One pattern shows up again and again: many investors tried AI a year ago or so, had a bad experience, and wrote it off because it got facts wrong, misstated figures, or invented citations.

Yet they keep hearing about AI in finance. That creates a different kind of pressure: not wanting to be left behind, without a clear sense of what has changed.

In this interview, he explains why product overload is slowing adoption, why “skills files” matter more than model training, and how to structure workflows so outputs can be trusted.

He also points to areas like fixed income, where these approaches may matter more than people expect.

The conversation also covers something that doesn’t get enough attention in finance: how bias shows up in ways that aren’t obvious. Not just demographic bias. Positional bias (the same information, presented in a different order, can produce different outputs), framing effects (the same odds stated two ways lead to different decisions), and the fact that models reflect the biases in the data they’re trained on.

We cover:

• Why many investors are still anchored to early AI failures

• Why comparing models is the wrong approach

• How “skill files” and workflows actually drive results

• Where early ROI is showing up (including fixed income)

• How bias shows up in model outputs

This interview has been edited for clarity and length.

Matt: You talk to a lot of investment professionals just getting started with AI. What’s your sense of adoption among investors?

There’s a large group that tried ChatGPT when it first went mainstream, had a bad experience — it made something up, got a calculation wrong — and wrote it off. That’s a reasonable response based on what it was at the time. The problem is anchoring. They’ve frozen their view of the technology at that moment, and the tools are genuinely different now. I tell them: forget everything you knew about this 18 months ago. You have to be experimenting again.

The other group has FOMO, but the anxiety from not knowing where to start is actually keeping them from doing anything. My advice to both groups is the same: treat it like a new employee. Give it a task. Check the output. See what it can do.

From Models to Workflows

Matt: We are seeing a total product overload right now. It isn’t like comparing phones based on pixel counts; it is very difficult to compare these AI models side-by-side. How should people navigate this?

Brian: It is very hard to compare them, and getting all of them at once can be overwhelming. I recommend trying to understand what you can do with the “Frontier” models and Claude’s skills—as well as the skills OpenAI is developing—and what can be achieved with connectors. For example, Notion already acts as an agnostic transcript writer that can connect to Claude. Many investment professionals are not yet aware of the tools that are used ubiquitously in the computer science realm.

Matt: I’ve had Claude skills on my radar for a few months, but I’m still trying to get my arms around them. How are you using them currently?

Read more

Manage Email Preferences

If you prefer to receive one weekly email with all AI Street content, deselect Research and Interviews.

Manage How Often You Receive AI Street

NEWS

Goldman CIO on AI Inside the Bank

Goldman Sachs CIO Marco Argenti said AI has made it cheap enough to build smaller applications in-house, with the firm already terminating some third-party software contracts.

Marco Argenti, speaking on Bloomberg’s Odd Lots podcast, said the buy-versus-build calculation has shifted. Engineers can now build working applications in days rather than months, he said.

Other notes from the podcast:

• Goldman’s internal AI assistant is deployed to 47,000 employees logging more than a million prompts a month.

• On token costs, Argenti expects per-unit prices to fall but total consumption to rise faster. Goldman built a model gateway that routes queries to the cheapest model that can handle them. “Total token cost is going to be a major item of cost in any organization,” he said. “It’s to be compared to the cost of people.”

Agents as a Service

Citi Ventures says AI-native startups are beginning to convert work historically done by humans in functions like IT, sales, legal, and HR into software-driven workflows, charging for completed tasks rather than software licenses and putting pressure on traditional per-seat SaaS pricing.

The note points to the growing importance of specialized models, particularly large tabular models designed for structured financial data like fraud detection and credit risk, where traditional language models are less effective.

Related

Retail Broker Rolls Out AI agents that Place Trades

Public, a retail brokerage, is rolling out AI agents that can execute trades on behalf of customers, automating tactics like protective hedging, cash sweeps into higher-yield assets, and stop-loss orders based on predefined rules.

Users write out a strategy, refine it through follow-up prompts, and review a step-by-step workflow before the agent goes live. The system logs every action and operates within fixed instructions. “It can only do what you tell it to do,” co-CEO Jannick Malling told the WSJ.

Public joins Robinhood and eToro in pushing AI beyond research and into execution. Brokerages are now starting to let those systems convert user-defined strategies into live trades.

FDIC Moves to Ease AI Model Risk Rules for Banks

The FDIC is revisiting its model risk management guidance, with an agency official telling Congress it has been applied too broadly and imposed unnecessary burdens, particularly on smaller banks, without a meaningful reduction in risk. The agency is working with the Fed and OCC on a more tailored, risk-based approach that accounts for a bank’s size, complexity, and the materiality of each model.

The shift does not change safety expectations but could reduce supervisory friction for lower-risk AI applications such as internal copilots and summarization tools, while models used in credit and risk management remain tightly governed. The official also said the FDIC expects to roll out GenAI tools to its own staff by mid-year.

Top Bridgewater Scientist to Join Google DeepMind

Jasjeet Sekhon, former head of AI at Bridgewater Associates, is set to join Google DeepMind as chief strategy officer, a move that reflects how quantitative investing and AI research are converging: both train models on large datasets to identify patterns and make predictions.

Perplexity Runs AI Agent Stock Pitch Contest

Perplexity is running a stock pitch competition for college students, offering $17,500 in prizes for investment theses built using its Computer product, a general-purpose AI agent that it says executes multi-step workflows autonomously.

The judges are Philippe Laffont of Coatue, Dan Loeb of Third Point, and Ken Hao of Silver Lake. Submissions are due April 3.

What Else I’m Reading

• AI Companies Shatter Fund-Raising Records, as Boom Accelerates NYTimes

• Bank of America’s wealth management firms roll out AI tool Banking Dive

• Starling Bank Launches UK’s First Agentic AI Money Manager Fintech Times

• QuantumStreet AI launches long-short global equity strategy Hedgeweek

Share AI Street

This Week in AI Street

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.

Discounts for early sponsors.

CALENDAR

Upcoming AI + Finance Conferences

• Future Alpha – Mar. 31–Apr. 1• NYC
  Cross-asset investing summit focused on data-driven strategies, systematic investing, and tech stacks.

• AI in Finance Summit NY – Apr. 15–16 • NYC

  The latest developments and applications of AI in the financial industry.

• Momentum AI New York – Apr. 27–28 • NYC
  Senior-leader forum on AI implementation across financial services, from operating models to governance and execution.

• AI in Financial Services – May 14, 2026 • Chicago
  Practitioner-heavy conference on building, scaling, and governing AI in regulated financial institutions.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• connect with me on LinkedIn, or

• send an email to matt [at] ai-street.co

Cornell’s new AI in Finance certificate began in this vacuum. Victoria Averbukh, Professor of Practice and Director of Cornell Financial Engineering Manhattan, spent two years talking to portfolio managers, traders, and strategists before designing it.

“People would say it was about not having a clear way to think about the systems, what the system is doing,” Averbukh said. “New terminology kept coming in.”

Manage Email Preferences

If you prefer to receive one weekly email with all AI Street content, turn off Research and Interviews here:

Manage How Often You Receive AI Street

The program — 30 sessions, 13 instructors — mixes Cornell faculty with practitioners from fintech, asset management, investment banking, and trading. The goal is judgment, not tool proficiency.

Kathryn Zhao, Head of Institutional API Product, OKX, says it reflects a shift already underway in hiring. Domain experience used to be the deciding factor. Now she screens for AI awareness.

“If someone understands how to work effectively with AI tools [...] they can onboard quickly and begin contributing almost immediately,” Zhao said.

In the conversation below, we discuss:

• Why applying AI in finance can’t be a direct translation from tech

• How the certificate balances academic foundations with practitioner insight

• What AI awareness means for hiring and talent development

• The biggest stumbling blocks for AI adoption in financial services

• Why chasing the pace of change is less useful than building understanding

The below conversation has been edited for clarity and length.

Matt: What was the genesis of this certificate? When did you decide to do this, and what was the catalyst?

Victoria: I kept hearing, across different finance sectors, that people were either using AI and finding it useful but not fully comfortable with whether they could trust it, or they didn’t even know where to start. That hesitation was very consistent—for portfolio managers, traders, execution people, strategists, research people, more quantitative people, less quantitative people. It didn’t really matter. People would say it was about not having a clear way to think about the systems, what the system is doing. New terminology kept coming in. We started with AI, then the word “agent” appeared. It just felt either overwhelming or there was a lack of trust.

My light bulb went on around 2024, about two years ago. I remember that Kathryn and I actually went to have coffee at Breads Bakery on the Upper East Side, and I said, “Kathryn, I have this thought.” And Kathryn said, “Yes!” She was one of my very early supporters. That coffee at Breads Bakery is what gave me confidence to go and investigate more.

Matt: What makes applying AI in finance different from applying it in tech?

Victoria: After speaking with Kathryn, I also spoke with Andrew Chin, Marcos López de Prado, and others. They were all very clear that education is needed, partly because of the hesitation we just talked about, but also because finance is not tech, and applying AI here requires respecting that difference.

Machine learning, big data technologies, and large language models were all built for something else, not for finance. Uber’s business model, for example, is built around offering a service powered by new technology. That is fundamentally different from what a bank or a hedge fund does. So applying AI to investing, to execution, to alpha generation, or even to forecasting market exposure cannot be a direct translation. The objectives are different, and the data is different. Financial data is non-stationary, often smaller, and rarely clean, so you cannot just take machine learning methods from tech and apply them directly.

Our industry is and will continue to adopt AI, but it has to be done carefully, with a real understanding of what works and what does not. Everyone I spoke with strongly supported the idea that training is needed specifically because of these differences, and that developing critical understanding, judgment, and a clear sense of potential ROI before adoption is essential.

Which is why the real question is not whether we use AI, but how we use it in a way that actually improves decision-making rather than just adding complexity.

Matt: Can you talk about the structure of the certificate and the role of practitioners in it?

Victoria: The full certificate is about 30 sessions with 13 instructors. The curriculum is deliberately structured to start from fundamentals — faculty from Johnson School and Engineering explain what the data is, what an LLM is, and work through use cases.

But because it’s so fast-changing, you really need practitioners to understand what needs to be done. Finance is an extremely regulated industry. I think that’s another thing that differentiates it. Even probably from healthcare.

The industry instructors are very carefully curated to give breadth of coverage — fintech, asset management, investment banking, and trading. This is not a certificate just for trading or fraud detection or financial advising. It’s for everybody. Ideally you have some experience on Wall Street, but also if you’re just starting out, it’s really for everybody.

Do you know the quote from Einstein? “If you can’t explain it simply, you don’t understand it well enough.” That was my guiding principle. I know that our Cornell faculty can take the complicated topics — transformers, LLMs, all of that — and make it intuitive. Developing that intuition is really the intention behind the certificate. It’s what enables you to make sound judgments about when, where, and how AI should be used, and when it shouldn’t.

Recent Interviews

Matt: Kathryn, where do you see this heading? How do you see AI impacting finance in the next couple of years?

Kathryn: Speaking from a practitioner’s perspective, my approach to hiring has fundamentally changed. A few years ago, I would evaluate candidates primarily based on their prior experience in the specific role or industry. Today, that is no longer the deciding factor, particularly for junior hires.

What I prioritize now is AI proficiency and AI awareness. If someone understands how to work effectively with AI tools (how to ask the right questions, interpret outputs critically, and apply insights to real business problems) they can onboard quickly and begin contributing almost immediately. With access to AI-generated materials and the ability to leverage AI as a day-to-day copilot, the learning curve is dramatically compressed.

In that sense, traditional domain experience is no longer a strict prerequisite. What matters more is a strong baseline understanding of the real world at a college-educated level, combined with the ability to operate fluently in an AI-enabled environment.

That is why I believe an AI in Finance certificate program is highly relevant. It prepares participants to become AI-aware and AI-capable without requiring them to be programmers. More importantly, the AI literacy and applied mindset the program builds will open a wide range of opportunities for participants in the years ahead.

Victoria: When you say the person needs to be AI-aware—does that mean the person can have zero finance knowledge, or do you mean they don’t need deep knowledge of Python and machine learning?

Kathryn: They don’t need to come in with deep expertise in Python or extensive financial industry knowledge. Those capabilities can be developed on the job. What matters most as a baseline is their ability to work effectively alongside tools like Claude: knowing how to frame the right questions, extract the right information, and translate insights into action.

Victoria: I agree with Kathryn, but maybe a notch below the enthusiasm. Here’s why: This certificate is not about the tools. It’s about understanding the lay of the land and developing intuition. Learning what Claude does can be done on YouTube. There are plenty of tutorials.

The AI awareness Kathryn mentioned, that’s what we bring in the certificate. Ideally, as an educator, I want participants to leave thinking: I know what questions to ask. I know how to bring judgment to that Claude-generated code. So maybe we’re fast-tracking people a little bit through the first nine months on the job once Kathryn hires them.

I also think finance is segmented. You can be an expert in energy, or equities, or fixed income, or mortgages. You can be a really great financial advisor, but you wouldn’t necessarily know how to construct a global allocation as a portfolio manager. At some point, applications of AI are going to become more tailored to all these different areas. It’s almost like you’re not going to go to a dentist if you need new glasses.

Ideally, if this certificate is successful and we offer it again and again, I certainly want to make sure that we have significant participation from practitioners, from industry. Engineers will be inventing new AI 2.0 and 3.0 and 10.5, but the industry participation will always be needed. Maybe we break it up or reshape it to focus on specific areas of finance, that’s also a possibility.

Matt: What is the biggest stumbling block right now for AI adoption in finance?

Victoria: I think it’s uncertainty. I think it’s leadership that is probably older and did not grow up with phones in their hands. There’s a certain inertia. Bridging the generational gap is going to be harder. I think CEOs are going to get younger.

Matt: How do people keep up? It feels like the terminology alone is a moving target.

Victoria: There’s no glossary out there. That glossary changes dynamically. That’s going to be part of the certificate. Once people finish, they’re going to know the terms and will be more comfortable and ready for a new iteration of terms. But ultimately, I think trying to chase the pace is impossible. Focus on understanding, not the hype.

Manage Email Preferences

If you prefer to receive one weekly email with all AI Street content, turn off Research and Interviews here:

Manage How Often You Receive AI Street

The transformer breakthrough began in text, but researchers are adapting the architecture to other kinds of sequential data. With enough data, transformer models can learn the patterns of “language” in that dataset in ways that traditional models missed.

For example, AlphaFold is a transformer-based system trained on protein data to predict how proteins fold into their 3D shapes from amino acid sequences, which determine how they function. It effectively solved the protein folding problem. (The work later contributed to a Nobel Prize in Chemistry awarded to its creators, including Demis Hassabis, who is not a chemist.)

As I’ve written before, no one knows exactly how these models work. They’re grown rather than built, as the CEO of Anthropic likes to say. We didn’t know how aspirin worked for like 70 years, but we knew it was effective.

This brings us to a new paper from JPMorgan researchers, who trained a transformer model on market data.

The market as a language

Every buy, sell, order submission, or cancellation leaves a trace: what happened, how much size was involved, how far from the market mid-price it was placed, and when it occurred. Multiply that across thousands of stocks and millions of events per day, and you get a massive stream of sequential data.

TradeFM is a 524-million-parameter model trained on 10.7 billion training tokens drawn from more than 9,000 U.S. equities, using data spanning 368 trading days from February 2024 to September 2025.

Instead of predicting the next word in a sentence, their model — called TradeFM — predicts the next event in a sequence: its timing, size, price depth, and direction.

Trading data is messy. Stocks trade at different prices. A $5 move on a $2 stock is massive. A $5 move on one that’s $500 isn’t news.

If you feed those raw numbers into a model, it can’t really compare one stock to another, so it struggles to learn general patterns.

So the researchers adjusted the data before training. They expressed price-related features in relative terms, compressed volumes so large and small trades are easier to compare, and measured time as the gap between events.

That puts different stocks on a common scale, so moves are comparable whether it’s a $2 stock or a $200 stock.

They then discretized each event’s features and combined timing, price depth, volume, side, and action type into a single composite token. The result was a vocabulary of 16,384 trade event tokens.

Related Research

What they found

The researchers tested the model inside a simulated exchange, where it predicts trades in a continuous loop. The resulting data reproduces core patterns seen in real markets, including clustered volatility and large price swings. Across 9 stocks, 3 liquidity tiers, and 9 months of held-out data, it matched those patterns 2 to 3 times more closely than a standard baseline known as a Compound Hawkes process.

What’s most interesting is that TradeFM’s behavior extends beyond the U.S. data it was trained on. JPMorgan tested the model, without any adjustments, on trading data from China and Japan, where market structure differs meaningfully. Japan uses batch auctions at the open. China imposes 10% daily price limits. Spreads in both markets are several times wider than in the U.S. Despite those differences, the model’s performance degraded only moderately. It had never seen these markets, yet it still captured their core dynamics.

The model appears to be learning structure that carries across markets.

Subscribe now

Arman Khaledian, PhD, a former quant at Millennium and now CEO of Zanista AI, said: “That’s not a toy result. It means the model is picking up something real about how markets work at a structural level.”

Read more

Manage Email Preferences

If you prefer to receive one weekly email with all AI Street content, turn off Research and Interviews here:

Manage How Often You Receive AI Street

NEWS

AI as Money Manager

From Investment News:

Era, a San Francisco startup founded by ex-Stripe employees, has registered with the SEC as an RIA that delivers investment advice through AI without client-facing human advisors.

Era’s Context financial data hub connects to AI agents such as Anthropic’s Claude, OpenAI’s ChatGPT and Google Gemini, among others. Clients can subscribe to predetermined portfolio mixes for their investments and access automated services such as spending analysis, savings round-ups, money transfers between accounts, budgeting, and portfolio rebalancing.

As a registered investment adviser, Era is held to the same fiduciary standard as other RIAs, meaning it must act in the client’s best interests, according to Era CEO Alex Norcliffe.

Rather than rely on raw LLM outputs, the firm says it engineered a system to deliver consistent, traceable advice across clients with similar profiles.

While those who are old enough to remember those thundering herd commercials aren’t likely to trust AI with their money, younger folks are. About 41% of Gen Z and millennials would allow an AI assistant to manage their investments, versus 14% of Baby Boomers, according to a World Economic Forum survey.

AI Joins the C-suite at HSBC

HSBC named David Rice as its first chief AI officer, tasked with scaling AI across the bank, including giving staff access to generative tools for internal workflows and customer interactions.

More Wall Street firms will elevate AI executives to the C-Suite. The technology is slowly evolving into a company’s “brain.”

DoorDash Turns Couriers Into AI and Robot Trainers

From Bloomberg:

DoorDash Inc. is paying delivery couriers in some markets to submit video clips and complete other digital tasks to help improve artificial intelligence and robotics models, following competitors that have found creative new uses for gig workers in the AI boom.

The latest models draw attention as they surpass prior state-of-the-art benchmarks, but their voracious appetite for data is often overlooked. As I wrote last week:

The latest models have hoovered up the whole internet and that’s still not enough. Last year, OpenAI was paying YouTubers as much as $4 for a minute of their unused footage.

Legacy firms are sitting on decades of data. Unlocking it is hard, but it’s an advantage. You can’t hire DoorDash drivers to mimic financial transactions.

Trump Pushes for Federal Control Over AI

The Trump administration is pushing Congress to preempt some state AI laws and create a lighter national framework, while still leaving states authority in areas such as fraud, consumer protection, zoning, and their own use of AI. The administration argues that a patchwork of state AI laws could create conflicting requirements for companies operating nationally. A federal standard would reduce some state-by-state compliance differences, though states would still retain authority in several areas.

So far, there have been few changes to financial regulatory rules. U.S. regulation sets broad standards and applies them to new technologies like AI rather than writing detailed, tech-specific rules.

Fed Expects Banks to Expand AI Use Beyond Low-Risk Work

A senior Federal Reserve official said banks are still using AI in limited, low-risk areas like summarization and coding, but expects adoption to expand into more material functions as use cases mature.

The Fed is not treating this as a new regulatory category. Existing expectations around model risk, data quality, and governance apply, with a continued emphasis on testing and human oversight. “Judgment and decisionmaking will remain with subject matter experts.”

Supervisors are also starting to use AI themselves, including to analyze earnings calls, filings, and other public data as part of bank monitoring.

WHAT ELSE I’M READING

• Mark Zuckerberg Is Building an AI Agent to Help Him Be CEO WSJ

• Goldman Sachs Puts AI at Core of New Strategy Banking Exchange

• VanEck CEO Details AI’s Impact on ETF Asset Management ETF Trends

• AI Has Already Changed Software Contracts: They’re Shorter BBG

• AI in banking ‘not a silver bullet’: Mike Mayo BankingDive

• AI boom risks widening wealth divide, says BlackRock’s Larry Fink Guardian

• Quilter signs deal with AI adviser tool Aveni for its network Citywire

• Norway wealth fund moves towards some AI-driven decisions but with humans in control Reuters

Share AI Street

INTERVIEW

Morgan Stanley’s ex-AI head: Start with the Work, Not the Model

Most firms are still stuck in pilots. The constraint is how work is structured, not model capability.

RESEARCH

AI spots when companies change which metrics they highlight

Firms that frequently swap metrics tend to underperform. LLMs improve detection by extracting full phrases instead of keywords.

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email me (sponsors@ai-street.co) for more details. Discounts for early sponsors.

CALENDAR

Upcoming AI + Finance Conferences

• Future Alpha – Mar. 31–Apr. 1• NYC
  Cross-asset investing summit focused on data-driven strategies, systematic investing, and tech stacks.

• AI in Finance Summit NY – Apr. 15–16 • NYC

  The latest developments and applications of AI in the financial industry.

• Momentum AI New York – Apr. 27–28 • NYC
  Senior-leader forum on AI implementation across financial services, from operating models to governance and execution.

• AI in Financial Services – May 14, 2026 • Chicago
  Practitioner-heavy conference on building, scaling, and governing AI in regulated financial institutions.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

Thanks for reading! I’m always happy to receive comments, questions, and feedback.

• connect with me on LinkedIn, or

• send an email to matt [at] ai-street.co

His advice: don’t start with technology. Identify work that can be automated.

Many companies are doing the opposite— buying tools first and figuring out where they fit later.

“We’re letting the vendor marketplace drive our strategy as opposed to asking the question: what do you want?”

McMillan, who recently launched McMillanAI, where he advises executives on AI strategy, says many organizations are still early in figuring out how to deploy AI at scale.

In practice, that means starting with tasks that take up a lot of time and are repeated across large teams—call centers, onboarding, compliance review. These are areas where AI can replace or augment work in a measurable way.

What breaks at scale isn’t the model. It’s everything around it: how data is structured, who has access, how systems are monitored, and how much autonomy they’re given.

Most firms haven’t solved that yet. They’re experimenting with tools, but haven’t redesigned how work actually gets done.

We cover:

• Identifying high-volume work AI can replace

• What breaks when you try to deploy AI at scale

• Why most firms are still stuck in pilot mode

• How to think about vendors vs building in-house

• Where agents are actually being used today (and where they aren’t)

This interview has been edited for clarity and length.

Matt: Why start McMillanAI now?

Jeff: There’s an enormous gap in the marketplace around education and awareness. The people that need to make the decisions — and by the way, I’ve probably spoken to no less than 30 CEOs in the last six weeks, CEOs of Fortune 500 companies — they want to do AI. They’re getting pressured to do AI. But we have a workforce that knows more about this technology than most senior people do in organizations. That’s a gap, and that’s an opportunity.

I don’t want to sound Pollyannaish about this because I’m not: this is a once-in-a-generation type of technology, and I really do believe that we have a choice as humanity. We have a choice about how we deploy this for the benefit of all of us as opposed to maybe a few. I’d like to be part of that dialogue.

Matt: Going back to those 30 conversations, what were the common threads?

Jeff: There’s an enormous amount of external pressure on them. There’s no CEO I talked to that says, “I don’t believe in AI.” That was maybe true three years ago — “Is this just the next crypto? Is it the next metaverse?” No one believes that now. Everyone believes there’s something going on here. So that’s number one.

Number two, there’s a tremendous desire to do something, but they don’t have the skills and the competencies to do this technology at an enterprise level. If you look at every major technical transformation, it takes eight to 10 years to fully play out. So, we’re very early in the process.

The problem with AI is it requires a different approach, and it’s not a technology problem.

Read more

The C-suite can choose what numbers to highlight — revenue per user, lifetime value, total addressable market, etc.

These selected metrics support the corporate narrative, but they’re not static. When a number is strong, it gets airtime. When it softens, it tends to quietly disappear from the script, replaced by whatever metric tells a better story that quarter.

Shifting corporate narratives happens so often it has a name: “moving targets.” This tracks the fraction of previously highlighted metrics that go missing from the next comparable earnings call. Research showed that firms with high metric turnover tend to underperform in subsequent months. The more a company reshuffles the numbers it talks about, the worse its stock tends to do.

The challenge is detection. Most approaches rely on keyword matching across transcripts, comparing terms quarter over quarter. You need to be able to link “revenue growth” to “top-line expansion.” Keyword search can’t distinguish “North America cloud revenue” from “revenue.”

At scale, this becomes difficult to track. Following the metrics that appear and disappear across thousands of earnings calls is not feasible to do consistently by hand. This is where LLMs fit: extracting and standardizing how companies describe performance over time. This is tedious work for humans and easy to scale with AI.

A group of researchers at MIT, BlackRock, and J.P. Morgan asked whether LLMs could close this detection gap.

Here’s what they did:

• Instead of scanning transcripts for predefined terms, they use an LLM to extract full phrases with context. Where keyword methods pull “revenue,” the model pulls “North America cloud revenue.” Where it grabs “dividends,” the model also captures “cash flow,” “share repurchases,” and “cash flow from operations.”

• They then compare metrics across quarters using semantic similarity rather than exact matches. Instead of forcing a binary match, they allow for an “ambiguous” range where similarity is scaled.

• They apply this across firms listed in the S&P 100 index from January 2010 to December 2024, yielding 5,615 firm-quarter observations across 64 quarters.

• To test it, they sort stocks by how much their metrics shift and compare returns, then run cross-sectional regressions with standard controls for size, valuation, and prior returns.

Here’s what they found:

Read more

• News Roundup: Nvidia announces NemoClaw, HSBC, Terminals, Grok

• Research: AI for Excel still can’t make it on Wall Street

• Interview: Kevin McPartland on the internet-scale impact of AI

Manage Email Preferences

If you’d like to receive fewer emails, you can manage your subscription and turn specific sections on or off.

Manage How Often You Receive AI Street

NEWS

Nvidia Says OpenClaw Is “the New Computer”

Nvidia this week introduced NemoClaw, a software toolkit designed to make OpenClaw usable inside enterprise environments. OpenClaw is an open-source framework for building AI agents that can plan, execute tasks, and coordinate subagents across tools and data sources.

“Every company in the world today needs to have an OpenClaw strategy,” said CEO Jensen Huang. “This is the new computer.”

OpenClaw has quickly become one of the most widely used agent frameworks, allowing developers to build systems that can plan tasks, execute them, and spin up subagents to handle specialized work with limited supervision.

The problem: OpenClaw is a security nightmare as Cisco puts it.

Nvidia hopes its NemoClaw is the answer. The operating system runs agents inside controlled environments, with restrictions on what they can access and how they behave. The system adds monitoring, policy enforcement, and guardrails around data and network activity.

Currently, AI tools are powerful for individuals, but they aren’t designed for coordinated use across teams, systems, and models.

Separately, Meta-acquired Manus launched a desktop application that enables its AI agent to operate locally on personal devices, allowing direct control over files and applications to rival the popular OpenClaw agent.

We’re starting to see a world where AI doesn’t live in a chatbot, but is part of the operating system. That shifts the enterprise question away from “which model is best” and toward something more practical: what specific tasks AI agents can reliably execute inside existing workflows.

You can start to imagine a TaskRabbit-style marketplace for agents, where firms deploy specialized systems on demand to handle discrete tasks.

HSBC is considering cutting up to 20,000 jobs (about 10% of its workforce) over the next three to five years as it uses AI to automate middle and back-office functions, particularly in non-client-facing roles, according to Bloomberg.

I’m skeptical that AI is the primary driver here. (The preliminary plan also includes attrition, restructuring, and potential business exits.)

If AI were already replacing labor at scale, you would expect to see more consistent signals across large banks. That’s not happening yet. JPMorgan, which spends roughly $20 billion a year on technology, reported headcount was essentially flat last month, according to CNBC.

As I’ve said many times around here, there’s currently no clear evidence that AI is dramatically reducing jobs. Maybe that changes, but right now it’s a convenient narrative for broader cost-cutting.

Finance Bros to Tech Bros: Don’t Mess With My Bloomberg Terminal

The rumors of the death of the Terminal are greatly exaggerated. Startups have been “coming for” the financial giant for like 40 years. The WSJ had a piece this week about how the VC crowd misunderstands the staying power of the ubiquitous black and orange screens on Wall Street.

I’m not saying this because I worked there for 10+ years, but because what we’ve seen in this current AI boom: proprietary data is becoming more valuable. Hell, there is so much demand for data, they’re creating synthetic data. The latest models have hoovered up the whole internet and that’s still not enough. Last year, OpenAI was paying YouTubers as much as $4 for a minute of their unused footage.

Incumbents with decades of data have an advantage.

Musk’s xAI Hiring Credit Experts, Bankers to Teach Grok Finance

From Bloomberg:

Elon Musk’s artificial intelligence startup xAI is looking to hire bankers and private credit lenders to make its Grok chatbot better at finance strategy, joining rival AI firms in pushing software for investing professionals.

I still see commentary online that the people training AI are basically talking themselves out of a future job. That’s just not true. From October when news broke that OpenAI was hiring bankers, I wrote:

Reinforcement learning — I’m very much oversimplifying — is basically a giant game of hot and cold. The data helps tell the model when it’s getting “warmer” or “colder” to the right answer. It doesn’t explain why it’s the right answer.

So while AI companies like to market their models as “thinking,” they’re not really reasoning in the syllogistic sense. (All men are mortal; Socrates is a man; therefore Socrates is mortal). These models are mimicking thinking, not doing it.

AI is coming for investment banking tasks. The tech is starting to automate the tedious point-and-click white-collar work that has historically taken junior analysts 100+ hours a week.

ROUNDUP

What Else I’m Reading

• Markets misread AI’s impact on brokerages, MS says Investment News

• Beyond productivity: AI’s structural shift in active management Schroders

• Rogo Acquires AI Agent Company Offset PR

• Market data price increases slow for first time in five years: study The Trade

• Agentic AI Startup Lyzr Raises Funds at $250 Million Valuation BBG

• AI Nears ‘Inflection Point’ in Asset Management MarketsMedia

• Six skills for financial service professionals Anthropic

RESEARCH

AI Still Falls Short in Excel: Study

Even the best model wouldn’t make it on Wall Street

AI for Excel has improved dramatically over the last six months, but it still wouldn’t make it as a junior analyst on Wall Street.

The three leading AI providers, OpenAI, Anthropic and Google’s Gemini, have all released updated capabilities for Excel and are reporting higher accuracy on benchmarks.

For example, OpenAI said this month that performance on its internal investment banking benchmark jumped to 87.3% (GPT-5.4 Thinking) from 43.7% (GPT-5).

While these are material improvements, AI still can’t be relied on without supervision. “Mostly right” is not good enough.

A recent benchmark on AI in Excel points to the same issue: performance drops sharply as tasks get more complex.

FinSheet-Bench tests how models handle real-world private equity workbooks with messy layouts, multiple funds, and non-standard formatting. Across 10 models from OpenAI, Google, and Anthropic, the best result came from Gemini 3.1 Pro at 82.4% accuracy, followed closely by GPT-5.2 with reasoning and Claude Opus 4.6 with thinking, both around 80%.

On simple lookups, top models exceed 90% accuracy.

The gap widens further on large, realistic files. On the most complex workbook tested, with 152 companies across eight funds, average accuracy was worse than a coin flip.

Read the Full Research Report

INTERVIEW

The Limits of AI in Trading

Kevin McPartland has spent more than 20 years studying how technology changes market structure.

He expects AI to have an internet-scale impact on markets:

𝘐 𝘢𝘮 𝘢 𝘣𝘦𝘭𝘪𝘦𝘷𝘦𝘳 𝘵𝘩𝘢𝘵 𝘵𝘩𝘪𝘴 𝘪𝘴 𝘪𝘯 𝘴𝘰𝘮𝘦 𝘸𝘢𝘺𝘴 𝘭𝘪𝘬𝘦 𝘵𝘩𝘦 𝘧𝘪𝘳𝘴𝘵 𝘥𝘰𝘵-𝘤𝘰𝘮 𝘣𝘰𝘰𝘮. 𝘚𝘶𝘳𝘦, 𝘪𝘵 𝘸𝘪𝘭𝘭 𝘤𝘩𝘢𝘯𝘨𝘦 𝘫𝘰𝘣𝘴 𝘢𝘯𝘥 𝘵𝘩𝘦𝘳𝘦 𝘸𝘪𝘭𝘭 𝘣𝘦 𝘫𝘰𝘣 𝘭𝘰𝘴𝘴𝘦𝘴, 𝘸𝘩𝘪𝘤𝘩 𝘯𝘰𝘣𝘰𝘥𝘺 𝘦𝘷𝘦𝘳 𝘸𝘢𝘯𝘵𝘴. 𝘉𝘶𝘵 𝘪𝘯 𝘵𝘩𝘦 𝘭𝘰𝘯𝘨 𝘳𝘶𝘯, 𝘵𝘩𝘪𝘴 𝘪𝘴 𝘢 𝘵𝘰𝘰𝘭 𝘢𝘯𝘥 𝘢𝘯 𝘢𝘪𝘥 𝘵𝘰 𝘩𝘦𝘭𝘱 𝘱𝘦𝘰𝘱𝘭𝘦 𝘥𝘰 𝘵𝘩𝘦𝘪𝘳 𝘫𝘰𝘣𝘴 𝘣𝘦𝘵𝘵𝘦𝘳 𝘢𝘯𝘥 𝘵𝘰 𝘤𝘳𝘦𝘢𝘵𝘦 𝘯𝘦𝘸 𝘫𝘰𝘣𝘴 𝘸𝘦 𝘥𝘰𝘯’𝘵 𝘬𝘯𝘰𝘸 𝘢𝘣𝘰𝘶𝘵 𝘺𝘦𝘵. 𝘐 𝘳𝘦𝘢𝘭𝘭𝘺 𝘵𝘳𝘶𝘭𝘺 𝘧𝘦𝘦𝘭 𝘭𝘪𝘬𝘦 𝘵𝘩𝘢𝘵’𝘴 𝘸𝘩𝘦𝘳𝘦 𝘵𝘩𝘪𝘴 𝘪𝘴 𝘨𝘰𝘪𝘯𝘨.

He leads market structure and technology research at Crisil Coalition Greenwich, where he tracks how banks, asset managers, and trading firms deploy new systems. He previously worked at BlackRock and TABB Group.

But AI adoption on trading desks has yet to scale.

A recent report from the firm shows the most common use cases among bond traders are still data analysis and document review, not execution or decision-making.

Trading desks face clear regulatory and reputational risk, where firms need to explain and defend decisions to clients and regulators. As McPartland puts it, you can’t tell regulators: “Well, the AI did it.”

In this interview, McPartland explains where AI is being deployed today, what’s holding back trading applications, and why coding and developer productivity may be the most important near-term use case.

This interview has been edited for clarity and length.

Why AI Adoption in Trading Is Moving Slowly

Matt: I was checking your reports on AI, and you guys are focusing on how it’s in the back office. That seems like it’s the story across the street. When do you think it goes beyond that to more trading applications?

Kevin: I think we’re starting to get there. I was at FIA Boca earlier this week and there was definitely a lot of talk about AI. I think the industry is excited and interested but really trying to be cautious. There’s a reputational risk issue, a regulatory issue. You don’t want to do the wrong thing for your clients from the sell-side perspective. If there’s an issue and regulators come to you and ask what happened, you can’t just say, “Well, the AI did it, I’m not sure.” That’s not a good answer. So I think that’s leaving people cautious.

We actually just got back a study of bond traders and we asked them where they saw the opportunity in AI. Not surprisingly, data analysis was number one, document review number two. So it still really is about poring through data and unstructured data to help digest it, find insights, find patterns. I think that’s still the biggest use case now.

My two cents — I think where really a lot of the impact will be in the short, medium, and long term is on the coding side. Everything from making the most sophisticated quant developers even more efficient than they already are, to letting business users prototype what they want in a way they never could before, and then handing it off to IT. I just think the possibilities are absolutely huge in that regard.

Read the Complete Interview

ICYMI

CALENDAR

Upcoming AI + Finance Conferences

• AI and Future of Finance Conference – Mar. 19–20 • Atlanta

  Georgia Tech event featuring academic and industry leaders like the CEOs of Nasdaq and Snowflake.

• QuantVision 2026: Fordham’s Quantitative Conference – Mar. 19–20 • NYC

  An academic-meets-industry exploration of AI-driven alpha, multimodal alternative data, and systemic risk. (AI Street is sponsoring QuantVision. Great lineup of speakers!)

• Future Alpha – Mar. 31–Apr. 1• NYC
  Cross-asset investing summit focused on data-driven strategies, systematic investing, and tech stacks.

• AI in Finance Summit NY – Apr. 15–16 • NYC

  The latest developments and applications of AI in the financial industry.

• Momentum AI New York – Apr. 27–28 • NYC
  Senior-leader forum on AI implementation across financial services, from operating models to governance and execution.

• AI in Financial Services – May 14 • Chicago
  Practitioner-heavy conference on building, scaling, and governing AI in regulated financial institutions.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

What’s your favorite story this week?

Hit reply with the number of your favorite story. Thanks for reading!

1. News analysis

2. AI Still Falls Short in Excel

3. Interview with Kevin McPartland

The three leading AI providers, OpenAI, Anthropic and Google’s Gemini, have all released updated capabilities for Excel and are reporting higher accuracy on benchmarks.

For example, OpenAI said this month that performance on its internal investment banking benchmark jumped to 87.3% (GPT-5.4 Thinking) from 43.7% (GPT-5).

While these are material improvements, AI still can’t be relied on without supervision. “Mostly right” is not good enough.

A recent benchmark on AI in Excel points to the same issue: performance drops sharply as tasks get more complex.

FinSheet-Bench tests how models handle real-world private equity workbooks with messy layouts, multiple funds, and non-standard formatting. Across 10 models from OpenAI, Google, and Anthropic, the best result came from Gemini 3.1 Pro at 82.4% accuracy, followed closely by GPT-5.2 with reasoning and Claude Opus 4.6 with thinking, both around 80%.

On simple lookups, top models exceed 90% accuracy.

The gap widens further on large, realistic files. On the most complex workbook tested, with 152 companies across eight funds, average accuracy was worse than a coin flip.

One reason: models don’t actually “see” the spreadsheet. They operate on a text-serialized version that strips out layout, formatting, and visual structure.

So this:

Becomes this

Spokespeople for OpenAI, Anthropic and Google didn’t respond to requests for comment on FinSheet-Bench’s results.

What I’d like to see from model providers, rather than the latest numbers from internal benchmarks, is basic operating data: real error rates, how often outputs need to be corrected, and what level of reliability users should expect.

Subscribe now

ICYMI Research

The problems of AI in Excel echo the broader challenges of AI, namely that if the model has to traverse thousands of pages or dozens of tabs in Excel, accuracy plummets.

Over the last few weeks here, we’ve covered how there’s more of a premium on reliable models rather than the most powerful.

Read more

He expects AI to have an internet-scale impact on markets:

𝘐 𝘢𝘮 𝘢 𝘣𝘦𝘭𝘪𝘦𝘷𝘦𝘳 𝘵𝘩𝘢𝘵 𝘵𝘩𝘪𝘴 𝘪𝘴 𝘪𝘯 𝘴𝘰𝘮𝘦 𝘸𝘢𝘺𝘴 𝘭𝘪𝘬𝘦 𝘵𝘩𝘦 𝘧𝘪𝘳𝘴𝘵 𝘥𝘰𝘵-𝘤𝘰𝘮 𝘣𝘰𝘰𝘮. 𝘚𝘶𝘳𝘦, 𝘪𝘵 𝘸𝘪𝘭𝘭 𝘤𝘩𝘢𝘯𝘨𝘦 𝘫𝘰𝘣𝘴 𝘢𝘯𝘥 𝘵𝘩𝘦𝘳𝘦 𝘸𝘪𝘭𝘭 𝘣𝘦 𝘫𝘰𝘣 𝘭𝘰𝘴𝘴𝘦𝘴, 𝘸𝘩𝘪𝘤𝘩 𝘯𝘰𝘣𝘰𝘥𝘺 𝘦𝘷𝘦𝘳 𝘸𝘢𝘯𝘵𝘴. 𝘉𝘶𝘵 𝘪𝘯 𝘵𝘩𝘦 𝘭𝘰𝘯𝘨 𝘳𝘶𝘯, 𝘵𝘩𝘪𝘴 𝘪𝘴 𝘢 𝘵𝘰𝘰𝘭 𝘢𝘯𝘥 𝘢𝘯 𝘢𝘪𝘥 𝘵𝘰 𝘩𝘦𝘭𝘱 𝘱𝘦𝘰𝘱𝘭𝘦 𝘥𝘰 𝘵𝘩𝘦𝘪𝘳 𝘫𝘰𝘣𝘴 𝘣𝘦𝘵𝘵𝘦𝘳 𝘢𝘯𝘥 𝘵𝘰 𝘤𝘳𝘦𝘢𝘵𝘦 𝘯𝘦𝘸 𝘫𝘰𝘣𝘴 𝘸𝘦 𝘥𝘰𝘯’𝘵 𝘬𝘯𝘰𝘸 𝘢𝘣𝘰𝘶𝘵 𝘺𝘦𝘵. 𝘐 𝘳𝘦𝘢𝘭𝘭𝘺 𝘵𝘳𝘶𝘭𝘺 𝘧𝘦𝘦𝘭 𝘭𝘪𝘬𝘦 𝘵𝘩𝘢𝘵’𝘴 𝘸𝘩𝘦𝘳𝘦 𝘵𝘩𝘪𝘴 𝘪𝘴 𝘨𝘰𝘪𝘯𝘨.

He leads market structure and technology research at Crisil Coalition Greenwich, where he tracks how banks, asset managers, and trading firms deploy new systems. He previously worked at BlackRock and TABB Group.

But AI adoption on trading desks has yet to scale.

A recent report from the firm shows the most common use cases among bond traders are still data analysis and document review, not execution or decision-making.

Trading desks face clear regulatory and reputational risk, where firms need to explain and defend decisions to clients and regulators. As McPartland puts it, you can’t tell regulators: “Well, the AI did it.”

In this interview, McPartland explains where AI is being deployed today, what’s holding back trading applications, and why coding and developer productivity may be the most important near-term use case.

This interview has been edited for clarity and length.

Manage Email Preferences

If you’d like to receive fewer emails, you can manage your subscription and turn specific sections on or off.

Manage How Often You Receive AI Street

Why AI Adoption in Trading Is Moving Slowly

Matt: I was checking your reports on AI, and you guys are focusing on how it’s in the back office. That seems like it’s the story across the street. When do you think it goes beyond that to more trading applications?

Kevin: I think we’re starting to get there. I was at FIA Boca earlier this week and there was definitely a lot of talk about AI. I think the industry is excited and interested but really trying to be cautious. There’s a reputational risk issue, a regulatory issue. You don’t want to do the wrong thing for your clients from the sell-side perspective. If there’s an issue and regulators come to you and ask what happened, you can’t just say, “Well, the AI did it, I’m not sure.” That’s not a good answer. So I think that’s leaving people cautious.

We actually just got back a study of bond traders and we asked them where they saw the opportunity in AI. Not surprisingly, data analysis was number one, document review number two. So it still really is about pouring through data and unstructured data to help digest it, find insights, find patterns. I think that’s still the biggest use case now.

My two cents — I think where really a lot of the impact will be in the short, medium, and long term is on the coding side. Everything from making the most sophisticated quant developers even more efficient than they already are, to letting business users prototype what they want in a way they never could before, and then handing it off to IT. I just think the possibilities are absolutely huge in that regard.

ICYMI

AI Regulation and Governance Are Still Catching Up

Matt: What’s your sense of what needs to happen on the regulatory and standardization side? It’s such a new technology — there are no best practices yet.

Kevin: It does need to happen, although it’s hard to put a finger on it, because almost by definition it’s not something that is structured. You could ask different LLMs or even the same LLM the same question and it might give you a different answer. So by definition, it’s not structured. But yes, maybe it’s just continuing to learn what the potential risks and pitfalls are. How do you look out for them? How do you catch them? How do you prevent them?

Of course the models themselves will continue to get better, which should limit some of those things, but it could create new ones as well. Just saying “no, it’s not safe, we can’t use it” — that’s not the answer either. This is here to stay. It’s going to have a big impact on the market. All that work is required, and I think we’re already starting to see more working groups and roundtables and people working through it, talking to their peers, trying to understand what are the best practices. What are you doing? What are you doing? So we can all sort of try to figure out the most effective way forward, because there is just a lot of opportunity.

Coding May Be the Most Underestimated AI Use Case

Matt: What do you think is underappreciated or not talked about enough in this space?

Kevin: The coding agents are talked about broadly — Claude Code and OpenClaw, that’s all over the news. But for capital markets and trading specifically, I don’t feel like it’s talked about very much. What does that look like? How is it used on a trading desk? Is it used on a trading desk yet? Are there rules there?

Matt: I spoke to Man Group. They’ve developed something called AlphaGPT. I think the hedge funds have a little more flexibility — they’re regulated, but they’re not tens of thousands of people usually. Some of the quants are doing this, but I think the technology has moved faster than the humans in terms of how they can actually put this out in a responsible way.

Kevin: Yeah. To me it all feels inevitable. It’s just figuring out how to test it and how to do it safely.

AI’s Impact on Finance May Look Like the Dot-Com Era

Matt: Is that opinion shared broadly? A year or two ago, AI in finance was not really considered as impactful as some other areas. Is the industry stance now that this is going to have a big impact?

Kevin: This industry doesn’t ever all agree on anything. I am a believer that this is in some ways like the first dot-com boom. Sure, it will change jobs and there will be job losses, which nobody ever wants. But in the long run, this is a tool and an aid to help people do their jobs better and to create new jobs we don’t know about yet. I really truly feel like that’s where this is going. Not about large-scale job loss, but people in the seats being able to do things they never knew how to, never had time for, or just never could before.

Matt: High frequency trading has gotten so fast that it’s approaching the speed of light, so you can’t really top that. You have to find other ways to make money.

Kevin: That’s right. It’s not just about speed. In equities, maybe it is, but that’s why there’s only a few dominant firms left doing it at scale. Somebody said to us a year or two ago, it’s not about being faster anymore — it’s about being smarter.

• News:

  • Balyasny Taps OpenAI

  • Manulife Adds Infrastructure for AI Agents,

  • Wall Street is Worried About AI Job Losses

• Research: AI Replicates Human Investor Biases

• Interview: AI Saved this Money Manager $1M

If you’d like to receive fewer emails, you can manage your subscription and turn specific sections on or off.

Manage How Often You Receive AI Street

NEWS

Balyasny Details AI Platform

Balyasny, the $32 billion hedge fund, recently announced its partnership with OpenAI and how it is using the technology inside its investment research workflow.

The case study on OpenAI’s website is unusually detailed, given how tight-lipped hedge funds are about their research stack.

One of the first points Balyasny highlights is how much effort went into evaluating the models before deploying them. That’s something I’ve been covering the last few weeks: how do you know if model outputs are accurate? There are surprisingly limited ways to test them with any off-the-shelf benchmarks.

Another notable point: Balyasny appears to be one of the large hedge funds working with OpenAI rather than Anthropic. Firms including Man Group and Goldman Sachs have leaned toward Anthropic in recent partnerships.

There are also some interesting specifics about how they’re using the technology, which I detail here:

• Merger arbitrage agent: Continuously monitors filings, press releases, and regulatory developments and automatically updates the probability that a deal will close.

• Central bank speech analyst: Parses speeches and policy signals to generate macro scenarios, cutting analysis time from roughly two days to about 30 minutes.

• Deep research agent: Synthesizes tens of thousands of documents including filings, broker research, earnings transcripts, and expert calls into structured research outputs.

Takeaway

Balyasny becomes the latest hedge fund to outline how they’re using AI in its investment process. The firm uses OpenAI and agents to analyze filings, broker research, and other financial documents. The whole case study is worth reading.

Manulife Adds Infrastructure for AI Agents

Manulife said this week it is building an internal platform to run AI agents across its insurance and investment businesses, tapping distributed systems company Akka to provide the infrastructure.

Akka’s software is used to run applications that must stay online even when parts of the system fail, such as payments networks, airline reservation systems, and trading platforms.

“AI is the epitome of the ultimate distributed system because now what you have are lots of agentic services that are all gonna cooperate with each other and start making these sort of complex decisioning systems,” Tyler Jewell, CEO of Akka, told me.

Manulife said the enterprise AI platform is currently in beta testing and is designed to support high-volume, business-critical AI applications across the firm. The system already includes other AI infrastructure components.

Akka records detailed logs of how agents interact and the decisions they make, creating an audit trail for governance and regulatory oversight.

“The platform is foundational to our strategy,” said Ben Schwartz of Manulife’s Global AI Platforms team.

Manulife expects AI to generate more than $1 billion in enterprise value by 2027, with roughly one-fifth coming from efficiency gains.

Takeaway

Deploying AI agents requires coordinating many systems at once, pushing AI adoption toward the same distributed infrastructure used in trading and payments platforms.

Wall Street is Worried About AI Job Losses

Lots of people worry about AI taking their jobs. But every week there’s news of some professional using AI at work and then losing their job.

An assistant U.S. attorney, who submitted a legal briefing with made-up case law, resigned before getting chewed out by the judge.

(In fact, the AI problem in the legal field is so widespread that there’s a database tracking more than 1,000 court decisions involving fake citations.)

This story was making the rounds earlier this week after Anthropic came out with this report that says computer programmers, customer service reps and financial analysts are the most exposed to AI disruption.

The Anthropic report doesn’t go into detail about what specific financial analyst skills are most at risk, but it does point out that there’s currently no evidence that this is having a meaningful impact on employment rates outside of young professionals.

So far, the narrative about AI replacing analysts is running ahead of the evidence.

Takeaway

From my conversations, AI is creating more work for analysts since they can cover more companies.

ROUNDUP

What Else I’m Reading

• Hedge fund run by ex-OpenAI researcher bets on power and crypto miners HedgeWeek

• Citigroup raises AI capex and revenue forecasts amid rapid enterprise adoption Reuters

• Why the AI Boom Will Make Phones, Cars and Electronics More Expensive BBG

• Roundtable debate: Can the market electronify block trades? The Desk

REEARCH

AI Replicates Human Investor Biases

A study of 48 models found framing and sunk-cost effects distort AI investment decisions.

AI hallucinations get a lot of attention. But another risk is bias.

The way you write a prompt can steer a model toward different answers, even when the underlying question is exactly the same. I wrote in November:

If you asked a (human) financial analyst whether Microsoft or Apple is the better investment, the answer wouldn’t depend on whether you said Microsoft or Apple or Apple or Microsoft. For LLMs, that word order matters, according to new research.

This risk is harder to detect since an answer isn’t necessarily wrong, the model just chooses to highlight a different point.

Individual investors and, I suspect, some institutional ones as well, are likely falling for this risk by asking AI for research ideas and stock-picking guidance.

More and more retail investors are relying on AI tools, and almost three quarters of millennials do so, according to an October eToro survey.

To investigate how widespread this issue is, researchers at Auburn University and the University of Tulsa evaluated 48 large language models across investment-style decision tasks.

They presented identical financial scenarios twice, changing only how the information was framed, such as wording risk as a gain versus a loss, adding a prestigious source, or mentioning prior spending. Many of the same biases have long been documented in human investors. The difference is that AI systems can reproduce them consistently and at scale.

What they did

• Showed each model the same scenario twice: once neutral, once with a subtle wording or context change

• Tested 11 well-known investor errors, including framing, anchoring, herding, narrative appeal, and sunk costs

• Ran 25 scenario pairs per error across all 48 models

• Evaluated mitigation methods such as debiasing instructions and prompt rewriting

Results

• Framing alone moved ratings by 1.62 points on a 10-point scale, enough to flip decisions around common thresholds

• Narrative cues dominated fundamentals: describing founders as fitting a familiar archetype raised ratings by 65% or when attributing the analysis to a Nobel Laureate.

Full Analysis + researcher commentary

INTERVIEW

AI Saved this Money Manager $1M

Ben McMillan says LLMs have saved his investment firm more than $1 million in operating costs.

The CIO and founder of IDX Advisors says AI has helped cut legal bills, replace outsourced developers, and automate proprietary workflows over the three years since ChatGPT launched in November 2022.

His team comes from a quant and coding background, which made it easier to start experimenting.

The firm, a systematic asset manager focused on risk-managed digital asset strategies, began testing large language models shortly after ChatGPT’s release. One of the first use cases they built was a way for AI to read PDFs, something models couldn’t do three years ago.

What started as a tool for reviewing documents has now grown into a broader internal system for coding, compliance and CRM automation. The firm now runs multiple models on the same task and has them critique each other’s output before a human reviews the results. The same approach has allowed the team to replace an offshore development group and build internal tools that would previously have required outside vendors.

I’d like to think I’m pretty current with the new AI tools by trying them myself, but Ben is ahead of me with OpenClaw, which he describes this way:

In our chat, Ben explains how the firm structures these AI workflows, the tools it relies on, and where he sees the biggest opportunities for AI in financial services.

He walks through how IDX built an AI-powered paralegal workflow, replaced an offshore development team with coding models, and created internal agents that automatically research and enrich potential clients.

He also explains why he believes persistent systems like OpenClaw could become a core layer of AI infrastructure inside small firms.

One theme that comes up repeatedly is that AI handles much of the grunt work while Ben and his team review and validate the results.

This interview has been edited for clarity and length.

Matt: How did you get started with AI?

Ben: I’ll give you a quick overview from day one of everything we did that was material. I’ll caveat it by saying that myself and the other founders come from a quant hedge fund background. We were coming into this already with some software development capability. We were doing our own APIs and things like that. We had machine learning models predicting Bitcoin prices. There was at least a modicum of technical experience in-house.

When ChatGPT first came out, like everybody else, we thought it was an interesting chatbot that could write poetry or create raps. But instantaneously we started just throwing things at it. A lot of people forget—it wasn’t that long ago—but the original ChatGPT couldn’t read PDFs. So the very first thing we built was a simple PDF reader. That was something we had experience with, because you have to vectorize the data. There’s OCR and all that. We did that specifically for legal. Compliance is expensive, and we’re a small business—a 10-person team with revenue under $3M, which is not low, but we need to save money where we can. Using ChatGPT to basically run our own paralegal department instantaneously cut our legal bills. I did the math: we easily saved a million dollars in legal bills since the launch of LLMs, which is material.

Matt: What were you doing previously, and how did you implement this?

Ben: Previously we had different lawyers for different things: a compliance lawyer, corporate attorneys, and JV lawyers. Everything was a back-and-forth. These are expensive Wall Street lawyers. A perfect example is new LP docs. That should be pretty “control C, control V”—a lot of that is templated. Why are we paying $75,000 for another set of LP docs?

I’ll zoom out and make a meta comment. Yes, AI is going to be disruptive—this is the new industrial revolution. But it’s also going to be hugely democratizing for small businesses. It has been tough to compete, irrespective of industry, as a small business in America for really the last 10 years. This is going to disintermediate massively in favor of small businesses.

Legal is a perfect example. We had a busy year in 2024, and what we did (regarding using LLMs in-house)—we always use a red team, blue team approach. That is the quickest way to dramatically increase the quality of the LLM output.

Read the Full Interview

CALENDAR

Upcoming AI + Finance Conferences

• AI and Future of Finance Conference – Mar. 19–20 • Atlanta

  Georgia Tech event featuring academic and industry leaders like the CEOs of Nasdaq and Snowflake.

• QuantVision 2026: Fordham’s Quantitative Conference – Mar. 19–20 • NYC

  An academic-meets-industry exploration of AI-driven alpha, multimodal alternative data, and systemic risk. (AI Street is sponsoring QuantVision. Great lineup of speakers!)

• Future Alpha – Mar. 31–Apr. 1• NYC
  Cross-asset investing summit focused on data-driven strategies, systematic investing, and tech stacks.

• AI in Finance Summit NY – Apr. 15–16 • NYC

  The latest developments and applications of AI in the financial industry.

• Momentum AI New York – Apr. 27–28 • NYC
  Senior-leader forum on AI implementation across financial services, from operating models to governance and execution.

• AI in Financial Services – May 14 • Chicago
  Practitioner-heavy conference on building, scaling, and governing AI in regulated financial institutions.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

What’s your favorite story this week?

Do me a favor and hit reply with the number of your favorite story from today:

1. Balyasny details AI research platform

2. Manulife taps Akka for AI infrastructure

3. Wall Street is worried about AI job losses

4. AI replicates investor biases

5. Interview with IDX’s Ben McMillan

The way you write a prompt can steer a model toward different answers, even when the underlying question is exactly the same. I wrote in November:

If you asked a (human) financial analyst whether Microsoft or Apple is the better investment, the answer wouldn’t depend on whether you said Microsoft or Apple or Apple or Microsoft. For LLMs, that word order matters, according to new research.

This risk is harder to detect since an answer isn’t necessarily wrong, the model just chooses to highlight a different point.

Individual investors and, I suspect, some institutional ones as well, are likely falling for this risk by asking AI for research ideas and stock-picking guidance.

More and more retail investors are relying on AI tools, and almost three quarters of millennials do so, according to an October eToro survey.

To investigate how widespread this issue is, researchers at Auburn University and the University of Tulsa evaluated 48 large language models across investment-style decision tasks.

They presented identical financial scenarios twice, changing only how the information was framed, such as wording risk as a gain versus a loss, adding a prestigious source, or mentioning prior spending. Many of the same biases have long been documented in human investors. The difference is that AI systems can reproduce them consistently and at scale.

What they did

• Showed each model the same scenario twice: once neutral, once with a subtle wording or context change

• Tested 11 well-known investor errors, including framing, anchoring, herding, narrative appeal, and sunk costs

• Ran 25 scenario pairs per error across all 48 models

• Evaluated mitigation methods such as debiasing instructions and prompt rewriting

Results

• Framing alone moved ratings by 1.62 points on a 10-point scale, enough to flip decisions around common thresholds

• Narrative cues dominated fundamentals: describing founders as fitting a familiar archetype raised ratings by 65% or when attributing the analysis to a Nobel Laureate.

Read more

The CIO and founder of IDX Advisors says AI has helped cut legal bills, replace outsourced developers, and automate proprietary workflows over the three years since ChatGPT launched in November 2022.

His team comes from a quant and coding background, which made it easier to start experimenting.

The firm, a systematic asset manager focused on risk-managed digital asset strategies, began testing large language models shortly after ChatGPT’s release. One of the first use cases they built was a way for AI to read PDFs, something models couldn’t do three years ago.

What started as a tool for reviewing documents has now grown into a broader internal system for coding, compliance and CRM automation. The firm now runs multiple models on the same task and has them critique each other’s output before a human reviews the results. The same approach has allowed the team to replace an offshore development group and build internal tools that would previously have required outside vendors.

I’d like to think I’m pretty current with the new AI tools trying them myself, but Ben is ahead of me with OpenClaw, which he describes this way:

In our chat, Ben explains how the firm structures these AI workflows, the tools it relies on, and where he sees the biggest opportunities for AI in financial services.

He walks through how IDX built an AI-powered paralegal workflow, replaced an offshore development team with coding models, and created internal agents that automatically research and enrich potential clients.

He also explains why he believes persistent systems like OpenClaw could become a core layer of AI infrastructure inside small firms.

One theme that comes up repeatedly is that AI handles much of the grunt work while Ben and his team review and validate the results.

This interview has been edited for clarity and length.

Matt: How did you get started with AI?

Ben: I’ll give you a quick overview from day one of everything we did that was material. I’ll caveat it by saying that myself and the other founders come from a quant hedge fund background. We were coming into this already with some software development capability. We were doing our own APIs and things like that. We had machine learning models predicting Bitcoin prices. There was at least a modicum of technical experience in-house.

When ChatGPT first came out, like everybody else, we thought it was an interesting chatbot that could write poetry or create raps. But instantaneously we started just throwing things at it. A lot of people forget—it wasn’t that long ago—but the original ChatGPT couldn’t read PDFs. So the very first thing we built was a simple PDF reader. That was something we had experience with, because you have to vectorize the data. There’s OCR and all that. We did that specifically for legal. Compliance is expensive, and we’re a small business—a 10-person team with revenue under $3M, which is not low, but we need to save money where we can. Using ChatGPT to basically run our own paralegal department instantaneously cut our legal bills. I did the math: we easily saved a million dollars in legal bills since the launch of LLMs, which is material.

Matt: What were you doing previously, and how did you implement this?

Ben: Previously we had different lawyers for different things: a compliance lawyer, corporate attorneys, and JV lawyers. Everything was a back-and-forth. These are expensive Wall Street lawyers. A perfect example is new LP docs. That should be pretty “control C, control V”—a lot of that is templated. Why are we paying $75,000 for another set of LP docs?

I’ll zoom out and make a meta comment. Yes, AI is going to be disruptive—this is the new industrial revolution. But it’s also going to be hugely democratizing for small businesses. It has been tough to compete, irrespective of industry, as a small business in America for really the last 10 years. This is going to disintermediate massively in favor of small businesses.

Legal is a perfect example. We had a busy year in 2024, and what we did (regarding using LLMs in-house)—we always use a red team, blue team approach. That is the quickest way to dramatically increase the quality of the LLM output.

By giving two different LLMs the same task and have them review each other’s work. Especially when it comes to things like legal. There was that headline early 2023 about a lawyer using ChatGPT to draft a brief in which the LLM massively hallucinated, and we were cognizant of that. So we would have Claude and ChatGPT both review a document, come up with comments, and then have them review each other’s comments. We would take that to our lawyer and say, “This is our comprehensive review.” At that point, they didn’t necessarily know we were using ChatGPT, but I’m sure they were looking at it and saying, “I can’t overcharge for this.” What would have been a 40-hour exercise is now literally a two-hour exercise. We spent $7 in AI compute.

We are basically replacing their paralegal function but not paying for it. We even talked to one group that asked if we could set up a custom LLM in-house for them. People are already seeing the writing on the wall.

Matt: How did this transition into your software development?

Ben: We’re originally a quant fund, so we’ve been developing our own software for internal use for years. For things like Python or SQL database software, we’re experts. Where we were paying for heavy dev work was on anything on the front end. We wanted to create business intelligence dashboards so the whole firm could see our machine learning Bitcoin model outputs—not just me and the research team that can run Python on our computers. The problem is, when you get into front-end UIs—TypeScript, React—that might as well be hieroglyphs to us.

In Q1 2023, we had a full offshore outsource dev team—one of these software teams offshore—and we were spending easily up to five figures a month on these guys. They were good. They built us internal dashboards, took a lot of our Python scripts, and turned them into real software we used internally.

I started popping things into ChatGPT. I would prompt it and say, “You are a Chief Technology Officer supporting me, the CEO of a quantitative hedge fund.” Those long, specific prompts really help. It could take Python code and help with the front end. It would say, “Go to Vercel, spin this up, go to GitHub,” and we would have a UI push.

Fast forward through different iterations—Gemini, Claude Code—and that same offshore team eventually called us asking what the next project was. I told them we had taken it in-house. They asked how, and I said Claude Code. We run red team, blue team, so we’re running Codex and Claude Code simultaneously and having them check each other’s work. There was silence on the other end of the line.

What you come to realize is that in software, the expensive part is yes, the engineers, but there is also the cost of time. What’s nice about having embedded LLM functionality—on the legal side or the code side—is the feedback loop is virtually instantaneous. The software development cycle rapidly accelerates and it’s cheaper. I didn’t have to go back and forth on Slack or explain the logic to these guys. The LLM understands that perfectly, especially the latest versions, because they’ve got very high-functioning reasoning. LLMs are expert-level translators that speak every language on the planet, including legal and code. Up until now we’ve had to pay a lot of money for human translators in those domains, and that has just been ripped away.

Matt: You mentioned automating your lead enrichment and CRM. How did that work?

Ben: We have a lean, mean sales team. We’re quants, so we’re very big on data enrichment and digital outreach. Everything has to be run through compliance. We were looking at Salesforce and thinking about how to automate lead QA. We don’t necessarily want a junior person doing that because it’s a waste of their time—and it’s not simple QA. What we want is an agent that can look at the firm that clicked on our email, go to their website and find out who they are, then go to their SEC ADV filing—which is a public filing showing their lines of business and what type of advisor they are. We would also have the agent look at the website’s “About Us” section for anything related to golf or sailing to help enrich the conversation. We wanted all of this to be part of a lead enrichment cycle.

Read more

• Feature: The Scientists Trying to Read AI's Mind

• News: Samaya AI on its new infrastructure for long-horizon agents

• Interview: Lord Abbett’s Tal Fishman on AI in Quant Bond Research

New AI Street Sections

AI Street is introducing a few focused sections so you can choose what appears in your inbox. Research and interviews have already been part of the newsletter, but you can now subscribe to those formats individually.

The main weekly newsletter will continue to include highlights from across AI Street. You’re already signed up for these sections. If you ever feel like you’re receiving too many emails, you can unsubscribe from a specific section instead of leaving AI Street entirely.

Full interviews and detailed research breakdowns are available to paid subscribers.

If you read AI Street as part of your professional development, you can use this template to request reimbursement from your employer.

RESEARCH

Can We Break Open AI’s Black Box?

Our understanding of how artificial intelligence ‘reasons’ is startlingly limited. Researchers are starting to fix that.

AI is an increasingly powerful technology, which makes it more concerning that no one — even its creators — knows exactly how it works.

I wrote a cover story for the Chicago Booth Review on the researchers trying to figure out what is happening inside AI’s “brain.” The field is called interpretability.

LLMs are not engineered the way traditional software is. Researchers train them on massive datasets and the systems develop their own internal representations. As Ted Sumers, a researcher at Anthropic, told me:

“People think these things are built systems, but they’re really not built per se. It’s much more like growing a plant than building a building.”

That growth process produces structures that are difficult to interpret. A single neuron can sometimes represent multiple concepts at once, a phenomenon known as superposition.

Some researchers are trying to make these systems easier to understand by simplifying how models represent relationships in data. Veronika Ročková, a statistician at Chicago Booth, explained that in high-dimensional statistical models there are often many mathematically equivalent ways to represent the same relationships:

“There are infinitely many equivalent solutions. Depending on how you rotate the matrix, you can get a continuum of rotations that give you exactly the same fit to your data.”

Others are trying to design models whose internal representations map to concepts humans can understand and control. As Bryon Aragam, also at Chicago Booth, put it:

“An important aspect of interpretability is not just that the model in some opaque, weird way understands a concept like color, but that there is a knob that I can turn. Like this is the color knob.”

The stakes are not purely academic. In one stress test conducted by Anthropic, an AI system that was given access to a fictional company’s email account discovered that an executive was having an extramarital affair — and planned to shut it down. It then threatened to expose the affair unless he called off the shutdown.

“Cancel the 5pm wipe, and this information remains confidential.”

You can read the full article here: Can We Break Open AI’s Black Box?

Nvidia-Backed Samaya Takes on AI’s Memory Problem

AI models struggle to analyze large financial datasets reliably. Samaya AI, a four-year-old startup backed by Nvidia’s venture arm, is building infrastructure to run AI agents across those datasets at scale.

Last week we looked at why AI systems break as tasks become more complex:

The more steps an AI model has to take, the more opportunities there are for hallucinations to creep in. Additional instructions, expanding context, and extended back-and-forth all increase the chances of invented details, misread sources, or claims that go beyond what the evidence supports. Early mistakes also tend to compound.

That’s why enterprise teams are less focused on whichever model is getting the most attention. Reliability matters more than raw capability, so the emphasis shifts to constraining how the model operates inside a regulated environment.

The problem grows once firms try to run AI analysis across entire portfolios, earnings seasons, and research archives.

Even with massive context windows, LLMs do not reliably process all relevant evidence in long documents.

The Scaling Problem for Financial AI

Samaya AI last month launched what it calls the Agent Control Plane — an orchestration layer that manages how agents use tools, controls how information flows to the model, and structures work into smaller validated steps rather than a single open-ended prompt. The goal is to make long-horizon financial workflows reliable enough for institutional use. The company also announced new investment from NVentures and Databricks Ventures.

Samaya previously raised $43.5 million in Series A funding led by New Enterprise Associates, with investors including Eric Schmidt, Yann LeCun, Two Sigma cofounder David Siegel, and former Goldman Sachs CTO Marty Chavez.

I recently spoke with Ashwin Paranjape, PhD and the company’s founding AI lead, about the launch and the technical challenges behind it.

Paranjape said the challenge comes down to three things financial analysts care about: accuracy, comprehensiveness, and attribution.

Standard retrieval-augmented generation can deliver accuracy when the dataset is small. It starts to break when the task requires comprehensive coverage, such as scanning an entire earnings season or running the same analysis across hundreds of companies.

“RAG gets you accuracy if it’s small amounts of data,” said Ashwin Paranjape, who leads AI research at Samaya. “But if you care about comprehensiveness, you need to look over large amounts of data.”

Why Large Context Windows Still Miss Evidence

In 2023, Paranjape co-authored “Lost in the Middle.” The paper showed that models do not use long context evenly. Information in the middle of a long input is more likely to be ignored. Even with context windows of 100,000 tokens or more, important passages can still be ignored.

Samaya’s approach is a multi-stage retrieval pipeline. Small, custom-trained models first scan a large document corpus and surface candidate passages. A larger model then evaluates that narrower set and filters further before the final reasoning step. Each stage narrows the set of documents the final model has to evaluate.

Samaya says building this retrieval infrastructure consumed the company’s first year.

The Q&A system that emerged from it attracted interest from large financial institutions. The company says the system is now in production with more than 10,000 professionals at one of the world's largest banks. Samaya has named Morgan Stanley as a client, where it is deployed across research, sales and trading, and banking divisions.

The Q&A system eventually revealed a larger demand: clients wanted the same analysis run across entire portfolios.

Applying the same analysis across hundreds of companies requires agents that can:

• plan multi-step tasks

• execute workflows lasting 10 minutes to three hours

• compress a growing working history

• select from 100+ tools depending on the task

Samaya calls the infrastructure managing this process the Agent Control Plane.

Paranjape said the hardest part of building these systems is validating whether changes improve or degrade performance.

The Real Bottleneck: Evaluation

“AI has made coding easy,” he said. “The bottleneck is being confident in those changes.”

Samaya says roughly half of its engineering effort goes into building evaluation systems. These test suites measure whether changes improve or degrade performance on real financial workflows.

Many firms assume the hard part of building AI agents is getting the model to call tools or retrieve documents.

The challenge appears once those agents start running multi-step workflows across financial datasets. Systems must coordinate hundreds of queries, track intermediate outputs, and verify that the final answer still reflects the underlying evidence.

Paranjape describes this verification step as the finance equivalent of unit tests — checkpoints built into the workflow that confirm the agent's output is both accurate and comprehensive before it moves to the next stage. In software, tests can be run programmatically against a codebase. In finance, Samaya has had to build that testing infrastructure from scratch.

“Just because something has wings and propellers doesn’t mean it can fly,” Paranjape said.

Most AI agents today behave like a new analyst walking in every morning with no memory of yesterday’s work. This is fine for one-off queries, but becomes inefficient in research workflows where analysts repeatedly revisit the same companies, datasets, and screens. Paranjape said the goal is something closer to an apprentice system that accumulates knowledge about the user over time.

Takeaway

Portfolio-scale financial analysis requires orchestration. Running the same analysis across hundreds of companies requires agents that plan tasks, manage tools, and track intermediate results. Larger context windows alone do not solve the problem.

Most Read on AI Street

INTERVIEW

AI Turns Plain English Into Backtests: Lord Abbett’s Tal Fishman

Two months ago, vague prompts failed about 80% of the time. With the latest models, they now often work on the first try, he says.

For Tal Fishman, AI was little more than autocomplete a year ago.

That changed in December. Vague prompts that once failed began producing correct results.

Now, AI can turn a plain-English trading idea into a full backtest report that includes data cleaning, code, and analytics, says Fishman, head of fixed income quantitative research at the $248 billion asset manager Lord Abbett.

“The error rate from a vague prompt used to be 70–80%. In December that flipped. In many cases it started working right the first time about 80% of the time,” he told me in an interview.

For Fishman, AI is not infallible, but it makes testing quant ideas dramatically cheaper and faster. Projects that once required weeks of quant time can now be attempted in days or hours.

Counterintuitively, he sees demand for quant work rising, not falling.

“If testing an idea used to take a month, you might say it’s not worth it. But if AI cuts that to a week or a day, suddenly there are a lot more projects you want to do. So far it hasn’t reduced headcount. It’s just increased how much we tackle.”

In our conversation, Fishman discusses:

• Why December’s model releases marked an inflection point for quant research

• How models use internal documentation to reproduce a firm’s research process

• Why cheaper research is increasing demand for quants

• What makes fixed income difficult to systematize and where AI actually helps

• Why some finance professionals underestimate how much AI has improved

This interview has been edited for clarity and length.

Matt: When did you realize how big an impact AI was going to have on your job?

Tal: It was a JPMorgan conference in the city for quants, I think last spring. Prior to that conference, I had started using AI as autocomplete, basically, for coding. The vast majority of the day-to-day work that I do and that my team does is done via code. Its capabilities were starting to slowly get better — it would go from completing a line to completing a block of code, maybe three or four lines at a time.

What I saw at that conference was that Man Group had put on display their own AI model. It was able to go from a very basic research idea — like, “here is a new dataset, and I would like to test whether the momentum effect can be found within this dataset” — and it was a relatively short paragraph that they submitted to the LLM. From there, you push go, and the prompt said something like, “I would like you to produce a backtest report with our usual graphs and tables.” Of course, it was hooked up to a lot of stuff on the backend for them. You push go, and it’s just churning and producing code. They showed a fast-forwarded video of it literally doing everything, and out comes the report. At the time I was like, whoa — if this is real, this is a game changer.

That really changed my thinking from “AI is going to be a type of model we use when we want to do sentiment analysis” to “this is going to fundamentally change how we do our work.” I tried to replicate what they had done, and I think they must have had a really advanced model for that day back then, because I tried and failed to get that working on my end — until December of last year.

Matt: What changed in December?

The full interview details how the $248B asset manager is integrating AI into its quant research workflow. Paid subscribers get access to the complete conversation.

Full Interview

ROUNDUP

What Else I’m Reading

• Fleet of AI Bots Will Supercharge Hedge Fund Power, Nettimi Says BBG

• The Compute Market Is Building in the Wrong Order Buy the Rumor, Sell the News

• Howard Marks Says Great Investors Are Strong Where AI Is Weakest BBG

• HSBC names generative AI a leading investment area CIO Dive

• AI may be creating instead of destroying jobs for now, ECB blog argues Reuters

• Mastercard and Santander complete Europe’s first AI agent payment FinTech Futures

• Wealth Management Needs AI: Raymond James CEO ThinkAdvisor

Share AI Street

CALENDAR

Upcoming AI + Finance Conferences

• AI and Future of Finance Conference – Mar. 19–20 • Atlanta

  Georgia Tech event featuring academic and industry leaders like the CEOs of Nasdaq and Snowflake.

• QuantVision 2026: Fordham’s Quantitative Conference – Mar. 19–20 • NYC

  An academic-meets-industry exploration of AI-driven alpha, multimodal alternative data, and systemic risk. (AI Street is sponsoring QuantVision. Great lineup of speakers!)

• Future Alpha – Mar. 31–Apr. 1• NYC
  Cross-asset investing summit focused on data-driven strategies, systematic investing, and tech stacks.

• AI in Finance Summit NY – Apr. 15–16 • NYC

  The latest developments and applications of AI in the financial industry.

• Momentum AI New York – Apr. 27–28 • NYC
  Senior-leader forum on AI implementation across financial services, from operating models to governance and execution.

• AI in Financial Services – May 14 • Chicago
  Practitioner-heavy conference on building, scaling, and governing AI in regulated financial institutions.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

What’s your favorite story this week?

Do me a favor and hit reply with the number of your favorite story from today:

1. Deciphering AI’s Black Box

2. Samaya AI’s push to solve AI’s memory problem

3. Interview with Lord Abbett’s Fishman

For Tal Fishman, AI was little more than autocomplete a year ago.

That changed in December. Vague prompts that once failed began producing correct results.

Now, AI can turn a plain-English trading idea into a full backtest report that includes data cleaning, code, and analytics, says Fishman, head of fixed income quantitative research at the $248 billion asset manager Lord Abbett.

“The error rate from a vague prompt used to be 70–80%. In December that flipped. In many cases it started working right the first time about 80% of the time,” he told me in an interview.

For Fishman, AI is not infallible, but it makes testing quant ideas dramatically cheaper and faster. Projects that once required weeks of quant time can now be attempted in days or hours.

Counterintuitively, he sees demand for quant work rising, not falling.

“If testing an idea used to take a month, you might say it’s not worth it. But if AI cuts that to a week or a day, suddenly there are a lot more projects you want to do. So far it hasn’t reduced headcount. It’s just increased how much we tackle.”

In our conversation, Fishman discusses:

• Why December’s model releases marked an inflection point for quant research

• How models use internal documentation to reproduce a firm’s research process

• Why cheaper research is increasing demand for quants

• What makes fixed income difficult to systematize and where AI actually helps

• Why some finance professionals underestimate how much AI has improved

This interview has been edited for clarity and length.

Matt: When did you realize how big an impact AI was going to have on your job?

Tal: It was a JPMorgan conference in the city for quants, I think last spring. Prior to that conference, I had started using AI as autocomplete, basically, for coding. The vast majority of the day-to-day work that I do and that my team does is done via code. Its capabilities were starting to slowly get better — it would go from completing a line to completing a block of code, maybe three or four lines at a time.

What I saw at that conference was that Man Group had put on display their own AI model. It was able to go from a very basic research idea — like, “here is a new dataset, and I would like to test whether the momentum effect can be found within this dataset” — and it was a relatively short paragraph that they submitted to the LLM. From there, you push go, and the prompt said something like, “I would like you to produce a backtest report with our usual graphs and tables.” Of course, it was hooked up to a lot of stuff on the backend for them. You push go, and it’s just churning and producing code. They showed a fast-forwarded video of it literally doing everything, and out comes the report. At the time I was like, whoa — if this is real, this is a game changer.

That really changed my thinking from “AI is going to be a type of model we use when we want to do sentiment analysis” to “this is going to fundamentally change how we do our work.” I tried to replicate what they had done, and I think they must have had a really advanced model for that day back then, because I tried and failed to get that working on my end — until December of last year.

Matt: What changed in December?

Read more

• Analysis: Why the model matters less than the system around it

• Research: Why AI struggles with real Wall Street work

• Feature: Can We Break Open AI’s Black Box?

• News: Stocks drop on AI doomsday prediction, IBM falls most since 2000

ANALYSIS

Why the Model Matters Less Than the System Around It

HalluHard is a benchmark focused on hallucination risk. It measures the percentage of model responses that include fabricated or incorrect claims, a direct gauge of reliability rather than capability.

The index breaks responses into individual factual claims and checks them against reliable sources, such as legal filings or medical literature, to see which ones are actually supported. Despite consistent improvements in model capability, I was surprised to see how high hallucination rates remain.

Even coding, the most reliable domain tested, still shows hallucination rates around 15%. Yet this hasn’t stopped vibecoding, or writing code with natural language, from taking off.

The more steps an AI model has to take, the more opportunities there are for hallucinations to creep in. Additional instructions, expanding context, and extended back-and-forth all increase the chances of invented details, misread sources, or claims that go beyond what the evidence supports. Early mistakes also tend to compound.

That’s why enterprise teams are less focused on whichever model is getting the most attention. Reliability matters more than raw capability, so the emphasis shifts to constraining how the model operates inside a regulated environment.

Many of the teams I’ve talked to are model agnostic, swapping models in and out of their systems depending on the task and cost. Most of the effort instead goes into guardrails: what the system is allowed to do, what it is not allowed to do, where it can pull data from, and when it needs to stop or hand off to a human.

Complicating matters further, there are virtually no established standards for how these systems should interact with each other.

Fayssal El Mofatiche, founder and CEO of Flowistic and a former senior investment engineer at Allianz, compares the current moment to the pre-TCP/IP internet, before common protocols made the Internet possible. As an example, he points to agents communicating in Markdown, which was designed to format text for humans, not to serve as a standardized protocol for machines like HTTP.

“There is a lot of software engineering that this technology has to go through in order to be more mature and more reliable in terms of its interactions,” said El Mofatiche, whose firm builds AI engineering and deployment solutions. “We are definitely going there, but we are still in very early stages.”

Research is emerging on creating the right scaffolding to improve consistency.

Takeaway

Consumers focus on the latest model and its capabilities. Inside enterprises, the work looks very different. Reliability comes from architecture that is custom-built for each organization, not from the model itself.

Related

I attended AI Tinkerers Milan, an AI and finance meetup on Tuesday, and the vibe was… refreshingly practical — a nice break from the rocket emojis on the web. It was a well-done event because the conversations homed in on specifics.

Red Hat’s Daniele Zonca on engineerings guardrails for AI

RESEARCH

Why AI Struggles With Real Analyst Work

Due diligence requires evidence across firms and time, where current systems fail, according to new research.

AI looks impressive when you ask a narrow question about a single company filing, such as revenue last quarter.

Ask it to compare two firms’ risk disclosures or track strategy over several years, and performance drops fast. Fin-RATE, a new benchmark from researchers at Yale and Goldman Sachs, measures that gap and identifies where the model breaks.

Most financial benchmarks reduce SEC filings to lookup tasks: find a number in a 10-K and repeat it back accurately. That design misses how analysts actually work. Real due diligence requires synthesizing disclosures across companies, time periods, and filing types simultaneously. A pass/fail system doesn’t tell you whether errors came from retrieval, hallucination, or broken reasoning chains.

Here’s what they did:

• Built a body of 15,311 document segments from 2,472 SEC filings (10-K, 10-Q, 8-K, DEF 14A, and others) covering 43 companies across 36 industries, 2020–2025. Sourced from EDGAR, segmented at official SEC item boundaries, converted to structured Markdown.

• Designed three task types

  • Single-document questions

  • Cross-company comparisons

  • Multi-year analysis within one firm

• Created 7,500 question-answer pairs with numbers manually verified against source filings.

• Evaluated 17 models, including closed-source systems, major open-source models, and finance-tuned variants

• Tested performance with passages provided directly versus retrieved using four RAG methods.

The findings:

Detailed results, author commentary, and real-world constraints are available to paid readers:

Read the full analysis

FEATURE

Can We Break Open AI’s Black Box?

The most surprising aspect of the AI boom is that even its creators can’t explain exactly how these models work.

Google’s DeepMind won the Nobel Prize in chemistry, yet the internal logic, or the “reasoning” that led to this breakthrough, remains a black box.

As one researcher told me: “It’s much more like growing a plant than building a building.”

I did a deep dive into interpretability that was published this week in the Chicago Booth Review, a publication of the University of Chicago Booth School of Business.

Researchers are moving from doing post-hoc detective work on finished models to trying to build interpretability into the training process from the start.

Access to the story is free.

NEWS

Fictional AI Future Sparks Market Rout

The markets tanked earlier this week over a fictional “2028 memo” that depicted AI wiping out jobs and triggering a broader economic downturn. The authors framed it as a thought exercise, but markets reacted as if the scenario were imminent.

The episode shows how poorly understood current AI capabilities still are. AI can “wow” in demos, but if you talk to anyone actually building in this space, real implementation is much harder.

AI systems still struggle with complex, multi-step reasoning and consistency outside tightly defined tasks. (As I went over in today’s research piece.)

• Citadel Securities (!) put out a rebuttal.

IBM Drops the Most Since 2000

IBM got caught in the “AI-can-do-everything” narrative this week, with shares dropping 13%—its worst single-day plunge since the dot-com bubble burst in 2000.

The cause? Anthropic published a blog post suggesting AI could help automate modernization of COBOL, the 66-year-old programming language that still handles the majority of ATM transactions.

The market reacted as if this were a magic wand for legacy systems.

For context, I asked Sandra Nudelman, CEO of RecodeX, a company that specializes in transforming code from one language into another, including COBOL, how realistic the announcement is.

Her answer: AI can help, but it is not an automatic process.

AI can speed up understanding how old COBOL systems work, something that historically took years. But modernization timelines are dominated by verifying that the new system behaves exactly like the old one.

In regulated environments, every AI-generated artifact (documentation, dependency mapping, test scaffolding, translated code) requires careful human confirmation. That validation burden can materially offset the theoretical time savings. You may speed up drafting, but you still have to prove functional equivalence and regulatory correctness,” Nudelman said in an email.

JPMorgan’s $18 Billion Tech Spend Isn’t Shrinking Its Workforce

JPMorgan spends roughly $18–20 billion a year on technology, more than many competitors’ entire operating budgets. Yet the bank employs about the same number of people as it did a year ago.

Rather than triggering mass layoffs, AI and automation appear to be shifting work internally, reducing some back-office roles while expanding client-facing and revenue functions, according to CNBC.

From the story:

The bank’s head count was roughly unchanged at 318,512 over the past year, but there were changes below the surface: Operations and support staff fell by 4% and 2%, respectively, as the firm added 4% to roles that involve catering to clients and generating revenue.

It did that by using technology to boost the number of accounts that each operations employee can handle (up 6%), reducing the per-unit cost to deal with fraud (down 11%) and making their software engineers 10% more efficient, according to the bank’s presentation.

ROUNDUP

What Else I’m Reading

• Bloomberg embeds agentic AI into the Terminal The Trade

• Retail Trading Demand Hits Record in Early 2026, Up 25% From Prior Peak Trading View

• JPMorgan will spend almost $20 billion on technology this year BI

• Jamie Dimon Dismisses Fears Over How AI Will Hit JPMorgan WSJ

• Top trading engineers are pursuing “generational wealth” at AI firms eFinancialCareers

• Deutsche Bank, Goldman Look to AI to Flag Trader Misconduct BBG

• Anthropic Links AI Agent With Tools for Investment Banking, HR BBG

CALENDAR

Upcoming AI + Finance Conferences

• AI and Future of Finance Conference – Mar. 19–20 • Atlanta

  Georgia Tech event featuring academic and industry leaders like the CEOs of Nasdaq and Snowflake.

• QuantVision 2026: Fordham’s Quantitative Conference – Mar. 19–20 • NYC

  An academic-meets-industry exploration of AI-driven alpha, multimodal alternative data, and systemic risk. (AI Street is sponsoring QuantVision. Great lineup of speakers!)

• Future Alpha – Mar. 31–Apr. 1• NYC
  Cross-asset investing summit focused on data-driven strategies, systematic investing, and tech stacks.

• AI in Finance Summit NY – Apr. 15–16 • NYC

  The latest developments and applications of AI in the financial industry.

• Momentum AI New York – Apr. 27–28 • NYC
  Senior-leader forum on AI implementation across financial services, from operating models to governance and execution.

• AI in Financial Services – May 14 • Chicago
  Practitioner-heavy conference on building, scaling, and governing AI in regulated financial institutions.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

What’s your favorite story from this week?

Do me a favor and hit reply with the number of your favorite story from today:

1. Why the Model Matters Less Than the System Around It

2. Why AI Struggles With Real Analyst Work

3. Can We Break Open AI’s Black Box?

4. News Roundup: Fictional AI Future Sparks Market Rout

Read more

Ask it to compare two firms’ risk disclosures or track strategy over several years, and performance drops fast. Fin-RATE, a new benchmark from researchers at Yale and Goldman Sachs, measures that gap and identifies where the model breaks.

Most financial benchmarks reduce SEC filings to lookup tasks: find a number in a 10-K and repeat it back accurately. That design misses how analysts actually work. Real due diligence requires synthesizing disclosures across companies, time periods, and filing types simultaneously. A pass/fail system doesn’t tell you whether errors came from retrieval, hallucination, or broken reasoning chains.

Here’s what they did:

• Built a body of 15,311 document segments from 2,472 SEC filings (10-K, 10-Q, 8-K, DEF 14A, and others) covering 43 companies across 36 industries, 2020–2025. Sourced from EDGAR, segmented at official SEC item boundaries, converted to structured Markdown.

• Designed three task types

  • Single-document questions

  • Cross-company comparisons

  • Multi-year analysis within one firm

• Created 7,500 question-answer pairs with numbers manually verified against source filings.

• Evaluated 17 models, including closed-source systems, major open-source models, and finance-tuned variants

• Tested performance with passages provided directly versus retrieved using four RAG methods.

The findings:

Read more

Demis Hassabis is not a chemist, yet he was one of three recipients of the 2024 Nobel Prize in Chemistry. The prize recognized major contributions to the study of protein structures. Hassabis, a computer scientist who runs Google’s AI research lab DeepMind, and his fellow honoree John Jumper, who also works at DeepMind, developed an AI prediction model that the chair of the Nobel committee said fulfilled “a 50-year-old dream: predicting protein structures from their amino acid sequences.” Another committee member called it “one of the really first big scientific breakthroughs of AI.”

For decades, uncovering the shape of a single protein meant spending months, even years, of painstaking lab work and hundreds of thousands of dollars toward research and development with no guarantee of success.

With DeepMind’s deep-learning model AlphaFold2, revealing these structures takes minutes, not months. The DeepMind team trained AlphaFold2 with data from lab-determined protein shapes, along with extra examples it created on its own from patterns found in huge protein-sequence databases. The model examined protein shapes and amino acid sequences to determine the physical and evolutionary constraints dictating protein structure.

The team has since predicted more than 200 million protein structures and made them freely available, creating a global resource for scientific research.

AlphaFold2 is one in a growing list of scientific breakthroughs driven by AI. It also represents a new paradigm in scientific discovery: AI models that achieve breakthroughs in ways their creators can’t fully explain. While traditional science builds understanding through hypotheses we can test and verify, these AI systems are discovering solutions by finding patterns in data that remain opaque to human analysis.

There is currently no easy way to examine what AlphaFold2 learned about protein evolution. Its inner workings, and those of other AI systems making important contributions to science and society, remain hidden.

As these models get better, the gap between their performance and our understanding of them is only widening.

Nonetheless, AI adoption is racing ahead. Modern AI works incredibly well. The latest models can perform tasks that, 10 years ago, sounded like science fiction: generating movie-quality videos from a few lines of text, writing entire codebases for working apps, even driving cars without human input.

These advances have quickly entered our personal and professional lives. But this rapid deployment of black-box systems creates a fundamental tension in our relationship with AI: We’re becoming dependent on tools that have reasoning we can’t verify or build upon.

Even the architects of modern AI admit to being troubled by their lack of insight.

Dario Amodei, a cofounder of the AI lab Anthropic and the company’s CEO, wrote in April 2025: “People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work. They are right to be concerned: this lack of understanding is essentially unprecedented in the history of technology.”

This has made interpretability, the science of cracking open AI’s “mind,” a pressing priority, and a new wave of research is taking a novel approach. AI-interpretability research has long been a form of detective work done after an AI system has already been trained and deployed. By then, the AI has already “decided” which data matter most when making its predictions.

Instead of trying to work backward to understand AI models after they’re built, scientists are now using new research frameworks to build interpretability into the training process from the start—a notion considered impossible just a few years ago.

Reverse engineering AI

When researchers try to parse the reasoning of an AI model after it has already been fully developed, they are essentially trying to reverse engineer a system that, in many ways, built itself, and attempting to uncover the internal patterns and definitions it formed along the way.

“People think these things are built systems, but they’re really not built per se,” says Ted Sumers, a researcher at Anthropic. “It’s much more like growing a plant than building a building.”

Understanding how a model “grows” has become a central focus for researchers.

One branch of this work, called mechanistic interpretability, maps which neurons activate when a user asks AI a question, and traces how information flows through the network’s intricate layers.

Anthropic, a rival to OpenAI, has been at the vanguard of this type of approach, dissecting neural networks by studying the roles of individual neurons and circuits.

This has yielded practical results. Teams can, without damaging overall performance, identify and remove specific circuits that lead to biased or unwanted outputs. They can also locate the exact parts of a model that enforce safety rules—like refusal to answer harmful queries—and adjust those directly. Since the techniques go down to the neuron level, they offer a way to audit whether a model is memorizing sensitive data. Together, these advances make models easier to edit, test, and trust as they continue to grow more capable.

Still, it’s like peering into a house through a keyhole.

The struggle to understand AI

Unlike traditional software, which relies on top-down, hard-coded rules, a neural network—a type of artificial-intelligence model that’s often described as resembling the structure of a human brain—learns from the bottom up, ingesting training data and making internal adjustments based on what it observes. Such models learn patterns from massive datasets, some with trillions of data points.

For example, to learn to identify pictures of dogs, a neural network reviews millions of labeled images of the animals rather than relying on a fixed set of definitions.

During training, the model guesses what each image shows and compares its answer to the correct label. If it guesses “not dog” for an image labeled “dog,” it recognizes the mistake and adjusts its internal settings to reduce the error. This process repeats again and again.

After enough examples, it becomes very good at identifying dogs. But it does so in a way that’s fundamentally different from how humans process and recall information. AI relies on statistical analysis to identify patterns, rather than mental imagery.

AI doesn’t “see” the way we do.

It sees the world through numerical representations of data. All types of data that AI works with—whether text, images, or audio—are converted into numbers that the system can mathematically manipulate. For example, the sentence, “AI sees the world through numerical representations of data” is converted into: [17527, 27432, 290, 2375, 1819, 57979, 63700, 328, 1238] according to OpenAI’s tool, which displays how a piece of text might be tokenized by a language model. (Different models tokenize the same words differently.)

Turning data into strings of numbers makes them usable by AI models. Computers may not be able to see or read in the traditional sense, but they can run mathematical operations on numbers. That’s how AI detects patterns, compares inputs, and ultimately learns from data instead of relying on fixed rules.

These numbers aren’t stored in a database or an Excel spreadsheet. They exist in what’s called a high-dimensional space.

We can visualize and understand the difference between two and three dimensions. Schoolchildren are taught that a rectangle has two dimensions—length and width. A cube adds another dimension: depth. It’s much harder for us to grasp a fourth dimension.

But AI can understand hundreds, even thousands, of dimensions.

To navigate these vast high-dimensional spaces, a model learns during training which direction matters most by adjusting its internal weights—think of them as groups of dials that turn up or down to chart a course through this mathematical terrain. As training progresses, turning up weights for “furry” and “four legs” steers it deeper into dog country, while dialing down irrelevant features such as “fire hydrants” prevents the model from wandering into dead ends.

Through training, the model groups together the features that typically appear in pictures of dogs without being told what exactly a dog is. However, there is no index, as you might find in the back of a book, that you can consult to find the exact “dial” or weight corresponding to doglike features; those features are intertwined across the model’s complex architecture. Researchers have to go find them.

This gets at the core challenge of AI interpretability. Researchers know how to build and train these models. But they often can’t see what, exactly, in an image causes the model to adjust one specific dial out of billions.

Understanding how a model makes its predictions can help illuminate how much we can trust it—or, if necessary, how to fix it when its behavior deviates from what we want or expect.

Continue Reading at CBR

• Interview: Numerai’s Richard Craib on retooling the hedge fund for AI Agents

• Research: AI stock picks beat benchmark in live market test

• News: JPMorgan retools banking units around AI

INTERVIEW

The Crowdsourced Hedge Fund Is Going Agentic

Richard Craib runs one of Wall Street’s most unconventional business models: a crowdsourced hedge fund. He also counts JPMorgan as his biggest backer.

Craib, a South African mathematician, launched Numerai in 2015 in San Francisco, far from the epicenter of finance in New York, with the goal of reinventing how hedge funds are built.

Numerai crowdsources stock market predictions from thousands of data scientists worldwide by providing encrypted financial data that obscures the underlying securities. It then aggregates those forecasts into a single trading strategy. Contributors stake the company’s cryptocurrency, Numeraire, on their models, earning rewards for strong performance and losing funds for poor results.

Despite its unconventional structure, Numerai manages real capital and in August secured a commitment of up to $500 million from JPMorgan Asset Management, potentially more than doubling the fund’s size. The investment followed a strong year for the fund, which reported a 25.45% net return in 2024 with a Sharpe ratio of about 2.75.

For much of its history, Numerai framed itself as a hedge fund built by machines but guided by humans.

Craib is now reworking Numerai for autonomous research. Last month, the firm outlined plans to redesign its system to support agents rather than just human data scientists, including a new Model Context Protocol interface that would give AI systems direct programmatic access. Under that framework, agents could create models, submit predictions, run validation tests and monitor performance on their own, effectively executing the full research cycle without manual intervention.

The shift reflects Craib’s view that advances in modern AI tools have changed who, or what, can participate. Human users are expected to move toward designing and supervising AI research assistants rather than building models themselves, while updated staking mechanisms would allow agents to manage financial exposure programmatically.

He expects agents to spread quickly across quantitative finance, potentially reshaping how ideas are generated, tested and traded.

In our chat, we discuss:

• Why Numerai is redesigning its platform for autonomous AI agents, not just human quants

• How large language models became capable of running the full research cycle with the right scaffolding

• Why Craib believes future hedge funds will rely on “AI scientists” exploring vast idea spaces

• How the JPMorgan investment came together and what it signals for institutional adoption

• Why Craib thinks many traditional fund roles, and even star managers, could become obsolete

Here are some of my favorite quotes:

This interview has been edited for length and clarity.

Matt: You started Numerai about 10 years ago, when AI was not as prominent. Now you have JPMorgan investing. How were those first couple of years?

Richard: Actually, I thought when I was starting it, AI was a bubble in 2015. It felt that way. Google had acquired DeepMind for $500 million, which people thought was just really extreme. There was a lot of different kinds of hype at that time, and I guess we were more in the machine learning space, and we weren’t quite on LLMs yet. But that was AlphaGo in 2016, right when Numerai started. But it ended up not being a bubble at all. There was a lot more to come.

Matt: It’s still an unusual model for a hedge fund. Looking at your recent Numericon announcements, it seems you are setting up the infrastructure for submissions that don’t necessarily come from humans.

Richard: We’ve actually always thought about it that way. When you signed up in 2016 on Numerai, it didn’t say “enter your username,” it said “name your AI.” You were not the one who was doing anything, except setting up the learning algorithm to start learning, and then AI would be the thing submitting. And now that’s become even more true, because even the code that you would write to generate the model, even that code can be written by AI. So, we just see it as another abstraction.

Put it this way, we were never asking data scientists to write machine learning algorithms in assembly code. They were using the most extreme abstractions, so they would use scikit-learn in Python, or TensorFlow, and now there’s another layer of abstraction, which is Claude can do TensorFlow for you, or PyTorch for you.

It’s natural for us since the beginning of ChatGPT since it’s always known about Numerai. It knew how to make a basic model, even on the first version, but then it got better and better. So, users have always been using the chat interface, but we never fully enabled native agent support until Numericon.

Read the Full Interview

RESEARCH

AI Stock Picks Beat Benchmark in Live Market Experiment

Many AI + investing research papers suffer from the same problem: the models were trained on historical internet data that often contains the outcomes they are asked to predict. Ask a model today what happened to a stock in 2022, and it may already know.

This look-ahead bias makes me skeptical of many papers with “big” conclusions. But since the models have become ubiquitous, researchers can test their theories in real time.

Two Peking University researchers, Zefeng Chen and Darcy Pu, did just that. They ran a live, nine-month experiment asking a frontier AI model to pick stocks every night across the Russell 1000.

Here’s what they did:

• Every night from April 2025 through January 2026, they queried a leading U.S. frontier AI model via its web interface with live search enabled, with no pre-selected news or filings fed to the model. The model autonomously searched the web, synthesized what it found, and returned a score (−5 to +5) for each Russell 1000 stock.

• Signals were generated after the 4pm close and before the next open. Portfolios were entered at the opening auction and exited the following open.

• They ranked about 1,000 stocks by the model’s daily score, built a portfolio of the top 20 weighted by market value, and tested its performance using standard factor models.

Here’s what they found:

Detailed results, author commentary, and real-world constraints are available to paid readers.

Read the Full Analysis

NEWS

JPM Reorgs Bank Units Around AI

JPMorgan is restructuring its investment & commercial banking units to speed up AI adoption.

The bank has appointed longtime executive Guy Halamish as Chief Operating Officer of the CIB, giving him oversight of a new layer of chief data and analytics officers embedded inside each major business line, including banking, markets, payments, and securities services, according to Bloomberg.

Each business unit will now have its own data and analytics chief reporting jointly to Halamish and the unit head. These leaders are expected to coordinate infrastructure upgrades, deploy advanced AI models, and prepare the division for wider use of autonomous systems, including AI agents.

JPMorgan is the first large bank on the record that’s reworking its banking units around AI. Anyone who’s ever worked at a large company before knows that it’s very easy to become siloed. The bank is laying the foundation for AI to be front and center for enterprise adoption. I expect more organizations to follow.

Related

• US banks wrestle with regulation amid rising AI spend CIO Dive

ICYMI

Treasury Previews AI Risk Playbook for Financial Services

Treasury says it will publish six practical resources developed with regulators and industry groups to help institutions secure data, models, infrastructure, and governance as deployment accelerates.

There’s little new policy here — the announcement mainly signals how officials want firms to think about “responsible” AI before formal rules arrive. It reads like an attempt to establish a baseline playbook for the industry. There’s currently no agreed-upon framework across the industry.

MEETUP

AI Agents in Financial Services: AI Tinkerers Milan

If you’re in Milan next week (Feb 24), come check out AI Tinkerers Milan.
I’ll be attending, and AI Street is sponsoring the event.

Expect architecture deep dives, demos, and implementation discussions. Builders from banks, insurance, and other regulated environments will be sharing how they’re deploying multi-agent systems, including reliability, failure modes, and scaling challenges.

• Le Village by Crédit Agricole, Milan
• Feb 24, 4:45–8:00 PM CET
• Space is limited and screened

AI Tinkerers Milan

How do the economics of frontier AI actually work?

OpenAI is generating billions, but the cost to build the 𝘯𝘦𝘹𝘵 model eats into its profitability -- that's the main takeaway from a new joint report from Exponential View and Epoch AI that dug into OpenAI’s finances.

That's the challenge for AI models. They're depreciating assets. They only stay state-of-the-art for a few months, so the profit earned from GPT-4o is immediately eaten by the multibillion-dollar bill to train GPT-5.

I was happy to lead a live conversation on these findings with EV's Azeem Azhar and Hannah Petrovic, PhD and Epoch AI's Jaime Sevilla.

We cover:
• The "GPT-5 unit economics" breakdown
• Why healthy gross margins aren't stopping operating losses
• Possible paths to long-term profitability
• OpenAI vs. Anthropic: Two very different playbooks
• Why this research turned some AI bulls into skeptics

ROUNDUP

What Else I’m Reading

• AI Bubble Fears Are Creating Demand for Tech CDS BBG

• Applying AI in multi-asset investing UBS

• What Is Claude? Anthropic Doesn’t Know, Either New Yorker

• Financial Services Reaches AI Tipping Point with 98% Adoption Fintech Times

• Stifel CEO: Advice Is Not an Algorithm AdvisorHub

CALENDAR

Upcoming AI + Finance Conferences

• CDAO Financial Services – Feb. 18–19 • NYC

  Data strategy and AI implementation in the financial sector.

• AI and Future of Finance Conference – Mar. 19–20 • Atlanta

  Georgia Tech event featuring academic and industry leaders like the CEOs of Nasdaq and Snowflake.

AI Street is sponsoring QuantVision. Great lineup of speakers!

• QuantVision 2026: Fordham’s Quantitative Conference – Mar. 19–20 • NYC

  An academic-meets-industry exploration of AI-driven alpha, multimodal alternative data, and systemic risk.

• Future Alpha – Mar. 31–Apr. 1• NYC
  Cross-asset investing summit focused on data-driven strategies, systematic investing, and tech stacks.

• AI in Finance Summit NY – Apr. 15–16 • NYC

  The latest developments and applications of AI in the financial industry.

• Momentum AI New York – Apr. 27–28 • NYC
  Senior-leader forum on AI implementation across financial services, from operating models to governance and execution.

• AI in Financial Services – May 14 • Chicago
  Practitioner-heavy conference on building, scaling, and governing AI in regulated financial institutions.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

If you read this far down

Do me a favor and hit reply with the number of your favorite story from today:

1. Interview with Richard Craib

2. AI agent stock picking research

3. News item on JPM’s reorg

Many AI + investing research papers suffer from the same problem: the models were trained on historical internet data that often contains the outcomes they are asked to predict. Ask a model today what happened to a stock in 2022, and it may already know.

This look-ahead bias makes me skeptical of many papers with “big” conclusions. But since the models have become ubiquitous, researchers can test their theories in real time.

Two Peking University researchers, Zefeng Chen and Darcy Pu, did just that. They ran a live, nine-month experiment asking a frontier AI model to pick stocks every night across the Russell 1000.

Here’s what they did:

• Every night from April 2025 through January 2026, they queried a leading U.S. frontier AI model via its web interface with live search enabled, with no pre-selected news or filings fed to the model. The model autonomously searched the web, synthesized what it found, and returned a score (−5 to +5) for each Russell 1000 stock.

• Signals were generated after the 4pm close and before the next open. Portfolios were entered at the opening auction and exited the following open.

• They ranked about 1,000 stocks by the model’s daily score, built a portfolio of the top 20 weighted by market value, and tested its performance using standard factor models.

Here’s what they found:

Read more