and audit actual decisions against it, ensuring adherence to predefined methodologies and identifying behavioral deviations.

5. AI creates a critical learning challenge that requires new training approaches. To effectively use AI tools, one needs pre-existing knowledge to judge output quality. Future training should make analysts responsible for quality control of AI outputs, shifting focus from tedious data gathering to critical thinking about why results are good or bad.

6. Systematically monetizing an edge through proper sizing represents a major opportunity. Even sophisticated investors often lack systematic approaches to position sizing. AI can recommend optimal portfolio sizing based on expected value, volatility, and correlation, acting as a “co-pilot” that guides learning and improvement over time.

Introduction and Background

Rob Marsh: Welcome, everybody, to AInvestor. I am Rob Marsh, and today we’ll be diving into AI’s role in identifying, building, and sustaining an investor edge. Today, I’m joined by Michael Mauboussin, Head of Consilient Research at Counterpoint Global with Morgan Stanley Investment Management.

Michael, I think our paths first crossed in the mid-'90s when I was doing a tour with the Raptor group up in Boston for Tudor when Jim Pallotta joined. And it’s gone on from there as we’ve both progressed throughout our careers.

To start, I want to begin where your good friend Patrick O’Shaughnessy typically wraps up his interviews where he asks what’s the kindest thing anyone’s done for you? In that spirit, I’d just like to thank you for joining today. You’re too kind to be doing this. Thank you.

Michael Mauboussin: Well, Rob, it’s my thrill to be with you. I’ve always thoroughly enjoyed our conversations over the last quarter-century—and it’s shocking to say that it’s that long, or maybe even a little longer. I never fail to learn from our conversations. And so it’s a real pleasure to be with you to talk through some of these topics.

Rob Marsh: Well, again, thank you. Now, when I first asked you to be part of this, you were very humbled about the AI part of the conversation. And I have three counters to that.

One, you know a lot more than you let on. You also have a pretty deep bench within the family that supports you. Two, we’re not going to get very technical today. That’s not really the audience or what we’re looking to do. And third, even before AI is a consideration, you have to think about what problem you’re trying to solve, how you’re trying to solve it, and what the risks are. From my seat, you’ve spent a career helping the highest-caliber investors walk through that, so you can set your modesty aside for today.

Michael Mauboussin: Well, I will react to that, Rob. You said my kids are deep into AI, and I’ve learned an enormous amount from them. I ask them a lot of questions. And this is an issue that’s constantly on my mind because I think—and you and I share this view—that the world is going to change a lot in the next three, five, 10, 20 years, and it’s really incumbent on us to think through how that might happen.

And the interesting thing always to think about in the world of investing is what is immutable. So what doesn’t change over time, right? Buying things for less than they’re worth and selling them for more or whatever it is. And what is mutable? We have to be flexible about the mutable stuff, including things that allow us to do our process more efficiently or better.

It’s not false modesty. I don’t really know what I’m talking about, but it’s something I’m very attuned to, and I think it’s something we should all have a conversation about. As we go through this, a couple of things will pop up, but it’s such early days for me. I also feel a little bit generational, that I’m maybe a little too entrenched, so I have to think about opening up to make sure that I’m seeing everything for what it is.

Rob Marsh: As you say, generational wisdom goes a long way, especially when it’s wrapped in your humility and flexibility. So, one of the things I want to do as we get started is to provide a kind of rough outline for the conversation today. I think, first, we should start at the beginning, perhaps define the whole concept of an edge from an investor’s perspective. Why is it important? You’ve written a book, The Success Equation, which applies to this directly. Then, talking more explicitly about edge and framing it for the investor. You have your BAIT framework, which you’ve introduced before and written about, and then perhaps your experience with some of the investors you’ve worked with, stories around how these concepts have been put into action effectively. And from there, we could get into where AI can supercharge both the offense-oriented aspects of a process and the defensive aspects—which I would say is probably more applicable and helpful to more people than the right-hand tail.

Michael Mauboussin: Yeah. No, that’s awesome. Let’s jump in.

Defining Investment Edge

Rob Marsh: Okay. So what is an edge?

Michael Mauboussin: Well, I think the place I’d start is just an acknowledgment that if you’re an active, discretionary manager, it’s an extraordinary act of hubris. Because what we know—and these data have been around for literally decades, going back a century—is that active managers really struggle to beat relatively straightforward benchmarks, for example, the S&P 500.

So, an edge would be a situation where you have an investment with a positive expected value. It means you believe something that is different from what is priced into the market and that your belief is going to come to pass, or the market’s going to come to accept your view. And so, it’s extremely difficult. One of the reasons—and you pointed this out with luck and skill—is that even if the markets were perfectly efficient, just pretend for the sake of argument, and you had hundreds of thousands of people participating, you would have some that beat the market and some that did much worse.

On average, they’d be the market, to state the obvious, but you would attribute the outperformance and underperformance to either good or bad luck, by definition, in our little case study.

Rob Marsh: Yeah.

Michael Mauboussin: And when you realize that markets are pretty efficient—let’s just say that, we can qualify it—you realize that luck plays an enormous role in the results we see. And so our challenge, when we think about edge, is to be very disciplined about homing in on what is skill, and that’s going to be a function of a process. And then how do we find that, how do we identify and exercise that systematically.

So edge is everything. Let me just say one other thing about this that I think helps make it easier for people to recognize. Blackjack is very different than markets, but blackjack is a good example. Ed Thorp, who himself was a great investor, wrote a book back in the 1960s called Beat the Dealer about card counting. The idea is that if you play standard blackjack strategy, it’s about a half percent house edge. So, you’ll lose money over time, but you have a few drinks, have some fun, and that’s all good. But if you’re card counting effectively, you can swing those odds in your favor to anywhere between 1 percent up to 3 or 4 percent. In that case, you have a positive edge. If you can bet more and press your bets a little bit, you have a positive expectation. That’s just a really nice, vivid way of saying that most times you don’t have an edge, but from time to time, things come along that allow you to have some sort of positive expected value, and you want to be in a position to take advantage of it when those things show up.

So it’s really—it sounds very fundamental to say you need to think about this every day, but it’s always asking the question, “Why do I think I know something that everybody else doesn’t know? Why do I think this transaction has an expected value when the person on the other side of my trade doesn’t seem to think so?”

Rob Marsh: But as you’re talking through that, I recognize that as an experience. I’m even taken back to sitting on the back porch with my dad in Southern California when I was in my early 20s. I was saying I didn’t want to be an accountant and wanted to get into this investment thing, and he asked me, “Why you?”

Michael Mauboussin: It’s a fair question. Every morning when you wake up and look in the mirror, that should be the question you ask. That’s right.

Rob Marsh: Exactly. So with that background, maybe we could walk into your framework on how you’ve come to really define it and come up with a taxonomy, a schema, a process.

The BAIT Framework

Michael Mauboussin: Yeah. And look, I’ve drawn here from many, many other people, so to be clear, I don’t want to give myself much credit for this. I did organize it in the acronym BAIT. The idea is you have to have good bait to go catch the big fish. So I’ll give myself credit for that, but that’s about it. But Rob, just as we go through this, as you said, I think it’s not only important to have or pursue an edge, but it’s also very important to understand why you think market inefficiency or opportunities are there. Like, why is market efficiency being compromised or why are the opportunities there?

Rob Marsh: Yep.

Michael Mauboussin: And being quite overt about it. Ideally, writing it down and saying, “Here’s why I think I have the best of this particular situation.” So we end up using four different areas. And by the way, in some ways, you could say everything’s behavioral, but we’ve been talking about BAIT.

So, the B of BAIT is behavioral. The idea there is that humans are humans. When you read these wonderful histories of markets going back hundreds of years and you read what people were doing, it seems extremely familiar. It’s the exact same sets of emotions, the exact same things people go through today. Of course, we have all sorts of fancy technology at our fingertips, and people didn’t have that hundreds of years ago, but the emotions are absolutely consistent and identifiable. So, we can get into the details of behavioral, but that’s the first one. And as I said, in some ways, everything is behavioral.

The second one is analytical, which is the idea that you and I have the same information but we can analyze it with different degrees of skill. Probably the best, easiest illustration of that is thinking about institutions versus individuals. There’s a really large body of academic research that shows when institutions and individuals compete with one another, it is the institutions that win.

The example there is probably something like tennis. You go out and play with a professional tennis player—same racket, same court, same equipment, and so forth, but they’re just more skillful than you are in playing the game.

The third one is informational, and that’s literally having better information than other folks. That is tricky. Obviously, you would like to have information that other people don’t have, but you need to do all of this legally. People get in trouble skirting the edges of that. So there are two aspects that might be interesting.

One is complexity. There can be situations where information is essentially embodied in other things, and if you can extract it effectively, you can gain some insight. For instance, there’s a lot of interesting work on supply chains. Something happens at company A; what does that mean for company C downstream, for instance? Can you connect those dots in a way that others either are slow to do or can’t do well?

The other one is straight-up attention. It sounds silly, but our lives are dictated by what we pay attention to. For whatever reason, I do think markets feel like they have a kind of collective spotlight on what is important at the moment. And that could be for the economy overall, an industry, or a very specific company. So what you pay attention to is really important. I will just say this is me riffing, this is commentary. I don’t know that this is completely supportable, but I found it fascinating when we had, what was it, Nvidia Monday, in January, when the stock went down 15 percent—it was like a $600 billion market value loss in one day. And it was on the back of this information about DeepSeek, this Chinese model that came out. And what was interesting—you were mentioning my own family—I’d had conversations with my kids literally about the two technical papers that DeepSeek published, and one came out in December and one came out in January, like a week or two before Nvidia Monday. So there was nothing technical that was not in the complete public domain. The only thing that I saw that was different is that DeepSeek had a lot of downloads on the app store over the weekend, but literally in terms of what they were doing, the technical aspects of it, basically all was in the public domain. And it was just like somehow the market one day woke up and said, “Oh my god, this is a big deal.” And by the way, the stock has since recovered and so on and so forth. But that’s a fascinating example of attention.

Rob Marsh: I was befuddled as well because I’m looking at it and I’m kind of like…

Michael Mauboussin: Right, exactly. And the last one is technical, which I find fascinating. This is basically asking, are there people buying or selling for non-fundamental reasons—or for reasons where they don’t want to buy or sell? And as a consequence, I could take the other side of that trade as a liquidity provider because they’re being induced to do something. So that’s the basic setup. And again, as I say, before you make an investment, it is really ideal to write down, “Here’s what I think is going on, here’s why I think I have the best of this particular investment situation,” and to see whether your investment works out or does not, if you’re right for the right reasons or right for the wrong reasons or wrong for the right reasons and so on.

Now the other thing to say, and I’m sure everybody listening fully appreciates this, but everything you do in investing is probabilistic. So when we say positive expected value, we’re not saying we’re going to make money with 100 percent certainty. What we’re saying is if we make many of these types of investments over time, we are going to do well and generate excess returns, understanding that for any particular investment, even with the best work done, best process, we are going to lose from time to time. And that’s just part of the game. So I just want to be clear that this is not saying we’re going to get anybody to be perfect. It’s about a probabilistic assessment of things.

Mauboussin's BAIT Framework for Investment Edge

AI Applications in the BAIT Framework

Rob Marsh: Exactly. Well, thank you. It’s funny, as you were talking, I’m writing down notes on the side, and I printed out a table. We were talking about this before we recorded, where I asked Google Gemini for this. Their deep research said, “Review all of Michael’s books, writings, recordings, whatever you can find,” and gave context about what we were talking about, and it had to generate a report. One of the artifacts in that report was a table of your framework, and it has the edge type, the description, the source of advantage, and some key characteristics around it. And I’m just scribbling now as we were talking, like AI helps in this area or element or that. So…

Michael Mauboussin: Well, we should go. If you’re game for it, we could go through some of these, and then we can go back and forth, Rob. So if you have something on your list that I’ve not written down. So behavioral, just to go back to behavioral, I mean, I think there are two or three big themes there.

One is the basic idea of overextrapolation. This is again very well documented in academic research: investors tend to overextrapolate. So things are going well, they think they’re going on well forever. If they’re doing poorly, they’re going to go poorly forever. And they don’t take into consideration—I don’t even want to call it regression toward the mean—but basically that things don’t grow to the sky.

And then related to that would be something like sentiment. So can we measure sentiment? Again, too much optimism tends to be bearish, too much pessimism tends to be bullish. And then the last thing is the wisdom of crowds, and this is the one I’m most interested in, and I think this is where we have some opportunities to use AI.

The wisdom of crowds basically says when you have a collective of people under certain conditions, they’re going to come up essentially with the right price. And the key is the conditions. And the three conditions that are properly mentioned are diversity or heterogeneity. We need our agents, our investors, to be different: long-term oriented, short-term oriented, technical, fundamental, whatever it is, throw them in the mix. That’s good. The second is an aggregation mechanism, which of course markets do beautifully: double auction markets. And then the third thing is incentives: if you’re right, you make more money, and if you’re wrong, you lose money and you go away. So it means the smart people are the ones that continue to play the game.

So here’s where it becomes really interesting: can we start to use some AI tools to measure some of these things more effectively? As we know, natural language processing is a really powerful tool for things like sentiment and probably to some degree diversity—measuring diversity, although that diversity needs to be manifested in market behavior.

So it’s almost the decision rules of the investors, but those are great examples. Overextrapolation would be another one where we could start to use these tools and just say, by applying things like base rates, so understanding historical patterns for example, sales growth rates or company profit patterns, you could really say to an LLM, “Hey, here’s what’s happened for growth rates. People seem to be thinking this one company’s going to grow to the moon. What do you think is the probabilistic assessment of something like that?” Right? So, that might be a way for us to very quickly—because it’s going to be able to digest the base rate data very effectively, it’s going to be able to look at a lot of different situations and, I think, shine a spotlight on some of the things that might be of interest. Is that—would you have anything you wrote down that would add to that?

Rob Marsh: Yeah. And just to riff off of that, you snuck in the word “agent” there, which has become very commonly used in the thought application within AI. And you’re using it a little differently, but you’re talking about AI using it to measure, and I scribbled down ‘measure versus proxy,’ because one of the things that can be done is to create agents that have a certain framework and direction that can proxy that independence, go surface the information, and present it to you in such a way that you could simulate, if you will, the public markets and a public discussion with some reliability that you have independence of thought.

The Noise Problem and AI Solutions

Michael Mauboussin: So, Rob, let me just tell you one quick story on this which I think is fascinating. There’s a very interesting book, I think it’s an incomplete book, but a very interesting book about the topic of noise, by Danny Kahneman, Cass Sunstein, and Olivier Sibony. They’re great guys, all of them. And the idea of noise is that when you give people judgments to make, whether that’s many people making a judgment on the same fact pattern or you making the same decision over time, there’s a lot of noise, a lot of randomness.

For instance, in the book, they describe insurance adjusters. You have to settle this particular insurance case. They give a folder to 50 different insurance adjusters at the same firm, right? So, they’re trained the same way, they’re supposed to be doing the same thing, and they come up with these wildly different answers.

So, here’s an example. I did—it’s actually in the book, although it’s all anonymous, mercifully anonymous—but I did this at an investment firm where I worked. I developed a single page, and folks were up for this. It was actually a real company, but I hid—I scaled all the numbers so you couldn’t identify it just by looking at the numbers. So I knew the stock price. And I literally asked people, “Write down the price at which you would be indifferent between buying and selling.” So that’s kind of fair value, whatever it is. Now, the stock price was $25, and the lowest number—we had about 30, 35 people participate—the lowest number was $5 and the highest number was $130, and the average of everybody was $25.

Right? So, the market itself worked. But here’s the thing. Now, you work in an investment organization and you hand that folder to analyst A, and she comes up with $5. You hand it to analyst C and he comes up with $130. Like, what are you going to do with that? So, very much to your point, they say, “Well, how do you mitigate noise?” Well, one is if there’s an algorithm—we can talk about that, too. If there’s an algorithm, make sure that you write it down and hew to it. That’s great. But the second thing they said is you need diversity. Do exactly that. Have 30 analysts cover the same stock. Why do we not do that? Wildly expensive. Wildly inefficient.

But you just described a process that absolutely allows you to achieve that objective in a really fascinating fashion, very quickly, and I think would replicate the vast majority of what we’d be after.

Right? So let’s look at XYZ. Your profile is you be Seth Klarman, you be Warren Buffett, you be Peter Lynch, whoever it is. Let them go at it with the tools that the LLM thinks those people would use. And you create that wisdom of crowds within. By the way, if you have a view on the stock, which you very well may, it will automatically give you counter-views to your own take based on the research and thought process of someone you likely respect. So that, I think, is really cool, and we’re working on some of that stuff now in our own organization. It’s early days, but it’s super fun and fascinating. You have to be thoughtful about prompting, to be clear on that, but it’s a really exciting area to explore.

Pod Shops and Center Books

Rob Marsh: So, Michael, at the risk of going off on a tangent, as you were describing the 30 analysts and how expensive it would be, who’s doing that right now are the pod shops. And it helps speak to why the center books—or for the non-initiated, ‘center book’ being the aggregated portfolio based on all the different independent inputs—have been so successful. But that’s just an aside.

Michael Mauboussin: But no, Rob, maybe we should stay there for just a second because A, I agree with you. B, by the way, the pod shops charge very high fees, but they deliver above and beyond those fees. So they’re extraordinarily expensive from the point of view of fees, but they’re actually delivering their fees and then some. But the other thing that I think is really important, and this ties into our AI conversation, is the one thing you want to think about is where in my process is judgment important and where can I be systematic? I think what these center books do is allow them to extract alpha or excess returns from all these different pods, but they’re doing it in a systematic fashion. And so that is a really key thing.

So, interestingly, why can’t the individual pods do that? Well, they’re doing a lot of good stuff, but they’re not taking the full systematic approach that the center book is allowed to take. So it’s really interesting. That’s the other thing I would just say when we think about our investment process—from identifying mispriced securities to how we analyze, to how we build portfolios, to how we hold them to monetize that edge—where along that chain can we be systematic, even if we’re discretionary investors? Where can we be systematic versus where are we going to be discretionary? It’s a super fascinating question. And that sort of goes back to the systematic community, the quants versus the discretionary guys. They still are kind of two different tribes, right? They don’t really talk to each other, but the quants know there’s much they can learn from the discretionary people, and the discretionary people know there’s a lot they can learn from the quants. And maybe AI is that sort of glue that brings these two communities together or takes the greatest hits of both these communities. And I think that’s literally what the center books are doing. That’s essentially a manifestation of that. But can we roll that out to other organizations in a fuller way?

Decision Documentation and Process Auditing

Rob Marsh: Yes, absolutely. I think connecting dots perhaps between the proxy, or using AI to proxy many different views or even understand actual analysts’ and PMs’ and traders’ views, with something you noted earlier is the ideal of being able to write down, track, and audit the decision-making process, and where can AI be used for that? Where can its core competencies be applied?

And my mind goes twofold. One is just the ability, as analysts are doing models and recording phone calls and meeting notes and everything else, and it all goes into your reservoir, which then can get processed. You get the pattern matching, the semantic understanding, the fuzzy logic capabilities, all of that processing it. So that’s one vector. The other—and I was having a conversation a couple of weeks ago with a friend and former colleague from the Tudor days who, by the way, is a big fan of yours, extremely process-oriented—is using AI to mimic him, not so he or his team doesn’t have to do the work, but almost as an audit or a control mechanism. Is he following his own well-defined, predefined process?

Michael Mauboussin: Yeah. I love all that. By the way, there’s a very famous paper from Lewis Goldberg in psychology, and the basic punchline is the model of man is better than man himself, or the model of a person is better than the person himself. So we’ll gender-neutralize that. But the idea is, I come to you, Rob, and I say, “How do you want to, what is your process?” And you tell me, and I write it down. And then I observe, “Does Rob follow what he says he wants to do?” And the answer is there’s always slippage between those things. And we know that, by the way, we even know that in the investment management business. There are examples of firms that help money managers with position sizing in their portfolios, and they start off by asking the managers, “How do you want to do this? What are your constraints?” and so on. They write it all down, and then the investor inputs all this data into the software, which allows the company to track the difference between what they’re actually doing and what they said they wanted to do. They can see that slippage, and it comes out to somewhere around 100 to 150 basis points of performance per year, which is not insignificant. So that’s the first thing I would just say, that I think you’re exactly right, and just the discipline of hewing to what we said we’re going to do.

The Learning Challenge with AI

The second thing, Rob, that the reaction I have, and this is maybe my overarching struggle with AI and its application, and you were talking about gathering all this information. My biggest core struggle is this chicken-and-the-egg problem. You’ve been doing this for a long time. You’re a very thoughtful guy. You kind of know what works and what doesn’t work. So when you put something into an AI program and you get an output, you have a very good filter to judge whether it’s of good or poor quality. You have a nose and a sense of this because you’ve been through the process yourself. Now, you were slower perhaps, and you didn’t have the reach that the machine has, but you’d been through the process.

You know, I built financial models when I was an analyst, so I know how to build a model, and I know there are certain nuances in these things. Even when I was an analyst, I never let my junior analyst ever touch my models because I felt it was so fundamental that I understood what was going on. I would go through this line item, something wouldn’t make sense, and it was only—I had to see it for myself and be in the middle of it. So there’s this idea of pre-knowledge that’s necessary to use these tools effectively.

That’s why I say I might be just the grouchy old guy or generational, which is I think that if you’ve learned to write and you’ve learned to model, then using these tools is fabulous, because you can judge very quickly. You can take all the good stuff and leave aside or figure out the bad stuff.

I had breakfast with a friend recently, and he was mentioning one of his workout partners is a professor of physics at a university in California. She was saying that for her physics students, they’ve never had higher grades on the problem sets and they’ve never had worse grades on the finals. The reason is that for the problem sets, they can apply AI. It’s an objective function, so there’s an answer. But for the finals, they’re sitting in a room with blue books, and they’re not learning. So the key is to learn.

Getting the right answer is obviously important in a lot of things, but when there’s a lot of judgment, it’s not always just getting the right answer. The process of getting to the right answer is really important. So that’s the second thing I’ll say is just this—this is what I struggle with, this need to know what you’re doing in order to use these tools effectively. And I worry that some people are going to jump the gun on that and not know what they’re doing.

Decision Documentation and Feedback

The third thing I want to say, and you also touched on this, which was fabulous, is really documenting your decisions. And I think this is where AI can be incredibly powerful. So leaving aside gathering information, as you said, pulling things from different areas, that’s all powerful, but even documenting decisions. Now, the reality is most people don’t do it. It requires some discipline, but many people just don’t. But if you start doing this every time you make an investment decision—and by the way, a non-decision is a decision—you just document it. You can now speak into your device. It’s not like you have to spend tons of time doing this. You can speak into your device for a minute to summarize where you are and just set it aside. And then periodically go back and let the AI comb through those decision-making processes and then say, “Hey, let me just tell you, it seems like when this happens, you do that. When that happens, you do this. When you do this, this seems to work out well. When you have this, you say these things, it tends to not work out well.” Right? So, that’s an example of how it could probably reflect back to you in a very honest way feedback on your decision process that may help you.

And I think one of the reasons people don’t document decisions is that it’s going to be embarrassing. You’re going to be wrong a lot, and you don’t want everybody in the world to know about it. But if you’re doing it for yourself, you’re just trying to improve. If you’re doing it for yourself, this could be an invaluable tool for people.

Rob Marsh: Yeah. It’s funny, I was starting to shake my head a little bit side to side when you say people not wanting to be wrong. In many respects, the industry self-selects out of that. You’re not going to be around very long if you’re that fragile.

But I want to go back to where you’re talking about your concerns about learning, how you were talking about when you were an analyst, how you wouldn’t let your juniors do it, and that there’s a visceral knowledge that you develop over the years by going through the tedium, like actually going, finding the data, understanding why the data is not what you think it is, or your vendor’s changed the format. And there’s so much knowledge that builds up over time that you can’t even articulate how it manifests, but you see something on a big spreadsheet and you know something’s wrong. And how do you get through the tedium part of experience separate from the experience of being wrong?

Michael Mauboussin: What’s the answer, Rob? What is the answer to that? Seriously, I mean, that’s my biggest struggle. For example, I was visiting with a very successful multi-strat firm and they said, “We try to have our analysts cover around 40 stocks, not more than that, and we want them to bump that up with AI to 60 stocks.” I’m not sure exactly I see that path, but I see the logic of it because they’re doing it now. They are in the trenches. They are doing the tedious stuff. And so I could see how AI could help them on the margins to extend their capabilities. But if someone didn’t know what they were doing, I’d be scared to death to allow them to try to use these technology tools. Now again, for gathering information—I just say how do I use it for my own research?

It’s extraordinary. And you can say to it, “Give me the primary research, give me the link to the primary research,” and it’s there or not. So that solves a lot of these things about making up papers. You can just say, “Give me the link to the actual paper.” It’ll do it for you.

Rob Marsh: Right.

Michael Mauboussin: And for that kind of stuff, it’s extraordinary. It’s led me, it’s unsurfaced stuff. I mean, I’m pretty good at Googling things, but I’ve learned a lot. It’s unsurfaced a lot of stuff that I never would have—it’s found documents that I would struggle to find on my own. And again, I’d say I’m probably above average at doing this because it’s a lot of what I do all day. So, it’s been really good for that kind of stuff. So I think this chicken and egg thing is something that we need to think a lot about, especially for young people.

By the way, you mentioned the tedium of doing all these things. The other thing I have to say is that in the olden days, we worked in offices together, and there was a lot of—the claim is something like 70 percent of knowledge is tacit knowledge, which is what you pick up on the job. You and I have a cup of coffee or we’re talking about the Yankees game or whatever it is. There’s information being conveyed that could be very helpful that is not overt. So, that’s another aspect that we’re—you know, I think companies are losing some of that as well, that was probably underestimated in terms of its importance.

Business Edge and Training Considerations

Rob Marsh: I want to bring it back to the topic of edge, because I think it’s actually critical and very important. Some of the other dimensions. So when we’re talking about behavioral, analytical, informational, and technical, that’s largely within the context of investing, trading decisions, buy, sell, hold, sizing, and the like.

If you’re running an investment business, there’s a business edge, a longer-term edge, which is what you’re talking about: the training, the people. And I guess if I were sitting in that seat again and was responsible for a team of analysts and whatnot, I would probably look to kind of flip it a little bit to where they’re responsible for doing the quality control checks on the output and not only saying, “This is good or bad,” and passing it along, but explaining why it’s good or bad. The tedium time is less about tracking down the numbers per se and more about writing a lot—because, again, writing is thinking. So you don’t want to obviate the need for analysts to write. And maybe work it that way. Ideally, instead of getting reps over 30 companies, maybe you do get 60 and you have more meaningful conversations. Or maybe you still cover the same 30, but you just do it in a much more nuanced way, coming at it from different angles.

So the analyst doesn’t have to spend all their time just getting the model populated and having the numbers cross-check. They could actually have time to do some of the inefficient thinking with an AI co-pilot, but also their human boss, whoever they’re reportingto. And I think that might be a way to bridge some of the concerns you have about the learning process while still getting the leverage and efficiency benefits.

Michael Mauboussin: Yeah, I think that’s really smart, Rob. And I think if I were in charge of a group of people doing an analytical process, I always like to say there are certain tools that are very helpful to help people make good decisions. By the way, we should qualify—I want to be clear, there are a lot of different investment approaches under the tent here.

Rob Marsh: Right.

Michael Mauboussin: I teach at Columbia Business School, I teach fundamental analysis, and we believe that the present value of future cash flows dictates stock prices. That’s not the only way to think about the world. There are a lot of other ways.

So I want to be clear about that. But let me just mention two tools that I think are really helpful for people to make good decisions, and I think AI could be extraordinarily helpful in both of these. One is base rates. This is something that Danny Kahneman and Amos Tversky wrote about extensively. It’s something that Phil Tetlock talks about a lot when he talks about superforecasters. It’s something that’s very much part of what we try to do. A base rate is saying instead of me doing analysis on XYZ company based on my information gathering and talking to management and so on, I’m going to simply ask about this company as an instance of a larger reference class.

Rob Marsh: Go on.

Michael Mauboussin: I’m going to say, what happened with other companies like this in the past, and where does my judgment fall within that distribution of outcomes? It’s a little bit of a reality check, often a sobriety check, candidly. But understanding base rates, which can maybe be done much more effectively with AI than we’ve ever been able to do before, I think could be really helpful.

The second is this idea of a premortem. This idea was developed by Gary Klein, but it’s a very popular idea. Danny Kahneman loved it, Dick Thaler loves it. And so it’s this idea of saying before I make an investment, let’s get a group of people together. And by the way, we can get rid of the group part of this by doing this. Let’s get people together, and let’s pretend we made the investment and it turned out to be a fiasco. Each of us then writes down why it was a fiasco. And then we share those and we say, “Okay, are any of these things that we wrote down likely to happen? And if so, maybe we should think twice about this investment.”

Again, you could have the AI do the premortem—again, rather than having anybody do it in a room, you could have the AI do it, as you said, perhaps with these different personas.

Rob Marsh: Michael, if I could jump in real quick. I think there’s an advantage perhaps. It’s not as threatening from a personal perspective. It’s one thing to do a premortem, and if you’re a junior and I’m asking—and I’m sitting in the room with Michael Mauboussin—I might be more reluctant to say anything. So, I think you could get more honest feedback.

Michael Mauboussin: That’s a great point. And by the way, there’s a guy named Roger Martin who’s written a lot about strategy, and he talks about this issue. He says when you’re making a decision at the end of the day. He’s like, “This decision is going to go forward.” He goes, “But”—and he’s like, “People don’t like pessimists.” He goes, “But you’re getting points for being clever about how things can go wrong.” So, it’s almost like you get clever points versus pessimist points. He’s like that reframing is really helpful.

And the second thing is it allows you to create essentially contingency plans. So, you say, “If this bad thing happens, then we will do X.” And so you’re not scrambling when the bad thing happens. You’ve already thought through what you’re going to do. So those are two examples. Do you have about a premortem in action and just how valuable it turned out to be?

A Premortem Success Story

Michael Mauboussin: At an investment firm, I was always involved with the investment committee process. I would sit in all the meetings, and from time to time, I was on the actual investment committee. So this is a particular investment where I was on the investment committee, one of five. Our team came in and pitched an idea, and it was literally three hours of going back and forth, going through their work. And by the way, these guys did a fantastic job. They were really well prepared, they knew their stuff, they had thought through a lot of different angles.

But what happened is that the head of that investment committee, had a particular point of view and was guiding the conversation in a particular direction. It kept coming back to this grooved path that had been set early on by the CIO, and we couldn’t get out of this grooved path.

At the end of it, these poor people have done tons of work, hours and hours. I just said we as an investment committee have to do a premortem. And we did one based on email, and it was absolutely remarkable that two or three things surfaced among our investment committee that I would say were the most important things to think about for that particular investment. And we had not covered them in three hours of discussion. It wasn’t like we were bombing through this thing superficially, and you just don’t cover things that are really relevant. And so that was a real eye-opener for me, that you could spend—it wasn’t like we were spending an hour bombing through something and we were missing—this was hours and hours of detailed discussion without what ended up being two or three really important issues that were under the surface and that required this mechanism to get them to the surface. Probably that would be one very vivid example in my mind.

Rob Marsh: Wow. Yeah. I’ve more than once wished I had gone through that process before making certain decisions.

Systematic vs. Judgment-Based Decisions

Rob Marsh: But I want to go back to something you mentioned earlier about where you can be systematic versus where you need judgment. And I think this is where the rubber meets the road in terms of how AI can be most effectively deployed. You talked about the multi-strat firms and how they’re able to extract alpha through their center books by being systematic about portfolio construction, even though the individual pods are making discretionary stock-picking decisions.

Michael Mauboussin: Yeah, exactly. And I think this gets to the heart of what we should all be thinking about as we contemplate how to use AI effectively. The overarching thing would be that we should be deep in this process of contemplating where we can hand things over and where we can add value, try to demarcate those two areas, and proceed as appropriate.

One example, you and I have talked about this over the years, but one area I find completely fascinating is sizing. I mentioned blackjack before, but when Ed Thorp actually went to Reno to play blackjack, it was a two-part system. Part one was edge through card counting, but part two was bet sizing based on his bankroll and his risk appetite. And they were both systematic. So, the question is, can we be more systematic in terms of sizing than we are today?

And I think the answer is absolutely yes. Most people, even very sophisticated investors, are not very systematic about how they size their positions. They might have some rules of thumb, but they’re not thinking about it in terms of expected value, volatility, correlation with the rest of their portfolio, and so on. That’s an area where I think AI could be extraordinarily helpful in helping people be more systematic.

Rob Marsh: I’m writing down some of these notes and it strikes me that we could have a whole another hour-long conversation just on identifying the features of what’s truly systematic versus what’s judgment.

Michael Mauboussin: I agree with that. I think as humans, we tend to overreach a little bit. One example, many investors love to think that they’re good at timing—market timing or sector timing or whatever it is. And the reality is that most people are not very good at timing. So that might be an area where you say, “You know what, I’m just going to be systematic about this. I’m going to dollar-cost average or I’m going to rebalance on a regular basis,” rather than trying to time things.

Pattern Recognition and AI

Rob Marsh: Well, pattern recognition—maybe our third hour. Because that is an area where again, you talk about—you very much have a deep fundamental perspective. I come from the macro world where my big opportunity was building pattern-based trading systems and so was able to have a career based on some variations of that. But…

Michael Mauboussin: But that’s the thing I would just want to underscore for people. I want to be clear that I am not anti-pattern recognition. I am anti-thinking you know where there’s a pattern where there is no pattern. So what you did, Rob, is you found legitimate patterns that allowed you to make money. That’s fantastic. The problem is that humans are pattern-seeking machines, and we see patterns everywhere, including where they don’t exist.

Rob Marsh: I might suggest my former boss’s success is more a testament than mine. But you raised a very good point. It’s not about choosing sides. These are not political or religious debates. And I think one of your profound gifts you’ve provided through your work over the years is providing the structure and the frameworks to identify when things are appropriate, where they should work, where they shouldn’t, and why not.

Michael Mauboussin: Yeah.

Rob Marsh: It’s kind of the first principles around all of that which is so important. So I think there’s a lot of work to do. But the one area where I feel most limited is on sizing. I actually think that’s an area where there’s a lot of opportunity. You mentioned before the center books and how there’s an alpha capture at these multi-strategy firms where they can figure out how to take the best ideas and put them together, and they obviously use leverage in that mix as well.

Michael Mauboussin: I just think there’s an enormous opportunity for a lot of investors to be smarter about how they monetize the edge that they find. And I look at sizing as being probably the area where there’s the most opportunity for most people to improve. You could imagine building a systematic approach to sizing that takes into account your conviction level, the expected volatility of the position, how it correlates with the rest of your portfolio, your overall risk budget, and so on. And then you could just compare what you’re doing to what the ideal would be. And then over time, if it works out well, you can obviously monitor the performance of both those portfolios, and to the degree to which the systematic portfolio does better, you can migrate toward it.

The Chess Program Analogy

I’m reminded, there’s this great story where they were looking at generations of chess players. And Magnus Carlsen obviously is one of the greatest of all time, and they were able to document, by running chess programs simultaneously with watching the players’ moves, those players that were trained with chess programs versus those that were not. And what they found is that the players who were trained with chess programs made moves that were much more similar to what the chess program would recommend, comparing how they move vis-a-vis the chess program.

So what would that be for us? How do we develop our version of the chess program that allows us to learn the right moves given the situation and then learn from that over time to get better and better? So to me, the magic wand would be the flight simulator, the chess program that would allow us to learn as we did our jobs.

Conclusion

Rob Marsh: Well, we’re running up against our time limit, but this has been absolutely fascinating. I think we’ve covered a lot of ground on how AI can enhance the investment process while being mindful of the challenges and limitations. The key seems to be finding that balance between leveraging AI’s capabilities and maintaining the human judgment and experience that’s so critical to successful investing.

Michael, thank you so much for taking the time to share your insights. This is going to be super exciting. We’ll see how all this unfolds. So, I really appreciate you taking the time and your thoughtful questions and comments.

Michael Mauboussin: Rob, this was fantastic. Thank you so much for having me on. These are such important topics, and I think we’re just at the beginning of figuring out how to use these tools effectively. It’s been a real pleasure.

Common Behavioral Biases in Investing and Mauboussin's AI-Aided Countermeasures

For more information on Michael and his writings, you are encouraged to visit is website: https://www.michaelmauboussin.com/

Share

Disclaimer: The information contained in this newsletter is intended for educational purposes only and should not be construed as financial advice. Please consult with a qualified financial advisor before making any investment decisions. Additionally, please note that we at AInvestor may or may not have a position in any of the companies mentioned herein. This is not a recommendation to buy or sell any security. The information contained herein is presented in good faith on a best efforts basis.

Q1: What is an investment "edge"?

An investment edge is having a belief about an asset that is different from what the market is currently pricing, and having a positive expectation that your belief will prove correct and the market will eventually come to agree with your view. It's a disciplined process of identifying what you know that the market doesn't, which gives your investment a positive expected value. Given that markets are mostly efficient and luck plays a huge role, systematically identifying and exercising a true skill-based edge is critical for an active manager.

Q2: What is the BAIT framework for investment edge?

BAIT is an acronym Michael Mauboussin uses to categorize the four primary sources of an investment edge:

• B - Behavioral: Exploiting the predictable emotional and cognitive biases of other market participants. This includes things like overextrapolation of trends or reacting to sentiment extremes.

• A - Analytical: Having the same information as others but analyzing it with a higher degree of skill. This is like a professional tennis player competing against an amateur with the same equipment.

• I - Informational: Possessing better or more timely information than others, obtained legally. This can come from uncovering complex relationships (e.g., in supply chains) or simply paying attention to publicly available information that the market is currently ignoring.

• T - Technical: Capitalizing on situations where other market participants are forced to buy or sell for non-fundamental reasons, such as fund flows, margin calls, or regulatory constraints, allowing you to act as a liquidity provider.

Q3: How can AI help investors apply the BAIT framework?

AI can enhance each component of the BAIT framework:

• Behavioral: AI can process vast amounts of text from news and social media to measure sentiment, helping to identify extremes of optimism or pessimism that often lead to mispricings.

• Analytical: AI can apply base rates to a company's forecasts to check for overextrapolation and provide a more objective, probabilistic assessment. It can also synthesize huge datasets to help analysts process information more effectively.

• Informational: AI can quickly ingest and interpret alternative data sources (e.g., satellite imagery, web scraping) to surface signals before they are widely recognized. It's also incredibly powerful for documenting an investor's own thought process, which is a form of informational edge.

• Technical: AI can monitor market flows, positioning, and liquidity in real time to detect imbalances that might signal forced selling or buying from other participants.

Q4: What is the "noise" problem in investing and how can AI help?

The "noise" problem refers to the high degree of variability and randomness in human judgment. For example, if you give the same case to 50 different analysts at the same firm, you will get wildly different valuations. This inconsistency is "noise." AI can help mitigate this by simulating a "wisdom of crowds" cheaply and efficiently. You can create different AI agent personas (e.g., "Warren Buffett," "Seth Klarman") to analyze an idea from multiple, diverse perspectives. This quickly surfaces counterarguments and reduces the randomness of a single analyst's view.

Q5: Why is documenting investment decisions so important?

Documenting the "why" behind an investment decision at the time it's made is crucial for learning and improving. It creates an objective record that can be reviewed later to see if you were right for the right or wrong reasons. Most investors avoid it because it can be embarrassing to be wrong, but it's an invaluable tool for self-improvement. AI can make this process easier (e.g., via voice notes) and can later analyze these records to provide honest, unbiased feedback on your decision-making patterns, highlighting what works and what doesn't.

Q6: What is the biggest learning challenge AI creates for new analysts?

The biggest challenge is the "chicken-and-the-egg" problem. To effectively use AI tools and judge the quality of their output, you need a foundational level of pre-existing knowledge. An experienced analyst can spot a flawed AI-generated analysis because they have been "in the trenches" and built models themselves. There is a concern that junior analysts might use AI to get answers without going through the tedious but essential process of learning the fundamentals, leading to high scores on problem sets but poor performance when true judgment is required.

Q7: How can tools like base rates and premortems improve decision-making?

• Base Rates: This involves looking at a current situation as an instance of a larger reference class. Instead of only analyzing a company from the "inside view" (its specific story), you ask what happened to other, similar companies in the past. This provides an "outside view" that serves as a powerful reality check on forecasts. AI can be used to quickly gather and analyze the vast amounts of data needed to establish accurate base rates.

• Premortems: This is an exercise where, before making a final decision, the team imagines that the investment has failed spectacularly. Each member then writes down the reasons for the failure. This process helps uncover risks and hidden assumptions that may have been missed in the initial analysis. AI can facilitate this by running premortems with different agent personas, which can be less threatening for junior team members and lead to more honest feedback.

Q8: How can AI help investors with position sizing?

Position sizing is one of the biggest opportunities for improvement for most investors. Many rely on heuristics rather than a systematic process. AI can help create a more systematic approach by recommending position sizes based on a combination of factors, including an investment's expected value, its volatility, its correlation with the rest of the portfolio, and the investor's overall risk budget. This can act as a "co-pilot" or a "chess program" that helps the investor learn and migrate toward a more optimal way of monetizing their edge over time.

Q9: What is the future outlook for AI's role in the investment industry?

Michael Mauboussin believes AI will fundamentally reshape investing within the next 3 to 20 years. It will act as a bridge between quantitative and discretionary investing, blending systematic rigor with human judgment. AI will free analysts to focus on higher-impact decisions while functioning as a "co-pilot" to help investors learn, refine processes, and improve over time—similar to how chess engines have trained human players to excel.

Share

I. Central Concept: The Investment Edge

A. Definition

• A belief different from the market price with a positive expected value

• An act of "extraordinary hubris" for active managers

• Discerning skill from luck is the primary challenge

B. Core Question

• "Why do I think I know something the market doesn't?"

C. Analogy

• Card counting in Blackjack (Ed Thorp) – turning a house edge into a player edge

II. The BAIT Framework for Identifying Edge

A. (B) Behavioral: Exploiting Predictable Human Biases

• Source: Overextrapolation, sentiment extremes (fear/greed), herd mentality

• AI Application: Sentiment analysis on news/social media, identifying behavioral mispricings

B. (A) Analytical: Processing the Same Information More Skillfully

• Source: Superior models, unique interpretation, better weighting of variables

• AI Application: Synthesizing vast datasets, applying base rates, identifying hidden variables

C. (I) Informational: Having Better, Unique, or Overlooked Information (Legally)

• Source: Analyzing complexity (e.g., supply chains), using ignored public data

• AI Application: Ingesting alternative data (satellite, web scraping), documenting one’s decision process to create a unique internal dataset

D. (T) Technical: Capitalizing on Forced Transactions

• Source: Fund flows, margin calls, index rebalancing, regulatory constraints

• AI Application: Monitoring market flows, positioning, and liquidity in real time

III. AI's Role in the Investment Process

A. The "Glue": Bridging Quantitative and Discretionary Investing

• Takes the "greatest hits" from both camps

• Analogy: The "center book" at a multi-strategy firm systematically extracts alpha from discretionary pods

B. Mitigating Human Flaws (Defense)

• The "Noise" Problem: Reducing randomness in human judgment

• Solution: Simulate a "wisdom of crowds" via AI personas (e.g., "Warren Buffett", "Seth Klarman")

• Auditing the Process:

  • Problem: Investors experience "slippage" and deviate from their best process

  • Solution: AI codifies the ideal process and audits real-world decisions against it

C. Supercharging the Process (Offense)

• Position Sizing:

  • Problem: Most investors aren’t systematic

  • Solution: AI acts as a "co-pilot" recommending optimal sizing based on EV, volatility, correlation

• Idea Generation & Analysis:

  • Surfacing overlooked documents and research

  • Running base rate analysis and premortems efficiently

IV. Improving the Decision-Making Toolkit

A. Decision Documentation

• Importance: Crucial for learning and feedback

• AI’s Role: Enables voice-to-text capture and pattern analysis

B. Base Rates (The "Outside View")

• Process: Compare investment to a larger reference class

• AI’s Role: Rapidly gathers and analyzes data to generate base rates

C. Premortems

• Process: Imagine failure and diagnose causes

• AI’s Role: Use personas to facilitate honest, low-threat exploration of potential risks

V. The Human Element: Challenges & Training

A. The "Chicken-and-the-Egg" Learning Problem

• Dilemma: Quality of AI outputs is hard to judge without experience

• Risk: Junior analysts may become over-reliant on AI

B. The Future of Analyst Training

• Transition from data gathering to quality control

• Teach analysts to evaluate AI output rigorously

• Use AI to:

  • Cover more names (e.g., 60 stocks vs. 40)

  • Go deeper on existing coverage

VI. Key Learnings for AI Adoption (Michael Mauboussin's Advice)

A. Enhance Your Decision-Making Process

• Use AI to audit your process and identify "slippage"

• Voice-document decisions and analyze for bias/winning patterns

• Conduct AI-driven premortems with agent personas

B. Improve Your Analytical Toolkit

• Let AI act as a co-pilot for position sizing

• Rapidly establish base rates to ground forecasts

• Combat judgment "noise" using agent personas

C. Rethink Team Structure and Training

• Treat AI as glue between quant and discretionary teams

• Solve the learning dilemma with foundational training

• Emphasize critical assessment of AI-generated outputs

  ---

Share AInvestor

I. AI Investor Dialogues: Context and Mission

• Purpose: A series focused on enhancing investor returns through AI.

• Format: Features conversations with top fund managers and "builders" like Justin Whitehead. Rob Marsh introduces the series and conducts the dialogue.

• Goal: To create content that serves as a public good to the investor community by exploring AI approaches from ground-level processes to higher-level strategies to enhance or generate investor returns.

II. Pebble Finance: AI as a Transformative Tool in Investment

A. Founders and Background

• Co-founded by Justin Whitehead (CEO) and James Esdaile.

• Justin Whitehead's extensive background in fintech and AI:

  • Early intern at Factset during the dot-com heyday.

  • Built the portfolio analysis (PA) product at Factset.

  • Joined Kensho Technologies, shifting focus from buy-side to sell-side.

  • At Kensho, worked on teaching machines to distill events from the news for market analysis; served as CTO.

  • Self-identifies as a "builder" and a "nerd", remaining a hands-on keyboard engineer despite being CEO.

• Pebble's Aim: To use two decades of experience to make sophisticated financial analysis understandable and accessible.

B. Problem Pebble Finance Aims to Solve

• Democratize financial analysis for retail investors, wealth advisors, and institutional managers.

• Retail investing is often haphazard, lacking structure or market factor awareness.

• Aims to improve returns and trust by providing understandable, informed insights, especially in volatile markets.

• Provides tools for informed decision-making, combating emotional and uninformed reactions.

• Helps investors avoid poor decisions when feeling "blind" or reactive during downturns.

C. Core Technology: The "Explanation Engine"

• Analyzes news, research, and market data to find catalysts behind asset movements.

• Goes beyond aggregation; inspired by Kensho’s "market journal".

• Helps users understand why an asset is moving.

• Uses statistical tools to find non-obvious connections between companies:

  • Example: Biotech stock moves due to related company FDA approval.

  • Example: Spirit Aerosystems stock drops due to 737 Max issues.

• Produces 10,000–15,000 explanations per day, revised dynamically as new data emerges.

D. How AI is Utilized within Pebble Finance's Technology

• Uses NLP, generative AI, and traditional programming; AI is viewed as "just math underneath the hood".

• Applications:

  • Data extraction: From SEC filings and presentations using small, focused LLMs.

  • News processing: NLP used to cluster and summarize news; integrates direct data feeds.

  • Explanation generation: Tailored by audience type and level of sophistication.

  • Quality control: Combines human-in-the-loop vetting with machine checks to prevent inaccuracies—crucial in regulated sectors.

E. Business Model and Client Delivery

• Operates on a B2B model, primarily delivered via APIs.

• Clients can use cloud services or run the engine on their own infrastructure.

• Embeds insights into existing platforms.

Value by segment:

• Retail Brokers:

  • Boosts trust and engagement.

  • Offers new revenue streams (e.g., $15–$20/month subscriptions).

• Wealth Advisors:

  • Enables fast, informed client updates.

  • Offers timely insights for large client rosters.

• Institutional Managers:

  • Improves portfolio monitoring.

  • Integrates into internal systems for instant analysis.

  • Used by Factset for AI-generated commentary.

F. Stance on the "Build vs. Buy" Dilemma

• Industry prefers in-house solutions (influenced by players like Bloomberg).

• Pebble argues prototyping is easy, but production-level AI (fast, compliant, accurate) is hard and costly.

• Offers speed, compliance, and ongoing innovation.

• Emphasizes clients will continue building, but Pebble "moves the goalposts" faster and better.

G. Mitigating Technological Obsolescence Risk

• Regulatory complexity deters big tech (e.g., OpenAI, Google) from entering finance directly.

• Pebble focuses on specialized, high-speed, high-efficiency AI.

• Offers cost-effective solutions—e.g., cutting clients' LLM costs in half.

• Small, focused approach seen as a strategic edge.

H. Vision for the Future of the Financial Industry and Pebble's Role

• Predicts major transformation in financial services over the next decade.

• AI will allow firms to do more with less, reshaping staffing and workflows.

• Traditional research platforms will evolve with private LLMs and external AI integrations.

• Retail investing transformation:

  • Large underserved market (~70% of investors).

  • Subscriptions for automated investment understanding will disrupt current wealth management models.

  • The fight for this market has already begun.

• Pebble aims to be the "guy with the gasoline cans", accelerating this shift.

III. Key Learnings for AI Adoption in Investment (Justin Whitehead's Advice)

A. Experiment Personally and Privately

• Start experimenting individually to build comfort.

• Don’t try to overhaul investment practices immediately.

• Personal exploration leads to idea generation and time-saving discoveries.

• AI won’t replace engineers, but it will become a powerful tool—like hitting "return" on a keyboard in the ’80s.

B. Find Micro-Problems for AI to Solve

• Focus on small, targeted use cases.

• Examples: Researching clients, summarizing info, creating prep materials.

• Best used in non-engineering contexts for rapid value.

C. Learn to Trust and Verify AI Outputs

• Understand both strengths and limitations of AI.

• Always verify outputs, especially in regulated fields.

• Learn to spot when AI is producing faulty results.

  Share

  Thanks for reading AInvestor! Subscribe for free to receive new posts and support my work.

Q1: What is Pebble Finance and what problem does it aim to solve in the investment industry?

Pebble Finance, co-founded by Justin Whitehead and James, aims to democratize sophisticated financial analysis, traditionally available only to institutional investors, by making it accessible to retail investors, wealth advisors, and institutional asset managers. It addresses the "haphazard" nature of retail investing, which often lacks structured processes for diversification and understanding underlying market factors. Pebble uses AI and machine learning to analyze vast amounts of data and provide understandable insights, helping investors make more informed decisions, especially during volatile market conditions, thereby improving returns and building trust.

Q2: How does Pebble Finance's "explanation engine" work and what makes it unique?

Pebble's explanation engine is a core technology that analyzes news, research, and market data to identify "catalysts" driving asset movements. Inspired by the "market journal" experiment at Kensho, it goes beyond simple news aggregation. The engine uses statistical analysis and an understanding of connections between companies to highlight significant movements and their potential drivers, even seemingly unrelated events (referred to as "weed day" movements where a stock moves due to news about a related, but not directly obvious, company). This allows users to understand why an asset in their portfolio or watchlist is moving, fostering informed decision-making.

Q3: How is AI utilized within Pebble Finance's technology?

Pebble Finance employs a range of AI and machine learning techniques, including natural language processing (NLP), generative AI (like LLMs), and traditional programming. These technologies are used to:

• Build proprietary data sets: Extracting structured information from unstructured data (e.g., SEC filings, investor presentations) to create a knowledge graph.

• Process news and research: Clustering and pre-summarizing vast amounts of news and research data.

• Generate explanations: Tailoring explanations to different audiences (retail investors vs. institutional asset managers) with appropriate vernacular and depth.

• Ensure accuracy: A critical "human-in-the-loop" process, combined with machine vetting, is used to fact-check and prevent hallucinations, ensuring the quality and accuracy of generated explanations, especially crucial in the regulated retail brokerage space.

Q4: How does Pebble Finance deliver its capabilities to clients, and what is its business model?

Pebble Finance operates on a B2B model, primarily delivering its technology via APIs. This allows for seamless integration into existing platforms used by retail brokers, wealth management platforms, and institutional asset management tools. Clients can either use Pebble's cloud-based services or run a copy of Pebble on their own infrastructure, connecting Pebble's engine to their data. The goal is to embed Pebble's insights where users already are, rather than requiring them to adopt a new platform. The value proposition varies across segments: for retail brokers, it enhances client engagement and trust; for wealth advisors, it enables more informed and timely client updates; and for asset managers, it aids in efficient portfolio monitoring and global market awareness.

Q5: What is Pebble Finance's stance on the "build vs. buy" dilemma in the financial industry?

Justin Whitehead acknowledges the strong tendency in the financial industry to build technology in-house, historically influenced by major players like Bloomberg. However, he emphasizes that while prototyping AI solutions might seem easy, scaling for production, ensuring speed (especially sub-second latency, which is challenging for generative AI), and addressing legal and compliance risks are significant hurdles. Pebble's advantage lies in its focused expertise and continuous innovation, offering clients faster and more cost-effective solutions. He sees Pebble as a partner that can help clients get to market quicker and cheaper, while continuously advancing its capabilities to maintain a "defensible moat."

Q6: How does Pebble Finance address the risk of technological obsolescence, especially with the rapid advancements in AI?

Justin Whitehead believes Pebble is protected from obsolescence due to several factors:

• Regulatory complexity: Large general AI initiatives (like OpenAI, Google Gemini) are unlikely to directly enter the heavily regulated financial domain due to the complexities of dealing with bodies like the SEC and FINRA.

• Specialized focus: Pebble's goal is to provide specific, high-quality explanations for the financial domain, optimizing for speed, scale, operational efficiency, and security. This often means moving in the opposite direction of general-purpose LLMs, which aim for broad capabilities.

• Cost-effectiveness: By optimizing for financial domain needs, Pebble can offer more cost-effective and reliable solutions compared to general LLMs, which aligns better with client needs in a regulated environment. He envisions being able to cut clients' LLM bills in half while maintaining high margins due to specialized efficiency.

Q7: What is Justin Whitehead's vision for the future of the financial industry and Pebble's role in it?

Justin Whitehead foresees a significant transformation in financial services over the next decade. He believes:

• More with less: AI will enable financial operations to do "more with less," eventually impacting staffing needs as AI handles grunt work, monitoring, and even tasks like building DCF models and tracking earnings.

• Shift in research platforms: Traditional dashboard-based SAS platforms for research and asset management will evolve. Large institutions will likely adopt private LLM licenses, and companies like Pebble will provide specialized capabilities to augment these internal AI systems.

• Reshaping wealth management: The lower tier of wealth management, currently underserved (70% of investors), presents a huge opportunity for retail brokerages. Pebble aims to be a key player in this shift, enabling brokerages to offer premium, Netflix-style subscription experiences providing comfort and automated understanding of investments at an accessible price point, potentially disrupting traditional wealth models.

Q8: What advice does Justin Whitehead offer to individuals regarding the adoption of AI in their investment process?

Justin Whitehead emphasizes that while AI may not replace professionals, it is incredibly useful. His key advice is to:

• Experiment personally: Start experimenting with AI tools, even in private, to become comfortable with the technology. Don't immediately try to change professional practices.

• Find micro-problems: Identify small, specific tasks where AI can assist, such as quickly gathering information, researching clients, or acting as a "sounding board."

• Learn to trust and verify: Understand AI's capabilities and limitations, learn when to trust its outputs, and always verify information, especially by citing sources.

He believes that personal experimentation will spark ideas for practical applications and highlight AI's potential to save significant time and enhance decision-making.

Share

Justin Whitehead is the co-founder and CEO of Pebble Finance, a pioneering fintech-AI platform that helps investors make sense of portfolio performance with intuitive, data-driven insights. With over two decades at the nexus of finance and technology, Justin previously led development of FactSet’s industry-leading portfolio analysis tool, steered engineering at AI pioneer Kensho, and held technology leadership roles at BotKeeper before founding Pebble in 2021. Based in Cambridge, Massachusetts, Justin combines deep software development and product expertise with a clear vision: to simplify complex investment analytics for professionals and their clients alike.

Justin and I first met while we were both at Kensho, the AI-centric fintech I’ll refer to all too often in these interviews. It was an intense twenty-four months that I measure in dog years. We agreed early on to never hold back and never be offended. It was a formula that, fortunately, worked—as we often had different ideas about what was most important at the moment. Don’t read too much into the conflict, though, as the north stars for product and technology are rarely perfectly aligned.

Anyway, let’s get to it. Please enjoy my conversation with Justin.

Click here for FAQs and a Mind Map summaries of the key concepts from our interview with Justin.

In this interview, you’ll learn:

1. Why most retail investors are flying blind—and how AI can act as a portfolio co-pilot

2. The secret to scaling personalized investment insights across 10,000+ securities daily

3. Why explanations, not predictions, are the killer feature in AI-powered investing

4. How Pebble helps retail brokers monetize advice in a post-commission world

5. What wealth advisers really need to shine—and how Pebble arms them instantly

6. Why “build vs. buy” in financial AI isn’t a binary—it’s a dance

7. How small, focused LLMs outperform general models when speed, accuracy, and auditability matter

8. What happens when you embed AI in every layer of the investment stack

9. Why Justin believes the wealth management model of 2025 won’t survive until 2035

10. How Pebble is designed to be invisible, embedded, and indispensable

Some takeaways:

1. Retail investing is broken—and AI can help. Most retail investors lack a process. They chase brands like Apple or Tesla without understanding how to diversify or analyze exposure. Pebble uses AI to turn vague themes like “I want to invest in EVs” into full-spectrum portfolios mapped to supply chains, regulatory catalysts, and competitors.

2. The explanation engine is Pebble’s core differentiator. Forget predictions—Pebble focuses on helping investors understand why something is moving. Its engine identifies catalysts by scanning news, filings, correlations, and graph-based relationships. A biotech stock in San Diego jumps 30%? Pebble can trace it to an FDA approval for a peer in Boston.

3. Pebble produces 10k–15k explanations a day—with compliance baked in. Unlike static articles, explanations update dynamically as new information comes in. Human-in-the-loop QA ensures every piece meets regulatory standards—critical in partnerships with brokerages and institutions.

4. Every firm is building something—but Pebble moves the goalposts. Justin knows the instinct in finance is to “build first, buy later.” That’s fine—until compliance, speed, and scale catch up. Pebble lets teams prototype in-house, then plug into battle-tested infrastructure when it’s time to ship at production scale.

5. Retail brokers need new revenue streams. With commissions gone, the race is on to build Netflix-style subscriptions for investing. Pebble helps platforms deliver differentiated advice and context, creating retention and monetization opportunities at scale.

6. Wealth advisers need fast, clear answers. Whether it’s a tariff announcement or earnings surprise, advisers can instantly explain how a portfolio is affected. Pebble translates market complexity into client-ready insights—no spreadsheets required.

7. Institutional firms are stretched thinner than ever. With fee compression and geographic mandates expanding, asset managers need AI as “boots on the ground.” Pebble enables them to monitor portfolios, scan news, and detect movement—even across continents—without adding headcount.

8. The real AI challenge isn’t language—it’s infrastructure. Pebble combines small LLMs with traditional ML and human QA to achieve speed, precision, and auditability. Sub-second responses? Compliance-approved commentary? Scalable across clients? That’s not easy with an off-the-shelf model.

9. The future of wealth is automated, augmented, and embedded. Justin sees wealth management bifurcating—high-touch for the ultra-wealthy, and AI-augmented advice for everyone else. Pebble aims to power that bottom 70% of the market—at scale, invisibly, and affordably.

10. Don’t build another dashboard—build the insight layer. Pebble isn’t trying to own the front end. It wants to be the intelligence embedded inside your stack—whether you’re a broker, RIA, or asset manager. Pebble doesn’t care who gets the credit—it just wants to be the engine behind the best advice in finance.

    Thanks for reading AInvestor! Subscribe for free to receive new posts and support my work.

Introduction

Rob Marsh: Welcome to AI Investor Dialogues. I’m Rob Marsh. Today we’re diving deep into the practical application of AI in investment processes with Justin Whitehead, co-founder and CEO of Pebble Finance. Justin brings over two decades of experience building technology for Wall Street, from his early days at FactSet to his role as CTO at Kensho Technologies.

We’ll explore how Pebble is democratizing sophisticated investment analysis, the reality of implementing AI in regulated financial services, and Justin’s bold predictions for the future of wealth management. Whether you’re a retail investor, wealth advisor, or institutional asset manager, this conversation offers actionable insights on leveraging AI to enhance investment outcomes.

Justin and I had the good fortune to work together at Kensho, where I oversaw product and Justin was, as noted, CTO. We had many vibrant conversations where he managed to keep me in line most of the time

Justin Whitehead: “Vibrant”—that’s an operative word right there.

Rob Marsh: So Justin, let’s start by defining your investment process at Pebble, then walk through what makes a good process and how AI plays a role. We’ll touch on what you’re doing at Pebble and discuss AI as an asset class. To kick it off, give us a brief background on your experience with AI and why you’re such a good person to be talking about this.

Justin’s Background: Building at the Intersection of Finance and Technology

Justin Whitehead: I’m a builder and a nerd through and through. I’ve been operating at the intersection of finance and technology before the term “fintech” was ever coined. I got my start as an early intern at Factset during the dot-com heyday, then joined full-time just as the bubble burst. I was one of the fortunate few computer scientists graduating who still had a job.

My early experiences shaped both my leadership style and understanding of markets while building a product called Portfolio Analysis. If you’re an institutional asset manager anywhere in the world today, there’s a high likelihood you’re either using FactSet PA or have used it at some point in your career.

After a long career at Factset, I joined Kensho, focusing on the sell side of Wall Street instead of the buy side. The name of the game flipped from teaching machines to understand portfolios and explain performance to asset managers, to teaching machines to distill events from news and quantitatively study them.

Today I’m co-founder and CEO of Pebble. Don’t let that fool you—I’m still very much a hands-on engineer. What my co-founder James Esdaile and I started was taking those two decades of experience and turning them into something understandable for retail investors. We wanted to make powerful analyses accessible to individual retail investors, eventually expanding to wealth advisors and institutional asset managers.

This is where AI comes in. I’m a firm believer that AI is just a two-letter word. It opens doors, but nobody has an AI problem. Advancements in technology, including machine learning and AI, have given teams like Pebble more tools to break boundaries between what was previously accessible to certain demographics and make it available to the masses.

Understanding Investment Processes: From Institutional to Retail

Rob Marsh: What does “investment process” mean to you, and how does it play into the key challenges Pebble Finance is addressing?

Justin Whitehead: I’ll take the investment process from two angles. First, as an engineer who built technologies for institutional asset managers. When you say investment process, I think of FactSet Portfolio Analysis and factor attribution—basically getting the math to mirror the asset manager’s investment process.

These managers were all benchmark-relative. You start by roughly copying the benchmark, then make allocation decisions. “I think tech is going to be great, healthcare is going to take it on the chin,” and you adjust your weights. Then you dive into each sector making active decisions that differentiate your portfolio from the benchmark.

The fascinating thing is when I went to the other end of the spectrum and started interacting with retail investors—guess what? There’s no formal process. You stumble into a portfolio. There are lots of brand-name companies where people think, “I know what Apple is, so I’m going to buy Apple stock,” without thinking about semiconductors, consumer technology, or Apple’s supply chain.

Early on, people thought Tesla and electric vehicles were synonymous. But there’s much more than just Tesla in the EV market. If you wanted to invest in electric vehicles, what might be advantageous is taking a more diversified approach.

A lot of Pebble’s early work was taking retail insights—when someone said they wanted to invest in EVs—and teaching a machine to do research on that topic. We collected our own datasets to understand what types of public companies produce what products, how they’re categorized, how they’re related, and who’s in whose supply chain. We did this by teaching machines to read public filings and investor presentations.

Generative AI gave us more techniques to turn a natural language statement into a query we could run on our graph. While it’s still not the institutional investment process, we focused on understanding the retail investment process, realizing it’s more haphazard, then creating technological ways to bake in diversification by default.

The Problem with Lack of Process

Rob Marsh: I hear what you’re saying about retail. The lack of good process is a barrier to returns. In relationships, ignorance may be bliss, but it’s not when it comes to investing. I think of all the calls I’ve taken from friends and family during major market dislocations when people are concerned and feel blind. That’s when people tend to make poor decisions.

Justin Whitehead: We are all genetically wired to be horrible investors.

Rob Marsh: We are. Being able to be informed is crucial. If you’re going to buy Amazon, you could tell someone it’s been a great investment over 25 years, but you had to sit through three to five 80-plus percent drawdowns along the way. It’s fine to say, “I get it, I understand, here’s the next Amazon, I’ll be able to do it.” No, you won’t—at least not if you don’t understand that these things happen, why they happen, and why something happening today connects to historical patterns.

Having a process or tool that helps you make those decisions is critical. The right decision may be not doing anything—in fact, most likely will be not doing anything.

Pebble’s Explanation Engine: Connecting the Dots

Justin Whitehead: One of our most successful technologies has been our explanation engine. Give me a portfolio, ETF, mutual fund, model portfolio, or stock. How do you look at that asset? Read the news, read available research, look at market data, understand its correlations and connections to other market aspects.

We call it identifying the catalysts of movement. Sometimes the catalyst isn’t in the public eye yet. When we were working on Market Journal—the grandfather of Pebble Explain—we took your classic real-time portfolio dashboard where you’re managing 100 stocks.

There’s no way any individual can stay on top of all 100 stocks equally. If you’re using a Bloomberg or FactSet terminal, it’s blinking green and red with news flying by. You know if your screen is all red, it’s a bad day; if it’s all green, it’s a good day. But most days, it’s a mix.

What we did brilliantly with Market Journal was help a person look at 100 assets and instantly understand where they need to look—not what is moving, but where they need to look now.

We had real-time quotes coming in and were statistically analyzing securities every 15 seconds, comparing them to cohorts. Was this a statistically significant day for the company itself—a “me day”? Was this just moving with the market—if the S&P 500 is down three standard deviations, everyone’s going out with the tide—a “they day”?

The most interesting was the “we day”—something in between. We understand to a quantitative degree the connections between companies, both structured quantitative and unstructured context. You might have some random company in San Diego researching CRISPR therapeutics for cancer, and dozens of these micro-cap companies scattered around the US all researching the same thing.

When one company gets FDA approval, there’ll be news about that company but not about the other 12. Did they react? If you have this random San Diego biotech company up 30% on no news, you’re automatically asking why. In our Market Journal dashboard, we’d say this looks like a “we day” because we see this cohort of six degrees of Kevin Bacon all rising. If you look across that cohort, you’re talking about a San Diego stock, but in Boston, there’s another stock that just got FDA approval, and the math tells us that’s probably what’s causing it.

Rob Marsh: I like to call that “catching strays.”

Justin Whitehead: Exactly. When we showed that to asset managers, it knocked things out of the park. We could sort that ticker list by who’s got the most interesting “we days.” There was always something fascinating every day in an arbitrarily large universe.

I remember during the 737 Max debacle, Spirit AeroSystems was down on no news. I initially thought it was Spirit Airlines, but when I clicked on it, I realized it was Spirit AeroSystems. Why would they be down? Because they make the fuselages for the 737 Max, and it all clicked together.

It was wildly fascinating to see those connections emerge in the real world. How that’s used today in Pebble—we call it the catalyst engine. Consumers using the Pebble engine see news being connected from other securities outside the cohort. This is how we use math to connect dots that aren’t otherwise readily apparent.

How Clients Access Pebble’s Capabilities

Rob Marsh: How are clients and investors accessing that capability? How much is browser-based versus APIs?

Justin Whitehead: Everything is API underneath the hood. The team at Pebble consists of people with my type of background—10 to 20-plus year veterans at fintech startups and AI. We’re heavily focused on the backend, so our surface area is our API.

Clients use us in two ways: either what we have in the cloud where we’ve connected data and market data and you can use our APIs there, or you can run a copy of Pebble on your premises—your own bare metal or in your cloud. We connect our engine to your data, and everything stays internal.

How users interact with it depends on what category of user you are. We sell to retail brokers for retail investors’ benefit, wealth management platforms for wealth advisors’ benefit, and institutional asset management platforms for asset managers’ benefit.

For example, FactSet’s institutional asset managers have access to generate portfolio AI commentary, and that’s partly Pebble underneath the hood. FactSet also sells APIs for this. Many financial institutions are exploring chatbots—starting with consuming the intranet for HR policies, then evolving to incorporate financial data.

My goal is to go where users already are. I don’t want Pebble to build yet another platform for people to log into. I’m more interested in focusing on answering certain classes of questions at incredible depth, scale, and speed.

Business Models and Market Opportunities

Rob Marsh: Walk us through the different client types you are serving. How big is the opportunity?

Justin Whitehead: There’s a pretty big addressable market across different industries. You’ve got the same engine being leveraged in completely different contexts. Serving retail brokers with millions or tens of millions of clients is one kind of scale with our revenue model around that. Then there are asset managers and hedge funds—there aren’t 180 million hedge funds, but these are professionals using the same explainer engine in a different context with completely different economics.

I love that it’s the same technology serving complementary categories.

Rob Marsh: Would it be fair to say this is ultimately B2B2C? What you’re really providing is trust through engagement and knowledge—retail investors engaging with their clients through relevant and credible information?

Justin Whitehead: In retail brokerage, brokers are trying to create a premium investing experience. Right now they’re trying to create differentiated and better retail investing experiences because if you do that, you can attract more clients who will stick around, ultimately fueling where revenue is made later in wealth.

Everyone got “Robin Hooded” five to ten years ago—commissions are down, big loss of revenue stream, everyone has to have a slick app. Everyone’s looking to figure out: “I can’t get commissions anymore, but can I get a Netflix-sized subscription for creating a better investing experience?” I think that’s actually an interesting model, and you’re seeing some innovators already generating significant revenues from it.

In wealth, this is all about helping that advisor shine. If you have a wealth advisor, to you they’re your hero, your sherpa into capital markets. But on the flip side, you’re probably one of 30, 100, or 150 client relationships going the other way. That wealth advisor isn’t sitting there actively managing and monitoring your portfolio unless you have your own private family office.

There’s this disconnect between what you pay an AUM fee for and what you get. What we do is help that wealth advisor. Click of a button: “This is what’s going on in the portfolio. This is exactly what happened over the past week. Trump just announced tariffs—this is how much money it cost you.”

By enabling that advisor to have conversations grounded in the same news their clients are seeing, it brings this notion of “this person working for me is on top of my portfolio, connecting dots with things I’m reading and concerned about.”

In asset management, it’s about portfolio monitoring. There was an asset manager in Canada expanding their mandate into Europe without boots on the ground, asking “How do we stay on top of that geography?” There’s fee compression on the institutional asset manager side and wealth side—people have to do more with less. There absolutely is a need to stay on top of markets where you don’t have boots on the ground and allow machines to be those boots on the ground, synthesizing and triaging what’s going on in the world that you need to be aware of.

AI Implementation at Pebble

Rob Marsh: We really haven’t talked about AI, which I think is ideal because it’s about understanding problems, challenges, and opportunities first. How is Pebble using AI to deliver that optimal experience and outcome for your clients?

Justin Whitehead: AI—to blur it a little bit—includes everything from generative AI like ChatGPT and LLMs, but also natural language generation, natural language processing, and linear models. Comically, at the end of the day, everything is just math underneath the hood. I hate saying AI because to me it’s just math, but AI gets someone to pick up the phone call.

Let’s talk about “explain” specifically. I mentioned we do portfolio analysis—it’s just math. All that quantitative six degrees of Kevin Bacon is all math. How did we produce our own knowledge graph of data? How did we procure this dataset and make these connections? Actually, a lot of AI techniques and generative AI techniques.

We’re a small company—I can’t spend $200,000 on a revenue database. I’m a startup, and I don’t know if there’s actually a business model here. So I’m going to build this database first before I try to buy it.

We ended up building unique proprietary datasets by teaching machines to extract structured data from unstructured data—SEC filings, investor presentations. It’s web scrapers, custom data feeds, then training extremely small, focused LLMs to do data extraction and populate a graph.

So AI techniques are used in building datasets. Then we do the portfolio math, the quantitative math, and we’ve got a mathematical understanding of where we think the catalysts are. Now we’re reading news using a mixture of traditional NLP techniques.

In the early days, we were pulling RSS news feeds. Most enterprises don’t like that, so we’ve matured and now have firm data agreements with actual underlying news sources. But you’re using traditional natural language processing techniques to cluster and pre-summarize data.

Natural language generation tech prior to LLMs still sounded like it was coming from a robot. But that ultimately gets fed into LLMs to clean up and rewrite. We can take the same engine outputs, the same math, the same pre-summarizations done very quickly and efficiently, but if we’re writing an explanation to a retail investor, we’ll use one class of language. If we’re talking to a sophisticated institutional asset manager, we’ll use completely different vernacular and provide a different level of depth.

That’s where the LLM gets involved. Also, because we sell to retail brokerage—a regulated entity—we have to answer to FINRA, the SEC, and firms’ internal legal risk and compliance teams. We do a lot of human-in-the-loop plus machine on the quality of output we write.

Our explanation engine produces about 10,000 to 15,000 explanations per day. We’ve got a human-in-the-loop process plus machine that can fact-check and eliminate hallucinations. Each explanation is actually vetted, potentially rewritten several times to iron out factual inaccuracies. If you send a factual inaccuracy through a retail broker, someone’s getting in trouble, and that’s going to be us.

Having all that auditability and statistical accuracy in what we’re writing gives us a huge edge in that space.

The Dynamic Nature of Explanations

Rob Marsh: You’re producing about 10,000 explanations a day—how many overlap?

Justin Whitehead: Hardly any. Each explanation lives only until fresher news or a bigger price move arrives—sometimes just 30 minutes. Heavy news days push the total to 15,000; quiet days, much less.

Rob: From a trader’s angle, how quickly does a narrative settle so I can act on it with confidence?

Justin: Depends on the company. For example, Apple once flipped overnight from “China sales” to “iPhone flop.” Clients today mostly want context—year-to-date drivers, three-day movers, real-time portfolio dashboards. But if someone wants deeper regime analysis, we’re ready to explore it.

Build vs. Buy: Pebble’s Competitive Advantage

Rob Marsh: You mentioned you’re small so you had to build everything, but you’re talking with big organizations. What’s the build-versus-buy decision, and what should be the main considerations?

Justin Whitehead: For anyone who knows financial market data providers, there’s a strong tribal pull to build everything here, don’t buy anything. I blame Bloomberg for being notorious for not wanting to partner with anyone—they build it here first. A lot of other companies followed that rhythm. It’s changing now, but not for Bloomberg.

I can safely say I can’t think of a client or prospect that isn’t already trying to do what we do—build portfolio commentary, build security commentary. A lot of this accelerated when ChatGPT was released two years ago and everybody became an AI expert. Boardrooms across the country had to come up with their AI strategy, everyone’s building chatbots, they spend tons of money on AI.

It’s really fast and easy to get a prototype of anything up and running. Everyone gets there like, “We can get a security explanation, just throw some news into an LLM.” Yeah, this is easy sauce.

Rob Marsh: I do it every night myself with Big Price Moves Explained.

Justin Whitehead: Then they find what’s not easy sauce about it. Now try to scale it. Now you want speed, you want to serve these explanations at cloud scale within certain latency times. Subsecond request and generative AI are two incompatible terms—you don’t get speed for that. So you have to start thinking very differently.

When you start dealing with legal risk and compliance—“How do we audit and make sure what was just said is truthful and not a hallucination?”—you get into the ugly, unsexy parts of AI anything, especially LLM anything in financial markets. That’s where most companies get stuck.

When it comes to Pebble, our advantage is—I say this upfront with every client and it gives us credibility—“Listen, I’ve been in fintech forever, so I get the whole ‘build it here first.’ I expect you guys, even after coming on as a client for Pebble, to continue your build-it-here efforts. I don’t mind—you should be trying to do this.”

But I’ll help get you to market faster and cheaper than before. Honestly, it’s shame on me if I can’t defensibly keep a moat between what you’re able to do and what I’m able to do. I don’t have a defensible business.

I’m in the business of getting into these clients and then just moving the goalposts forward faster and faster, doing more things with more data, solving the compliance headaches, solving the speed headaches, the scale headaches. Eventually people are like, “All right, we’ll allow Pebble.”

I’m not going to give you technology and then leave it. I like competition, and this is what infects my mind on a daily basis. I am living and breathing and loving life right now doing all of this.

Technology Evolution and Competitive Risks

Rob Marsh: What’s the risk that these great things you’ve done can have the technology move out from under you—like OpenAI comes out and basically a lot of the engineering you do becomes readily doable by something else?

Justin Whitehead: You get this in other firms—“Why couldn’t Google just come in and do all this stuff in finance? They’ve got all the data.” Google does not want to deal with the SEC or FINRA or any more regulatory bodies than they already have. There’s enough hairy, not-so-fun pieces to the financial services industry that OpenAI, if you want to come and deal with it, I give up. You’re going to hate yourself after a couple years.

The other thing is we have different endgame goals. OpenAI, Anthropic, Google through Gemini, even Facebook through Llama—everyone’s looking for large general AGI. That’s the goal: generalized intelligence so these machines can answer everything for everyone.

Where we do what we do, we actually focus on going the opposite direction. We start with large LLMs, but I’m in the explanation business. I need to do a few things super, super well. We can get tons of mileage going the opposite direction of the OpenAIs.

I go for speed, scale, operational efficiency, and security. OpenAI’s costs are going one direction. All my clients aren’t looking to spend more dollars on Azure—they’re looking for ROI, and if you can decrease the denominator, that’s also good.

Will I be able to have a drop-in replacement engine for the LLM portion that I can charge a client half as much—cut your LLM bill in half—but be quietly smiling because I’ve got an additional 80% margin on the other side? That’s where we’re going.

Being smaller and more focused is a direction we’re going, not necessarily where the big players are going. Nobody wants to be regulated. Google could do a lot more in finance if they wanted to—they’ve got the money—but there’s a reason why they don’t. You get a little bit of that protection as well.

Investment Thesis: The Future of Financial Services

Rob Marsh: What are the key things investors should know about Pebble as an investment?

Justin Whitehead: I’ve got some theories about how the financial services market is going to change over the next decade. An investor in Pebble—I want you to be on board with challenging me, but ultimately thinking this is a reasonable way the world is going.

I have very distinct opinions about what’s going to be the future for SaaS platforms at research shops and asset managers. I do not want to be a company building dashboards and tools for Wall Street individuals because I think the nature of how people are going to conduct research is going to change materially.

I can imagine a Goldman Sachs buying private licenses to their LLM du jour, and then it will be companies like Pebble, Bloomberg, FactSet, S&P, and MSCI all competing to add capabilities that these internal chat/research systems can do at the big banks.

While a lot of talk right now is about doing more with the same, the reality is it will be more with less at some point. It’s a hard thing to sell, but let’s be adults—this will have an effect on the jobs you actually need to run existing operations.

Even if you’ve gotten a boost in growth, you can still move more. I think we’re going to see AIs doing a lot more of the grunt work and monitoring. “Hey, ChatGPT, build me a DCF cash flow model, watch this company, give me a Teams message when the company posts earnings and blows past X.” That thing just quietly works behind the scenes. It’s not a junior analyst doing that—it’s a machine you’ve expressed your will to, and it will sit there and do that 24/7 without forming a union or requiring bathroom breaks.

Rob Marsh: And leaving an audit trail. A lot of intermediate information that presently gets lost becomes feedstock for compounding knowledge.

Justin Whitehead: I don’t think that’s a five-year scenario. I think you’re getting into the 10-year scenario.

The other aspect is how this is going to change wealth management, sandwiched between institutional asset management and retail brokerage. There are firms with their heads in the sand thinking the business model that worked in 2025 will work in 2035. I do not think that’s correct.

I think we’ll have a new form of wealth management brought to market by retail brokerage. Seventy percent of investors are not being served by wealth managers, but they still have similar needs. They have unsexy AUMs, but if they’re willing to spend 15 to 20 bucks a month to binge-watch Stranger Things on Netflix, you better believe they’re willing to spend 15 to 20 bucks to get comfort and automated understanding that they’re not headed toward financial ruin.

That business model will start to take shape over the next five years and have interesting ripple effects in wealth. The wealth industry will absolutely always be there, but the lower tier of existing wealth is going to get eaten from the bottom.

If you’re one of these vertically integrated firms—you’ve got brokerage, wealth management, asset management—you better understand there’s a war. The war has already started. The first battleground is on the unsexy retail brokerage side of the business. You better start investing fast because Robin Hood is coming to eat your lunch.

From the company standpoint, I am excited as hell to be part of that transformation. It’s not going to be Pebble alone, but you better understand that I see fire and I’m the guy with the gasoline cans. I’m going to make this thing burn faster than anyone can imagine.

If that excites you, great. You should come talk to us.

Key Lessons for AI Adoption

Rob Marsh: What are the key lessons you’ve learned about increasing the probability of successfully adopting AI in the investment process?

Justin Whitehead: Interestingly, I hear from the little birds on the street—and I’ve experienced this myself as a professional programmer—there’s all this talk about AI coming to replace software engineers, which I don’t think is true. However, man, is it useful.

But I still see in the financial profession people who aren’t experimenting and adopting. Get out there and just start experimenting with it, even in private. Get used to AI. Don’t worry about changing your investment practice or style.

I’ve used coding assistants to give me quick throwaway code. I know enough because I’m a professional to know when this thing is giving me garbage, but it does save time. What I’ve found most accretive has actually been on the non-engineering aspects—having a sounding board.

When I’m about to pitch a client, “Help me understand what this client is all about. What have they been talking about in the news?” What would take hours of dedicated reading and studying up on who I’m talking to can all be done now in seconds.

Find little micro problems and give it a try with ChatGPT. Learn when to trust it and what to look for when not to trust it. Make sure it’s citing sources so you can trust but verify. The more people experiment with it, the more it’ll end up being good for adopters in the industry and start sparking ideas of what can be possible.

I’m weirdly bumping into a lot of professional investors who are just like, “This is not for me.” Have you tried it? Don’t try it in a professional context—try it in a personal context. I think you’ll be amazed.

I’d encourage your listeners to hack around and experiment. I think you’ll be pleasantly surprised.

Rob Marsh: I would reduce it almost to the equivalent of 1985 to 1987: hit return on the keyboard. Just do it. The world’s not going to melt in front of you.

Justin Whitehead: Exactly.

Conclusion

Rob Marsh: That’s a wrap on today’s AI Investor Dialogues. Justin’s insights remind us that successful AI adoption isn’t about replacing human judgment—it’s about augmenting our capabilities to make better, more informed decisions faster than ever before.

Key takeaways: Start experimenting with AI tools in low-risk environments, focus on solving specific problems rather than chasing the latest technology, and remember that in financial services, trust and auditability aren’t optional—they’re essential.

If you found this conversation valuable, please share it with fellow investors and practitioners. Until next time, keep building, keep learning, and keep pushing the boundaries of what’s possible in investment management.

Again, Justin, thank you for sharing.

Justin Whitehead: Excited. Thank you for bringing me on.