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

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.

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

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.

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

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

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 NumerCon 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 NumerCon.

Matt: What made you decide to focus more on this approach? When did it click?

Richard: In November, there was a tipping point that everyone in Silicon Valley felt. Models like Claude and ChatGPT Codex became capable of doing almost anything if you provided the right scaffolding.

That was the moment where it was like, okay, well, now we should really just lean into this, because you get the feeling that everyone will be here soon.

In the beginning of Numerai, there was a popular statistical programming language called R, and that was actually very popular, maybe half and half users used that. But then it moved to Python, PyTorch, almost completely, and I think it’s the same thing with this. You’re just gonna see very quickly people feeling they’re doing it wrong if they’re not using agents.

Matt: So, you see this more as the scaffolding and architecture. From the B2C side, it’s about which model does what, but for enterprises, they’re more concerned with the framework—the tracks it runs on.

Richard: This is the key thing. It’s not super well understood, but if we were to hire a PhD that was super smart, he would still make the basic errors in the first few weeks, because he wouldn’t know how to do a proper cross-validation backtest on financial time series data. And that’s the same with Claude. No matter how smart it gets, it doesn’t have Numerai Skills, which is the skills.md file we made. So, when you ask it, do a whole bunch of research on Numerai, sort of within the first hour, it’ll make three bad mistakes. It’s not really its fault, it’s not because it’s dumb, it just doesn’t know quite how we do things. And so, once it’s got access to the skills.md, it’ll be like, oh, well, if I need to do that, I’m just—I have to use the skills way of doing it. And so that’s how the scaffolding gets defined really nicely.

Matt: So, it’s tailored directions. And it sounds like you’d rather have good scaffolding with a lesser model than a good model without it.

Richard: Yeah, exactly. And the 'Skills' feature was released by Anthropic basically only two or three months ago. If you don’t have that, it’s almost like you haven’t had onboarding. It’s like Citadel University where you spend a month before they let you do anything. You have a whole bunch of training, and so you’re building up skills about how the organization works, how they do things, before they let you touch any production code. So that’s a training camp for AIs.

Matt: Have you heard of Man Group’s AlphaGPT?

Richard: Yeah, I don’t know what it really is, but I’ve seen that they’ve made announcements about it.

Matt: I spoke with Man Group’s Ziang Fang about it. He described it as an end-to-end idea test machine. Whatever passes through certain thresholds, the humans look at it. They’ve said the AI-generated ideas are passing their human benchmarks on the tests—it just lets them test more ideas and do more than they could before. Is that a similar concept to Numerai?

Richard: It’s one thing to say we have good infrastructure to test ideas. But it’s really, can you get to the point where you can span the full space of possible ideas? That’s the trouble with a quant signal. If you think about how many ways there are to order 6,000 stocks—because that’s what Numerai users are doing, ranking best to worst—there are 6,000 factorial. That’s 2 times 10 to the power of 200. There are more permutations than atoms in the universe. So, that’s why we still need the crowdsourcing, because we don’t know what to ask, or what model to build in the first place.

I think if you were 16 years old and you said, “Hey Claude, make me a quant fund from scratch,” and that’s the best prompt you could come up with, it would make you a very generic, mode of the distribution, quant fund infra. Whereas if you say, “I read this paper in 2023, which had these really strange ideas, and here’s the paper, and can you turn this into a Numerai idea in this way,” then that is a much more directed search into what’s possible. So that’s how we see users using this. They still have a role to play in, ‘This is the direction I want to go in,’ and I don’t want something average, because you don’t get paid for an average model on Numerai.

Matt: You’re looking for outliers.

Richard: We literally pay you for orthogonal alpha. So, if you make something crowded, that is the first thing Claude would come up with, and you’ll earn nothing.

Matt: HRT has said at some academic conferences that they’re training foundational models on financial time series—basically the language of financial data. What do you think about that approach?

Richard: We haven’t done that. I don’t think it’s that important. If you just train on the price time series, it’s crazy to say all you need is the price to predict the future. That’s a very 90s thing to say. You are getting paid by the market for adding strange new information to it, not the most commonly known information. So, we have 2,000 features, and not that many of them are based on price. You could maybe cobble something together, but that’s not the be-all, end-all.

The way I see it is more like these models are themselves AI scientists, and they weren’t a year ago. So why not just now run the scientific method more and more?

We have built language models. We built something called Numerai Predictive LLM, and we made it read news and then come up with a prediction from the news. That was an 8 billion parameter model. It actually doesn’t matter if you make it higher in terms of parameters. But that was a natural use case, because the current language models will not be able to predict from news what will happen, because they’re not trained to.

One example we give is, if you have a company like NVIDIA, and they make a press release that says, “we’re being investigated by the Department of Justice for monopolistic practices,” and the second paragraph is, “our revenue grew 150% year over year.” You ask ChatGPT, is this good or bad for the company? And ChatGPT will say, well, it’s neutral, because there’s some good things in it, and there’s bad. And actually, it’s extremely positive for the stock if you’ve fine-tuned the model. So, our model gets that right. It says this is amazing news, and the other models don’t. That’s one place where we are internally building features with language models, but it’s not on the time series of price.

Matt: How did the JPMorgan investment come about?

Richard: The thing about hedge funds is there’s quite a lot of short-term thinking, people want the first 3 years to be amazing. But we were like, let’s not even raise any LP money and raise venture capital, and then build something that no one can compete with in the long term. That’s why it’s been more of a tech company.

JPMorgan, the first meeting with them was something like 2018, 2019. We were probably below $100 million, maybe below $50 million, because they’ve had a lot of success investing in cutting-edge stuff. They’ve invested in early machine learning funds like Voleon and Voloridge, I believe.

In the early days, it was more like us saying, what do you want us to do to be fully institutional and ready for you? And they told us all the things they like. They like the 3-year track record, not a zero-year track record. They, in some ways, helped us make the product something that was institutional grade.

They’re not the first institutional investor—we have 3 endowments and many others besides. But they are the biggest one in terms of capacity, they want to invest $500 million.

Matt: Has that opened other doors?

Richard: Yeah, a lot. We now have a 6-year track record. We look very good compared to peers, and we’re getting better and better. Whereas other peers, maybe they got too big, and now they’re struggling to put up good numbers. But we’re snowballing, where our data is growing, our users are getting smarter, and everything’s kind of getting better, and risk management is getting better. So, yes, after the JP Morgan announcement, many of the big players have been reaching out to us. And probably in the next month or two, there might be other announcements, and we should be over at 1 billion quite soon. We’re almost $600 million, but $600 million is maybe 2 checks away from a billion.

Matt: You also have a different level of auditability and transparency compared to a typical hedge fund. There’s a lot more detail about how models are submitted and staked.

Richard: Yeah, and it’s interesting, because we’ve even had an endowment investor make a user account and submit a model. And he got to really see, okay, this makes sense. I didn’t understand this bit. And also, he got to see that he didn’t win. There were people who were a lot better than him. He sort of saw, okay, this talent is very good here. Anyone can watch our performance. Another thing you can watch that I like to watch is how well the metamodel is doing—the combination of all models—how well that’s doing against the benchmark models, which is just the free benchmarks we give away. But those models, we’ve tried our best to make them very good, so they’re our best internal model. And we said, here’s the baseline model, improve on it. Well, almost month after month, the edge widens, and it’s never looked as wide right now, where the crowd, the stake-weighted metamodel, is crushing the best we can do internally. Because the reality is, we’re good at data science. We have good data scientists. We’ve hired some of the top Numerai users over the years, but we don’t know how to beat everyone. And we don’t think we’ll ever beat everyone even with infinite AI scientist assistance.

Matt: Is that just the wisdom of the crowd?

Richard: No, I don’t even like that term, because the wisdom of the crowd is almost saying that the individuals are dumb, but the crowd is smart. I actually think it’s the opposite. It’s much more like an open source project, where about 1% of the users who’ve ever signed up to Numerai are the core contributors. And then the next 5% is also very important. Numerai, by asking people to stake their models, we are making it hard, on purpose, to do well. If you do badly, you will get your stake destroyed. So, Numerai is more like an API to find the best thousand data scientists in the world, versus let’s all make dumb guesses, and it’ll average out to something good.

Matt: So it’s more like winnowing it down to the best?

Richard: Yeah.

Matt: You’ve been building this a while. What’s been the most surprising thing?

Richard: The one thing I do think is true, and it makes total sense, is that the venture capital industry in this country is just amazing. We raised from the best VCs. Very quickly, they sort of saw a future where it’s like, okay, well, what if the way Millennium works is kind of gonna seem outdated in 2030? Where you hire all these people, and then pay them a lot, and then they read the newspaper and code. It’s weird. But the Numerai way was this new thing, and so it’s always been very easy for us to raise venture rounds. But I would say that the asset allocators, they’re more backward-looking. They say, well, Millennium has a 30-year track record. And we don’t trust AI yet. So that to me was quite tiring, in a way, to basically try to just educate, because when you heard about Numerai, there would often be three things you have to kind of know about. Blockchain, which no one knew about. Then machine learning. And then quantitative finance. So you had to have all three to like Numerai. You had to have a lot of knowledge of all three. And most people were kind of 1 out of 3. Now, I would say people are getting to 3 out of 3, because those are the technologies du jour.

Matt: What about the hype cycle? You’ve seen Numerai get labeled different things over the years.

Richard: Yeah, we’ve been a little bit bubble-averse. There was a similar time where people were talking about us as a blockchain company and hyping up our cryptocurrency. And I was just trying to put some cold water on that, because I just don’t want people to be disillusioned. That’s not really a hedge fund style. It’s supposed to be risk-adjusted, long-run. It’s not gambling.

Matt: Are you seeing more submissions since NumerCon?

Richard: Yeah, it has. NumerCon was less than a month ago, and there’s 150 to 200 MCP connections, and there’s only 500 staked users. They make many models per user. That’s surprising.

Matt: Are you going to get to the point where an agent is working for you and you’re just on the beach?

Richard: That’s the dream, but everybody has agents, too, including Numerai’s peer competitors. I really think that there are pods at Millennium that are paid $100 million a year with code that could be replicated in 40 hours by Claude. And so I don’t know what they do. I think that’s part of Numerai’s mission. We want fewer human beings in the hedge fund management industry. And I think we’ll get there, and Claude is helping.

Matt: What about the broader disruption to white-collar work? A lot of it turns out to be white-collar manual labor.

Richard: I think it’ll be looked back on in almost disgust by grandchildren. With trading, you don’t understand the human mind or intelligence if you think you can go out for breakfast at the St. Regis in New York, have coffee, and then on your way walking to work you’re like, “I should buy NVIDIA.” And then you go and buy it. It’s crazy that you have no information, except kind of the sort of amorphous blob of human thought.

Matt: But that’s not just the retail trader. You’re talking about hedge fund managers doing the same thing.

Richard: I’m worried more that it’s actually hedge fund managers who would be that person having coffee at the St. Regis, and they’d buy $100 million of NVIDIA based on their vibes. And you’re like, you know there’s 2,000 dimensions of data that Numerai has?

Matt: The majority of money managers underperform the benchmark.

Richard: At what point do you realize you had a lucky call, and it had nothing to do with you in some way? It was just an apparition in your mind? We’re very vulnerable to that type of thing.

Matt: Everyone on Wall Street wants to be the smart guy.

Richard: I’m not the smart guy, but I made a website to be friends with all the smart people.

An earlier version of this interview misspelled NumerCon.

Kirk McKeown spent about 15 years running what he calls the “factory”—some of the largest fundamental channel-check and data driven operations on the Street – first, at Glenview, and later, at SAC Capital and Point72. At its peak, his team was conducting several thousand calls each year. Kirk’s role evolved and ultimately, he ran all proprietary research at Point72 across calls and data. After years of managing this massive human-capital engine, he realized the “moat” in institutional finance was shifting from access to data towards the architecture used to structure it.

In 2021, he co-founded Carbon Arc, a platform built to structure data to be sold by consumption, as opposed to locking it up as a long-term asset. Carbon Arc unlocks data trapped on balance sheets and is fresh off a period of rapid institutional adoption. Now, Carbon Arc is betting that the future of alpha resides in a “refinery” capable of structuring 100 trillion transactions for the coming wave of 30 billion AI agents, solving problems not only on the Street, but for all types of businesses around the world.

I spoke with Kirk about his journey from the “manual” Wall Street factory to Carbon Arc’s “agentic” refinery. Here is what readers will learn:

• How the “several thousand calls per year” grind forged a mathematical framework for global market structures.

• Why Carbon Arc treats data as a derivative with time decay (Black-Scholes for data).

• The transition from “drilling” (data collection) to “refining” (knowledge graphs).

• Why the next 12 months belong to “automated agent onboarding” over retail chat.

This interview has been edited for length and clarity.

Matt: You had a long history in the hedge fund industry. What made you decide to jump ship and start something of your own?

Kirk: In 2012, I went to SAC to build what is now called Canvas. I had run a similar business at Glenview Capital, and built a large fundamental research business collecting information in supply chains. From 2006 to 2014, I did thousands of calls a year, myself. That kind of volume created for me strong principles around scaling research problems.

Research frameworks are patterns. The same story happens over and over. What’s happening in the U.S. government right now has happened three or four times. It’s just different names and different clothes. The world follows rules based on market structure, business models, management teams, and personality types.

I started to learn that hospitals and hotels are the same business because they both get paid on length of stay. I learned that TSMC and US Steel are the same business because while they have different end markets, they make money the same way. In the Global 2000, there aren’t 2,000 companies. I believe that there are four market structures and nine business models.

You start to find these scale points. For example, Tractor Supply is like a home center for rural areas. 25% of their business is animal feed and 25% is Texas. If you get a handle on animal feed in Texas, you get a handle on that business and can make a better risk-adjusted bet. I developed mental hacks and mathematical decision frameworks. It wasn’t because I was smart. It was because the “n” was so big.

By the time I left Point72, I was running proprietary research, managing the people responsible for generating actionable insights for the Firm’s investment teams to use as inputs in their process.

While working in research at these great firms, I started looking at the friction in the data market. It traded like 1930s equities: big block trades for bags of cash for market insiders with massive balance sheets. I looked at the legal and compliance frictions—it takes forever to get a data set approved. There were technical frictions. In 2016, Snowflake had been around for two years and Databricks had just come onto the scene. The infrastructure to manage this data at scale didn’t exist before 2010.

The pricing and commercial construct was built for 500 qualified buyers, not five million, so the clearing price was very high. If you could bring supply and demand closer together, smash down the cost of the insight, and sell the insight rather than the asset, you could achieve density and velocity of consumption.

Matt: How has the business evolved since you started it in 2021?

Kirk: We started building in 2021. Fast forward five years, and we have a two-sided consumption-based platform. Data asset owners bring their data, and we structure and graph it for the AI economy. We created composable infrastructure that allows people and agents to plug into the front of the stack. You can hit a modularized data structure to request an entity (like Lululemon), a framework (like revenue growth), or an asset (like credit card data). You compose that element and buy it for $5. We built an ontology that manages the modular analytical framework and a payment processor to meter it.

We started the stack in 2021, but when ChatGPT came along, we recognized we needed to be in graph. We tore down what we had built and started fresh. We rebuilt the platform as a knowledge graph. We have 100 trillion transactions structured in graphs. We modularize the entity structure around companies, brands, people, and locations.

For example, if someone wants to buy the average salary in 40,000 zip codes monthly to understand wallet structure, they can buy that aggregated from us. They don’t have to buy the whole data set for several hundred thousand dollars. I’m making a market for them and running the business like Goldman in the 90s.

Matt: What is happening to the price of data now that the cost of intelligence is decreasing?

Kirk: Companies like FactSet or Bloomberg have valuable data, but they don’t monetize it the right way. The structural commercial relationship between how agents interface with data and how they value it is changing. If you’re selling cases like Westlaw and an AI model ingests it once, they own it. Rewriting that licensing construct is hard because IP rules and laws were not built for agents.

Analytical platforms and database companies get hit because they get paid on compute. Compute is going to be socialized and optimized. It’s the wrong pole in the tent to get paid on. In the oil business, you don’t want to be a driller. You want to own the field or the refiner. Drilling is a bad business. Refining is a fixed-cost, high-volume framework. We are building a refiner.

Matt: How do your former colleagues on Wall Street react to these concepts like knowledge graphs and ontologies?

Kirk: I’ve been evangelizing this for a long time. This is just Wall Street from 1984 to 2025 on a truncated time horizon.

In 1973, the Black-Scholes paper was published. In 1983, Goldman launched the first quant desk, the beginning of the quant age on Wall Street. Between 1984 and 1990, early quant shops competed on models and saw 80% per annum alpha returns. Alpha degraded through the ‘90s as models competed. After the 1998 LTCM crash, ETFs formed because quants needed bigger liquidity profiles. Following the 2007 quant crash, factors came along, rates went to zero, and the factor market formed. Over that time, commissions went from $2 a share in 1983 to less than a penny today, while volumes went up 10,000x. In 1984, 50% of New York Stock Exchange trades were blocks. Today, it is 7%. Our stack is built to remove frictions to allow models to engage.

In 1985, when models started to proliferate, the traditional guys poo-pooed them. In 1995, when electronic trading came along, floor traders said it would never work because people liked talking to people. It’s the same thing as Blockbuster. There are resistors, but everyone uses data.

To us, OpenAI and Anthropic are just hedge funds. They are writing models to create lift in decisioning and competing on that lift. They are buying and selling scientists the way Millennium and Citadel do. I’d argue they are on the wrong capital structure—they should be raising GP/LP stakes rather than VC money. They don’t have a moat other than capital. Wall Street ends up winning the AI wars over the medium term because of regulation and their historical relationship with modeling the world.

When I worked at Point72, I had to find simple analogies to manage a big group. Data is a content business. You can’t own data end-to-end as an individual. You need engineers, scientists, analysts, and salespeople. Content must be relevant, differentiated, and accessible. Accessibility is asymptotic and relevance is table stakes. Differentiation is the only thing that separates us, and in content, that means more data and better questions.

OpenAI and Anthropic have trained on a relatively small amount of data, mostly scraped from the web. To manage the world’s inventories and inform global decisions, you need access to transaction data that shows how people spend their time and money, and their balance sheets. That is what Carbon Arc has built. We have 75 assets, three petabytes of data, and daily granularity for $150,000 a month in compute. We smashed the cost down. Now we are scaling both the supply and distribution sides.

Matt: Can large hedge funds or banks build this themselves, or do they face structural issues?

Kirk: They can build it, but they have a competitive problem. They monetize data through the market, so they won’t share their alpha back with data providers. We built a data transaction processor that creates liquidity for data providers and opens up their distribution. Data providers are coming to us because they can distribute broadly rather than doing one big exclusive check with a firm like Two Sigma, Citadel or D.E. Shaw.

Hedge funds want exclusive data and don’t want it proliferated. Data providers just want to get paid. Because data has historically been expensive and hard to work with, only global businesses and large funds could buy a million-dollar data asset. That market structure is what we are changing.

We launched platform 2.0 in mid-2025. We started last year with 35 customers and ended with 75, quadrupling revenues. Half of our customers are Wall Street buy-side and sell-side. We work with five of the top eight consulting firms, and we have good coverage in media and Hollywood. Companies like Paychex are both suppliers and customers. We are launching automated onboarding for agents on February 17th. We didn’t build this platform for eight billion people. We built it for 30 billion agents.

Matt: How do you see the market for small and mid-sized businesses (SMBs) and retail users?

Kirk: We are launching retail in March 2026. We will launch our MCP server for people with Robinhood or Kalshi accounts. We’re going on Reddit to offer a hedge fund data stack for $20 a month. For SMBs, a VP of Finance at a small healthcare business can pay $200 a month to do competitive analysis by plugging their Claude into our stack via the MCP to query credit card, paycheck, and healthcare claims data.

A consumption-based model needs volume. We give away publicly sourced data, like SEC data, for free. This is a cost game. Models are democratizing analytics. If you are building on top of public data and overpricing it, you will lose. We think about the business in terms of cost per megabyte and price accordingly.

Markets are forming for things that seem bizarre, like Kalshi’s contracts, but the real differentiation is composable contracts on anything, anytime, anywhere. We are in the third inning of a doubleheader. This technology cuts friction and makes things economically viable that weren’t before.

I am an AI bull, but I am concerned about the next 10 to 15 years. The dislocation in the labor market will take time to absorb. There are massive regulatory and ethical issues. Civilization-changing situations—like electricity in the 1880s or the rise of quants in the ‘80s—always involve these cycles.

Matt: Where does the “moat” for your business exist in the long term?

Kirk: The moat ends up being regulatory. Our General Counsel came from Schulte Roth & Zabel, the largest data compliance shop. She is standardizing compliance as a product. As agents proliferate, we must ensure legal and regulatory standards are met. Right now, it’s the Wild West, with major publishers suing Silicon Valley shops for scraping.

We are leading with scalable compliance frameworks. It’s like Stripe. You need regulatory infrastructure to scale. We model the business after Goldman Sachs. Someone once said to me that Goldman is a regulatory wrapper that allows you to do cool stuff in 100 countries and apply capital against it. They are an enablement platform that marries regulatory access and capital.

In the future, someone will emerge as the “Moody’s of data,” scoring inputs. Centralization will happen around core scale points, just as it did with Coinbase in blockchain. Maintaining a competitive advantage when things move this fast is hard. It’s an infrastructure build. Some people stand up application layer businesses on top of OpenAI and hit $10 million in revenue in six months, but that’s a gold rush. It’s not sustainable because there’s no underlying moat once others join. We are building the infrastructure. We think that scales. We think that sustains. We think that’s permanent.

We aren’t going anywhere any time soon.

Norway’s sovereign wealth fund runs the biggest pool of capital in the world.

And its CEO, Nicolai Tangen, might just be the biggest advocate of AI in investing, calling himself a “total maniac” about it.

Stian Kirkeberg is tasked with implementing Tangen’s vision across roughly $2 trillion in assets and about 8,600 companies as NBIM’s Head of AI and ML.

That scale brings a specific set of constraints: broad market coverage, strict ethical rules, and an organization that has to work reliably across thousands of decisions. An individual can quickly boost their own output with vibe-coded solutions, but that does not necessarily translate into a faster organization. When everyone becomes a coder, productivity can rise in pockets while technical debt quietly accumulates.

In this interview, Kirkeberg walks through how NBIM is navigating this transition. He explains their partnership with Anthropic, the move from a bottom-up ambassador model to a more centralized strategy, and how small autonomous teams are replacing traditional Scrum structures. He also gets specific about how they reserve GPU capacity from hyperscalers and how LLMs are being used to screen thousands of companies for ESG compliance.

By reading this conversation, you will understand the constraints that show up when AI-driven development scales, and why the biggest hurdle to ROI is not the model’s performance, but the organization’s ability to absorb what it produces.

This interview has been edited for clarity and length.

Matt: How did you get connected with Anthropic?

Stian: Last autumn, Nicolai invited Dario to his podcast. From there, the ball started rolling. While we were evaluating which tool to buy, Anthropic came out on top. At that time, it was OpenAI and Anthropic, and the others weren’t that great.

Matt: It took a decade for the move to the cloud to happen. This technology is still relatively new.

Stian: We were really fortunate to get this collaboration with Anthropic. We started with basic training and prompting for everyone. Then we set up an AI Ambassador Network which grew from 20 to over 70 people. My AI team had meetings with Anthropic twice a week. Ambassadors were tasked with finding a use case in their area, solving it with the AI team, and then showcasing it to the rest of the organization.

We built a lot of momentum with success stories. This was umbrellaed under “Tech Year 2025.” We created mandatory training for everyone in NBIM—seven different modules covering prompting, critical thinking, and responsible AI. We rolled out Claude, Cursor, and Copilot for everyone who wanted it. We had internal conferences in each office where people celebrated good stories and brought in speakers. We even had a Neo1 robot from a company called 1X.

After that bottom-up approach, we needed to identify the most valuable use cases for NBIM as a whole. Consultants interviewed the chiefs and ran workshops, identifying another 171 projects. Phase 3, which we focus on now, is about people delivering value on everything they’ve learned and the tools they’ve been given. We are pushing the cultural change this year to show the value of those investments.

Matt: In practice, where have LLMs proven most useful?

Read more

When generative AI began gaining traction on Wall Street, many firms responded cautiously, often firewalling off the technology from its employees. At Manulife Investment Management, the reaction was different. After years of investing in cloud, data, and machine learning infrastructure, the firm moved early to establish an AI framework across the organization, building governance, risk controls, and a process for prioritizing use cases.

I recently interviewed Robi Krempus, who leads AI for global wealth and asset management, which has $1.3 trillion in assets under management and administration. Earlier in his career, Krempus was a control systems engineer in the energy sector, working on nuclear power thermodynamics and other high-stakes modeling problems. That background now shapes his role overseeing Manulife’s AI platform. Rather than committing to a single vendor, his team has built a model-agnostic framework that allows the firm to move between providers such as OpenAI, Anthropic, and Google as the technology evolves.

In our chat below, Krempus explains why Manulife moved quickly, how his team co-designs tools with portfolio managers, and why the firm shifted from project-based experimentation to a platform strategy. He also discusses how Manulife evaluates large versus small language models, how it manages tech debt as models change, and where AI is already proving useful, particularly in extracting qualitative signals from standard financial disclosures.

This interview has been edited for length and clarity.

Matt: Manulife moved quickly when generative AI first emerged. At the time, many big firms were banning or firewalling it. What drove that decision?

Robi: We truly saw the opportunity. We had already established a strong data science and machine learning community at Manulife. When you build traditional machine learning models, it is often about forecasting or predicting a variable, which still requires a massive infrastructure. When generative AI came into the mix, we quickly understood that this is much broader and will impact everything—decision-making and how you think about intelligence. Organizationally, we saw the opportunity, and with the CTO, CIO, and Chief AI officers, we were certain this technology was not going to go away. Unlike emerging technologies like blockchain that take time to embed, it was quite apparent that this would be transformational.

Matt: How do you architect this technology? How do you organize it to get started?

Robi: In asset management, there were three ingredients where we believed this would really make a difference. One is strong leadership support. We had huge support from Colin Purdie, the Global Chief Investment Officer for Public Markets, and his leadership team. Secondly, we co-designed solutions with the investment professionals. We have CFAs on my team, but we are not managing the money; the investment professionals are. That co-design allowed us to tackle specific pain points together.

Thirdly, our mindset shifted from being project-based to a platform mindset. We wanted to establish a platform so that whenever we have an additional use case, we can give AI to the end user through that platform. We have seen adoption over 70%, and we hold weekly office hours where investment professionals can stay on top of new features.

Matt: Some firms use various models as an engine and build an application layer on top. Can you walk me through your thinking on building those applications?

Robi: From a Manulife perspective, we have a robust model risk management process in place. Before anything goes into production, it is vetted against hallucinations and quality. In working with investment professionals, quality matters a huge deal. If the LLMs do not produce an output that hits the investment context, it will not work. We architected our AI with feedback loops and tested various systems to increase output quality and reduce hallucinations. It is not a straight-through process to a reasoning model; spending time on the AI architecture to increase quality was really impactful.

Matt: Are you agnostic to the model? Can you swap different models in and out of your infrastructure?

Robi: Yes. That goes back to the ten years of investment we put into infrastructure and cloud. What is amazing now is the availability of all these models. Even when OpenAI released 3.5, we had access to it quite fast. The idea was to create a data framework that allowed us to productionalize models in a responsible way. We have a broad lineup available, whether it is OpenAI, Anthropic, or Google. It is fast and responsible.

Matt: What are the most common use cases right now, and how have they evolved?

Robi: We started with discrete use cases. One that seems obvious is earnings call transcription. We co-designed solutions with investment professionals to build standard prompts for things like red flags, concerns, or bull-and-bear situations. This was managed in a prompt library and helped support investment conviction.

What we did next was aggregate that data. If you take earnings calls and add outside reports or notes, it allows you to search across the board. You can search across your portfolio or a sector for specific topics. That has been very helpful for deeper intelligence across coverage. We also use it for sustainability, which is an efficiency play to quickly get information out of very long documents.

Matt: How are you thinking about small models versus larger ones?

Robi: We have teams that constantly task new models out. We have looked into small language models for operations or distribution areas and have seen a fit there. We are not deploying small language models into asset management right now because we are very pleased with what we can do with large language models. Organizationally, we work strategically on small language models regarding cost and scale, but it hasn’t impacted asset management yet.

Matt: How do you decide between building something internally versus using a third-party source?

Robi: We look at it as “buy, build, or reuse.” Because Manulife is a large organization, we first see if we can reuse something already built. We have a stream that constantly evaluates vendors to see if a solution makes sense. The last thing we want to do is manage internal tech debt. In some cases, we bring in vendors; in others, we build. The platform mindset matters here because it reduces tech debt while allowing us to fine-tune and differentiate ourselves in the marketplace.

Matt: What do you think is currently overhyped or underhyped?

Robi: I am particularly curious about autonomous coders in the bigger tech space. In the past, if you built a machine learning model and a new algorithm came out, you had big expectations for accuracy, but you still had to do so much feature engineering to improve it. Now, innovation and design matter because you have so many options in how you architect data and AI.

Regarding what is overhyped, the real impact is AI’s ability to deeply analyze structured and unstructured data in an automated way. In asset management, with the enormous amount of qualitative and quantitative data, that is where it gets interesting. When these things merge toward AGI, AI will be able to figure out insights and analysis from any sort of data using natural language. You won’t need to know Python to dig that information out.

Matt: How do you see AI impacting alternative data sets and how people use them?

Robi: There is a progression and a cultural change involved. Our mindset was not to wait and establish a perfect, integrated, scalable data infrastructure before building AI. Instead, we put AI into the hands of end users to learn and adjust based on feedback. While we haven’t fully tackled alternative data sets in our AI journey yet, there is an opportunity within standard data sets. For example, in sustainability reports or the footnotes of documents, there can be instrumental nuggets that take a lot of time to find manually. We ask ourselves how we can get critical information out of the data sets we already have readily available.

Matt: What has been surprising to you in building this out?

Robi: The agility really matters. Two years ago, ChatGPT 3.5 came out, and now the world is talking about the AI workforce and humanoids. What was initially a surprise is how fast you have to rethink things. We might build a solution in January, and then a new model like Claude comes out which is an excellent execution engine. The surprise is the constant reimagining and adapting. If you asked me a year ago, I would have been surprised by how far we have come.

Early users span hedge funds, insurers, and sovereign wealth funds. Bridgewater has used Claude to help researchers query internal documents and data. AIG has applied it to underwriting and risk analysis. Norway’s sovereign wealth fund, NBIM, uses it to work through policy and investment material at scale.

I was happy to speak with Jonathan Pelosi, who leads Anthropic’s financial services effort, about how firms are actually using the product. Here is what readers will learn from the conversation.

• How Anthropic is tailoring large language models for regulated financial workflows

• What Claude for Financial Services is designed to do in day-to-day finance tasks

• How skills and Model Context Protocol connect models to firm-specific workflows and data

• Why…

Read more

Inside Man Group’s AlphaGPT

Ziang Fang, senior portfolio manager at Man Numeric, on building AI for systematic investing

Man Group has built an internal AI system that generates trade ideas and subjects them to the same internal review as human research.

The system, AlphaGPT, proposes signals, writes the code, and runs backtests before any human sees the output. Only after that does it enter Man Numeric’s standard research and investment committee process.

The $214 billion hedge fund says the edge is speed and scale. AlphaGPT can produce viable research concepts in minutes rather than days, allowing researchers to evaluate far more investing ideas than would be feasible with a human-only process.

I spoke with Ziang Fang, Senior Portfolio Manager at Man Numeric, about his recent article detailing AlphaGPT’s architecture, how Man controls for lookahead bias and data mining, and the limits of AI in systematic research.

This interview has been edited for clarity and length.

Q: What has been the reaction to the AlphaGPT article?

A: I think it’s been very well received. The article on Man Institute, What AI Can (and Can't Yet) Do for Alpha, is one of the most read pieces we’ve published recently. At the same time, especially among our client base, larger organizations and allocators are thinking hard about how to adopt AI in their daily workflows.

A lot of people have been using AI as a chatbot for different things. But using it systematically, automating it, and applying it end-to-end is different. Many are interested in how we’ve built the process and how we bring it up to standard for delivering products and research outcomes.

Q: How did Man Group’s AI adoption evolve?

A: We did a really good job making AI accessible to everyone. Once ChatGPT became available, Man Group quickly rolled it out broadly. Once people had access, they started experimenting.

Suddenly, it felt new. You could bounce ideas off it or use it to prototype code. Before, you’d have to search online and dig through posts, which was slow. Now you put your idea in and get a prototype, even if it doesn’t work perfectly.

Last year, we started thinking about bringing everything together. If AI was already used across the research process, why not think about an end-to-end, integrated adoption?

Q: Why does the reasoning model matter for quantitative research?

A: Quant researchers want to show something that works, at least on paper. But you need a lot of vetting to understand the research process versus the final backtest. What matters is whether it works in live trading, not whether it looks good on paper.

The reasoning model gives us full transparency. At every step, when an agent makes a decision, it logs why it made that choice. That level of visibility is something you don’t always get from a human-driven process.

Q: What challenges did you encounter building this system?

A: Along the way we ran into a lot of issues—hallucination, lookahead bias, multiple testing, and many other things. One exciting part about AI is that as humans, we take a lot for granted in our daily work. Now we have to step back and reevaluate everything. You ask, why do we do things this way? It created another opportunity for internal debate about what the right approach actually is.

One interesting thing is that because the language model isn't part of the group, it doesn't develop the same blind spots. When you work alongside colleagues, you eventually start thinking alike. The model learns from us but doesn't sit next to us, so it can surface angles we might have missed.

Q: How does AI help with both volume and quality of ideas?

A: There’s been an explosion in data availability. No one can realistically go through thousands of alternative datasets, many of which are unstructured.

Previously, a researcher had to manually figure out how to handle all the alternative datasets, which took a long time and many steps are repetitive. LLM-based agentic workflow provides an opportunity to automate those tasks and help systematic teams to process information at much higher volume.

On quality of idea generation, think about a “researcher” that has effectively read every paper, article, and public code base. You’d want to hire that person immediately. But in reality, that person wouldn’t understand your institutional environment or what good quant research looks like. That’s where AlphaGPT comes in. It combines the raw capability of language models with our institutional context to ensure high quality idea generation.

Q: How do AI-generated ideas compare to human research?

A: We tested this by having the model redo research on datasets we already analyzed. It’s much faster, and it also surfaced ideas we hadn’t considered, which is great.

If you didn’t know whether a signal came from AI or a human, you probably couldn’t tell. The main difference is formatting. The AI output is more consistent.

That said, a lot of nuanced research still requires deep market understanding, judgment, and intuition. We’re not at the point where models can fully replace that.

Q: How does the human oversight process work for AI-generated signals?

A: Our view is that AI isn’t a silver bullet, especially in noisy financial markets. Every idea must be hypothesis-driven, with a clear economic rationale. Once you start backtesting, you can’t change the hypothesis without risking data mining.

AI follows the same rules. It has to state its hypothesis, explain why it makes sense, and stick to it. Flipping a signal after seeing results isn’t allowed for humans or AI.

Humans stay in the loop throughout. We review every step, and developers inspect the code line by line before implementation.

Read more

🎤 A Q&A with Aric Whitewood, CEO of XAI Asset Management, on building an evolving AI hedge fund.

INTERVIEW

Aric Whitewood runs XAI Asset Management, a systematic hedge fund built on self-updating AI that trades with minimal human intervention.

In this interview, we discuss:

• How his fund “evolves” with markets through a closed-loop, Bayesian framework that updates relationships as new data arrives while keeping human tuning rare.

• His background from aerospace and radar systems to leading early ML at Credit Suisse to launching a fully systematic fund.

• Why LLMs should be treated as a tool, not the center of “intelligence,” and where they fit alongside time series models, information theory, and neuro-symbolic methods.

• His concerns that too little focus on compute efficiency can inflate costs, encourage synthetic data shortcuts, and lead to stretch valuations.

Here’s how the strategy has performed since launch. The emerging manager's XAI Systematic Macro strategy returned 38.6% in 2022, 0.2% in 2023, 16.3% in 2024, and 20.3% so far in 2025. The results reflect a steady level of risk aimed at keeping annual volatility near 15% and are reported in U.S. dollars before fees.

This interview has been edited for clarity and length. 

Matt: Tell me about yourself

Aric: I started my career in the aerospace industry, after completing a PhD in radar systems and electronic engineering. I spent a number of years working on drones, back when drones weren’t yet mainstream, studying how swarms fly together. I worked on ship-based sensor systems, very interesting. Also, some jet and helicopter projects.

Then I moved into banking, which was extremely interesting as well. I ran pretty much one of the first machine learning teams at Credit Suisse, and eventually became Head of Data Science across sales, trading, all kinds of different functions.

I moved around—London, New York, then Zurich. And then I left at the beginning of 2017. It was a while ago.

Matt: Tell me about your fund.

Aric: The vision of the firm is to create a kind of multi-strat, but with AI creating all the pods. I know other people have claimed, ‘Oh, we have LLM traders, they do everything for you,’ but I’m not convinced by that. What we have is a real track record—actual trading of real assets and with double-digit returns over multiple years.

We’ve done it for macro assets, to some extent for stocks, and we’re now looking at options and other asset classes. The idea is to have pods, but all powered by a very consistent underlying framework—what I call a causal reasoning platform.

This platform pulls together elements of signal processing from my aerospace days, combined with AI and ML. It handles regime shifts and uncertainty. In fact, it embraces uncertainty. That was one of my early realizations: many quants see uncertainty as a nuisance. They widen their distributions, or they avoid it altogether because it doesn’t fit neatly into an equation.

But in signal processing, uncertainty is everywhere. In radar systems, you’re trying to detect targets with imperfect data, constant noise, and competing signals. Sometimes even your own radar system interferes with what you’re trying to see. In finance, the signal-to-noise ratio is just as bad and worse, it changes over time. That’s the challenge, but also the opportunity.

Our system makes uncertainty a feature, not a bug. It’s fundamentally Bayesian in nature. When you fly drones, you often use Markov decision processes to control them. The environment is uncertain, you never fully know what’s going on, but as you observe more data, you refine your understanding. That’s exactly what we’re doing in financial markets: continuously observing, updating, and adapting as prices come in and regimes shift.

We’re now considering raising VC funding to expand and commercialize this causal reasoning engine.

Matt: How might that work? Applying the structure you have in finance to these other scenarios?

Aric: Yeah, exactly. When you think about drones—and autonomous vehicles more broadly—you can represent their behavior in terms of regimes. For example, a car can be in a regime where it’s approaching a junction and turning, and that turn can have its own subtypes. For a drone, landing is one regime, taking off is another, and there are many others.

The same framework we use in finance, handling uncertain information in a Bayesian way, fits these scenarios naturally.

But the key point is that the underlying representation engine, the way we encode causal relationships and make predictions, can be shared across these domains. It’s a common framework that works whether you’re dealing with financial markets or autonomous systems.

Matt: How would you characterize your mix of AI? It sounds like it’s mostly traditional techniques. Are you using LLMs for some of the causal aspects?

Aric: We use a combination of time series techniques, other ML methods, information theory, compression, all kinds of things mixed together. The LLM work is a bit more recent, but more generally, we’ve done language processing work—data pipelines, creating our own sentiment indicators from high-quality news data. That worked fine. Now we’re experimenting with open-source LLMs to infer information from text.

The issue with applying LLMs to time series is that we don’t have enough data. There’s some research on how small an LLM can be while still producing reasonable output. You can get down to around a million or a few million parameters and still get text that’s not bad. But training a system with millions of parameters on our data would be a disaster — an overfitting mess. Time series are very high-dimensional.

I’d say the AI we rely on is closer to neuro-symbolic AI. It’s not new from a component perspective — but then again, LLMs aren’t really new either. They’re the result of years of iterative progress on architectures.

My point is that using the LLM as the center of reasoning for the whole architecture is risky. It’s not explainable, it’s not deterministic, it hallucinates — all the issues people have been talking about for a while. Whether that will change, I don’t know.

Matt: How do you view traditional AI versus LLMs?

Aric: When I ran a team at Credit Suisse, we looked at all kinds of different techniques. My view has always been that you match the technique to the problem you have. You don’t just take the “best” technique you think there is and try to apply it to everything. Sometimes a rule engine works very well—even though it’s from the 1980s and no one likes them anymore. But they can work extremely well.

At other times, an ontological system is really great because it can represent some of the knowledge stores you have. I’ve always been agnostic about these things.

I think a lot of AI researchers tend to go too far into pure computer science and don’t look at cognitive psychology or the human-driven research that’s out there, which is very interesting. Our approach is to actually draw some elements from that research.

Matt: So you’re reading psychology papers?

Aric: Yes, think about it. It’s not that we as humans are just continually pulling in data, processing it, and immediately spitting out an answer. There are more kinds of representations and reasoning going on. So why not be flexible in that approach?

I’ve seen people recently talking about world models [a machine’s internal “brain” modeling how the world works]. We’ve been doing that for years. We’ve effectively had a world model for financial markets running and doing trading for us for years. That seems like a very obvious thing to do. Why wouldn’t you?

Why would you assume that it’s just going to appear implicitly because you add more and more layers to your neural network? I’ve never understood that approach.

It almost feels like going down the world-model route is quite a lot of work. You have to think: how do I represent the information? Should I put it this way? How do I join it with everything else? It’s not easy at all. The other route — more compute, more data — feels simpler. You’re using the same techniques, just scaling up. But I don’t think that gets us anywhere. We see that with common sense failures, hallucinations, and all the other issues LLMs have.

Matt: How often do you tweak your model?

Aric: We don’t. We are completely systematic. Once we’ve trained the model, we run it — and it just keeps running. We might tune it after a year or two, but tuning doesn’t mean changing it completely. It might just mean adjusting parameters slightly if the model has drifted.

But generally, the system — the framework — takes in new information, updates its own parameters, updates its knowledge representation as new data comes in, and then trades. It’s completely a closed-loop system.

Matt: It’s evolving by itself?

Aric: Exactly. It’s an expanding window of knowledge, but it’s not like the number of relationships is exploding. It does increase, but because we compress information, some of the new data strengthens existing relationships and some creates new ones. The new ones might fade over time, becoming less relevant, or they might get reinforced and become long-term relationships that actually generate alpha. So, we have a system that continually updates its knowledge of financial markets.

Matt: That’s not true of all systematic hedge funds, right?

Aric: No, it’s not true of all of them. Many have a human in the loop somewhere. Many funds use a bit of systematic trading and a bit of human trading. We’ve really gone in and built the fund as an end-to-end, fully systematic system.

It’s quite rare. One large investor said to us, “We haven’t seen many firms like you.”

Matt: Do you think AI shrinks the number of discretionary money managers?

Aric: People have been predicting the death of the discretionary trader for a long time, but it still hasn’t happened. There’s still value there. There are things machines can’t do well.

Matt: Any predictions on how AI will evolve over the next few years?

Aric: We’re going to carry on with neuro-symbolic and our approaches and carry on hopefully making money, and, as I said, maybe moving into some other areas. But my view: there’s obviously been a lot of money spent on people to do superintelligence, with some people saying there are only 200 people in the world who can deliver on this, and so on and so forth. I view that as bubble mentality. There are literally thousands of good researchers in the world who know all about AI and can deliver all kinds of interesting things. There are loads of different techniques you could use. The problem is we, as an industry, parked ourselves in this little cul-de-sac: deep learning, transformer-based architectures, which are certainly powerful, but have their limitations.

Matt: What did you make of Deep Seek’s progress?

Aric: The main story is that they trained their model with far less compute power by being clever about how they trained it and introducing reinforcement learning.

So can you train a similar-quality model with less money? Yes, you definitely can. And there’s more and more research showing that you don’t need as large a model. You can get rid of about half of the weights and the model accuracy is only minimally affected. And you think, how much redundancy is there in this model? You really don’t need such a huge model. You don’t need so much money to train to produce that output.

Matt: The assumption that everyone’s making: oh, this is just what it costs.

Aric: It is an assumption. And then the other dangerous thing is going into this simulated-data area where you’re saying, I don’t have any more internet data to train on. I’ve used all the Reddit posts that exist and all the Wikipedia articles that are out there, and now I’ve run out. So what do I do? I’m going to generate some more data. I think that’s an incredibly dangerous path to take, because I really think the quality of the models will go down, not up. I think you need to use the data you have much more carefully and efficiently, not add more data. If you reach a brick wall and you’ve used all the data on the internet or whatever it is, there must be something wrong.

My aerospace background shapes how I think about compute. Back then, we were getting systems to run on FPGAs—embedded devices with limited memory—so efficiency was everything. We spent enormous effort getting fairly complex algorithms to work under tight compute constraints. Today, it feels like that isn’t a consideration. Instead of optimizing, we just throw massive GPU clusters at the problem and spend hundreds of billions on compute. I worry that this might not be good scientific practice. There should be more focus on efficiency rather than brute force. Does that not set off alarm bells?

We still don’t know if this approach will truly work out. My own gut feeling is that pure LLM approaches won’t lead to “superintelligence” or whatever term you want to use. I’m just not a believer in that outcome.

RECAPS

*In Case You Missed It:

• Spotting Accounting Shenanigans with AI: Transparently.AI’s Hamish Macalister

• Quantifying Intangible Assets with AI: Sparkline’s Kai Wu 

• Investing with Traditional AI & LLMs: QuantStreet’s Harry Mamaysky

• The Open-Source Project That’s Making SEC API Calls Cheap

• Effective Prompts For Investment Research

*Not investment advice

Hey, it’s Matt. You’re reading AI Stack, an interview series exploring how investors are adopting AI. In this edition:

🕵️‍♂️ The CEO of Transparently.ai, Hamish Macalister, PhD, on identifying red flags with AI.

INTERVIEW

I spent six years writing about white-collar crime for Bloomberg News. In that time, I learned that accounting fraud cases were among the longest for the SEC to investigate and the hardest to bring.

I was surprised. I naively thought, “Well, if the company is cooking the books, eventually folks will find out, right?” But that’s not always the case. Bad actors can use events outside their control—like COVID—to bury years of weak numbers.

It’s just hard to police accounting statements. And even if you latch on to what you think is a significant issue, sometimes it doesn’t matter because the company is massive. I once wrote about a company that stuffed six months of revenue into a quarter and the market basically shrugged. Granted, this detail does not inspire confidence.

Ambani’s Mobile Startup Packs 6-Month Sales Into a Quarter

A review of Jio’s unaudited results for the last year shows that the wireless venture and its parent relied on a series of accounting decisions that wound up portraying Jio’s financial performance in the best possible light.

To dig deeper into these challenges, I spoke with Hamish Macalister, co-founder and CEO of Transparently.ai, which uses traditional AI and large language models to assess signs of accounting manipulation.

Transparently.ai rates the accounting health of 80,000+ public companies on an A-to-F scale, flagging early signs of manipulation and potential failure. Founded in 2021, the Singapore-based company counts two of the Big Four auditors as clients and money managers overseeing $4 trillion in assets.

Macalister worked as a macro strategist at Citigroup, led quantitative strategy in Asia at Deutsche Bank, and later served as chief data scientist at Firth Investment Management. He also earned a PhD in finance, where his doctoral research on analyst forecasts laid the groundwork for Transparently.ai’s approach.

In this interview, you’ll learn:

• Why accounting manipulation is more common than most investors think.

• How avoiding high-risk companies based on these scores can generate meaningful alpha.

• Why auditors and analysts miss red flags—and how AI can surface them. 

This interview has been edited for clarity and length. 

How does Transparently.ai help investors evaluate accounting nuances across industries?

This is a perfect problem for machine learning because it’s very complex and multidimensional, but also one for which there’s a great deal of data. That combination makes it well suited to machine learning.

There may be relationships a person would struggle to identify, but a machine can. Another advantage is that the machine isn’t wedded to traditional ways of thinking. For example, it might pick up on signals that an activist short seller would look at, but from a different angle.

One of the red flags might be unusually high margins—possibly a sign the company is faking revenue or hiding costs. That’s a classic example of what an activist short seller might look for.

From a machine learning or AI standpoint, the system might learn something similar: unusually high margins can be a warning sign. Our system does flag that from time to time. But it can also flag unusually low margins if it detects that certain combinations of features—low margins alongside other factors—may indicate a company is doing something unusual.

Machine learning can identify very complicated patterns that may not be intuitively obvious. The one thing I’ll add to that is it cannot just be a black box—unless you’re a quant and all you care about is the black box. In that case, all you want is the numerical output: the risk, the number, the indicator.

But for most of the users we deal with, they want some sort of explanation behind this. So it’s critical to design the system not only to provide an indication that something unusual may be happening in a company’s accounts, but also to explain why and how. It should guide what you need to do next: what questions to ask management, what areas to investigate, and what procedures to implement if you’re an auditor, given the specific features of that company.

What was the a-ha moment for you to start this company?

The a-ha moment came when I was a quant fund manager marketing my fund, talking to private wealth advisors and others. I would casually mention accounting manipulation, since to me it was a very small part of the process.

But the reaction across the table was visceral—people would literally stop me mid-sentence: “Wait, stop there. How do you do that? I didn’t know that was possible.”

I kept hearing it again and again: “I didn’t know it was possible to quantify aspects of accounting manipulation, or to quantify the quality of the accounts.”

I heard it so many times that I started thinking: first, this is amazing, because clearly nobody seems to know about it. And second, while there’s actually quite a significant body of academic research in this space, very few people are aware of it—because very few people read accounting journal articles unless they have serious sleeping problems.

How big is this issue?

It’s a multi-trillion-dollar-a-year problem, and forget about what we say: there’s academic research that shows it. It’s a monster pain point for which, as far as we could tell, nobody else had come up with a solution.

Independent academic research finds that, in the U.S., on average 40%—four zero—of companies manipulate their accounts every year. That’s astonishing. Manipulation can range from something mild and permissible to outright fraud. At the extreme, the same research found that 10% of companies commit securities fraud annually. That’s mind-boggling.

Now, let’s just take that 10%. If you knew in advance which companies were doing this, you wouldn’t touch them with a barge pole. The numbers don’t matter if you can’t trust them. And if you can’t trust the numbers, your investment analysis breaks down.

But what we realized was that there wasn’t much understanding of just how widespread this problem is.

If you’re talking to, for example, the audit assurance team of a Big Four auditor, they know how big a problem it is because they see it firsthand. But for your typical investor or bank asset manager, while they recognize it as a pain point, there isn’t necessarily an appreciation for just how large it really is.

That’s why we started producing research showing, for example, the magnitude of return differentials between high-risk and low-risk companies in our system using true point-in-time data. This isn’t about a backtest.

We generate our risk scores and ratings for companies, then track their performance over the next 1, 3, 6, 9, 12, 24, and 36 months. We looked across all these different periods and compared the performance of high-risk companies versus low-risk companies.

What we found was that the alpha was far larger than we expected. This system keeps surprising us—every time we look at it from a new perspective, the impact is even more dramatic.

Importantly, we didn’t design it to do that. We designed it to identify corporate collapse, or the likelihood of collapse, over a two-to-three-year lead time. But what we discovered is that there’s very significant alpha in our work even for one-month holding periods, which was really surprising.

Instead of comparing the best and worst companies, which mimics a long-short portfolio, the question is: what happens if a typical billion-dollar fund simply avoids the worst companies above a certain risk threshold? What difference does that make to return and risk performance? Once again, the numbers are ludicrous. And because it’s just a matter of not holding something, there’s no issue with trading costs or implementation.

How are the ratings calculated? Are they deterministic, traditional AI with a generative component?

It’s partly deterministic and partly non-deterministic. The predictive AI component definitely has an element of being non-deterministic, though not as much as the generative AI component.

I remember this Journal story. They highlighted academic work showing that 0.4 doesn’t show up in earnings, because if you want it to be 0.5, you round up.

This is a very good example, and it’s well known in the academic literature. If you think about the distribution of earnings surprises, imagine a bell curve. The peak is usually around a small positive surprise. Then you get long tails in both directions. That’s what you’d expect.

But in reality, the distribution looks like that with one exception: you don’t see small negatives. Instead, there’s a dive down and then a jump back up around small negative surprises. That goes to your point: companies know you don’t want to report a small negative surprise. If you’re going to miss, make it a big miss, because either way the market will react badly.

So you get these biases in earnings surprises that create very odd mathematical patterns—yet they make intuitive sense. Companies don’t want to disappoint, and if they have to, they’d rather do it in a big way.

Can you talk about the landscape now? Are there any particular sectors that the model flags as being higher risk?

Read more

AI’s Biggest Cost Lacks Standard Pricing 

Five Minutes With Simeon Bochev, CEO of Compute Exchange

By now, you’ve heard how expensive AI is to train and run. McKinsey estimates the boom could require nearly $7 trillion in data center investment by 2030.

Numbers that big are hard to wrap your head around.

Even harder to grasp: buyers of computing power have almost nothing to benchmark against. There’s no Expedia for compute the way there is for flights. It’s hard to know if what you’re paying is “fair” or “market rate.” It’s similar to calling around plumbers for quotes to fix a busted pipe.

And how do you keep track of how much compute you’re using? Does that match your invoice?

Today, there’s very little standardization.

That gap is what Simeon Bochev is betting on. His startup, Compute Exchange, is building a marketplace to treat compute more like a commodity. I spoke with him about how this market could evolve, why he teamed up with DRW’s Don Wilson, and where he thinks compute belongs on the spectrum between oil, electricity, and currency.

As AI costs balloon and more companies scramble for scarce GPU resources, opaque pricing lets bad actors exploit uninformed buyers. The market is ripe for the kind of transparency common in today’s commodity exchanges.

❝  

“You shouldn’t need a PhD in infrastructure to make sense of this [market]. But right now, you kind of do.”

  Simeon Bochev, CEO Compute Exchange

This interview has been edited for clarity and length.

How did you connect with Don Wilson?

I was invited to a dinner hosted by Don to talk about the commoditization of compute. The CEOs of most of the neo clouds were there.

The reason my CEO wasn't there is because he's in Vermont and I was his leadership person in the office, in the headquarters in San Jose. And so I went and Don and I really hit it off. And we really talked about how there are these structural issues in the market.

I described one of them, which is pricing transparency, right between Amazon and Apple. But the issues go well beyond that, and I don't want to go into a monologue, so I'll skip the issues for now. But we identified a lot of structural issues on both the supply and the demand side.

And so we said maybe a market can solve that. And so I ended up being Don's point of contact within Lambda about building this market. And as we went down that path more and more, I'm like, actually, why don't we do this together? And so, with the blessing of Stephen (Balaban, CEO) at Lambda, we started Compute Exchange in April of 2024.

And now, more than a year and a half later, we launched the market in February, we've had over 80,000 GPUs in supply in about five months. To go from zero to 80,000 GPUs, I challenge anyone else to find someone that's grown that fast. And it's all in GPUs that we don't own. We don't take balance sheet risk on. We open them up to the market.

I see. So, you're the auction? You are the place to source GPUs.

We are the market. Our goal is to be the place where all compute in the world transacts.

So, think about how oil transacts largely in one place, whether you're buying a barrel to speculate as a market maker or you're a country buying a billion for, or maybe not a billion, but a million for strategic oil reserve. They all transact in an exchange. That's what we aim to be.

Don Wilson thinks that compute demand will outstrip oil in 10 years. Of course, he has vest interested with stakes in Compute Exchange, but that’s a bold statement. Do you see compute as oil? I tend to think of it as freight.

Sam Altman calls it the currency of the future. I talk about oil. Some people talk about electricity. I don't think there's an easy, this is the exact parallel, which is why part of the onus is on us to educate.

Here's how I look at it: You've got physical hardware that goes into data centers. That hardware, an H100, has a serial number. It can be the number 1 or a million and excluding any defective units, they’re the same thing. You put that hardware inside of a much larger, complex system of hardware. A server or a rack, networking, storage, liquid cooling, all this stuff coming together. That thing then goes inside of a data center where you pump in a lot of power.

Then, there are differences in the data centers, the power cost, where it's coming from, is it carbon friendly or not, et cetera. And then there's the software layer on top of that. That whole thing we call the AI stack. That thing is what we are trying to build a market out of.

It's not the GPU, right? It's not ‘Go resell an H100 alone,’ because that doesn't help anyone, outside of maybe a few countries that don't have access to that compute. And so the analogy of this thing doesn't really exist anywhere else in the world. One factor of it is power, right? And yeah, you can model power grid demand—and as GPUs consume more power, great. You can use that as a proxy for some energy demand. But that’s not the whole story. Just like oil isn’t the whole story. With oil, you have a limited range of octane grades. Most people don’t care whether it comes from Saudi Arabia or West Texas. Candidly, when they fill up their car, they care whether it’s jet fuel or unleaded.

We ultimately hope to get compute to that level of simplicity, but it's not there yet. So, I would say there's not a simple analogy, but there are elements of these other commodities markets that work.

I used to cover white-collar crime—SEC enforcement, financial shenanigans, all that. So, I’ve seen firsthand how a lack of transparency can play out, whether it’s with level three assets or elsewhere. You’re standardizing this market. And I have a feeling that at some point, there are going to be some stories uncovering shady behavior, where someone’s selling compute in a way that’s not exactly transparent.

Without using names, I’ll give you an example. We know some very smart young people who started a company and were being advised by someone they trust. That person is the co-founder of a company that sells GPUs.

Obviously, that person holds a position of power over them as a trusted advisor. And the price they secured through this person’s company was, at the time, about 30 to 40% above market.

To me, that’s icky. Imagine if I’m your parent, but I overcharge you by 40% for something.

Maybe I’m trying to teach you a life lesson—but that’s not what’s happening here. I’d love for independent folks to step in and find these things because it would make the market more efficient. My goal isn’t to point a finger and say, you’re a bad actor. They’re doing what they have to do to sell their compute.

At the same time, in this case, it shouldn’t fall on the users to figure out why it’s a bad deal. It should be obvious.

Even if I’m a sophisticated buyer, we say you shouldn’t need a PhD in infrastructure to make sense of this. But right now, you kind of do.

Would you consider naming who is setting the lease price?

Read more

Hey, it’s Matt. In this AI Street Markets:

🎙️ An interview with Kai Wu, Sparkline Capital founder & CIO, on how he’s using LLMs to better quantify intangible assets.

INTERVIEW

Quantitative investors have historically relied on accounting data and price metrics.

Kai Wu thinks they're missing the soft factors that drive stock performance today.

As Sparkline Capital founder and CIO, he uses AI to analyze patents, corporate communications, and other unstructured data to identify what he calls "intangible value"—the intellectual property, brand strength, and human capital that he believes traditional financial statements understate.

He started his career at GMO working on Jeremy Grantham's $40 billion asset allocation team, helping manage a $2.5 billion global macro hedge fund. In 2014, he co-founded Kaleidoscope Capital, a quantitative hedge fund in Boston that grew to $350 million in assets before selling his stake in 2018.

He founded Sparkline Capital that same year, spending time exploring where the investment industry was headed and discovering large language models—well before they became mainstream. He launched his first ETF in 2021 and has since built a suite of active ETFs centered on his intangible value framework.

In our conversation, Wu explains how he applies AI on centuries of patent data and culture indicators, why he thinks the line between quantitative and fundamental investing is blurring, and why transfer learning made text-based factor investing viable. He also shares his view on what investors can learn from Renaissance Technologies’ use of unstructured data.

This interview has been edited for clarity and length. 

Tell me about Sparkline Capital

The main business at Sparkline is asset management through ETFs. We’re still trying to create alpha using quantitative techniques but in terms of structure we are trying to skate to where the puck is going. A lot of assets and investor interest are moving into ETFs, specifically active ETFs.

Historically, ETFs were synonymous with index funds. But due to a variety of changes, we’re now seeing more active strategies put into ETF wrappers. That provides efficiency, operational benefits, and tax advantages compared with traditional hedge funds. There’s a lot of interest in that category.

I launched my first fund four years ago, a second one about a year ago, and now I’m building out a suite of products centered on the concept of intangible value. I believe that if value investing, in the Ben Graham and Warren Buffett sense, is going to thrive in the digital economy, then we need to adapt the definition of intrinsic value to include intangible assets.

The techniques we use—LLMs and unstructured data—are what make this possible. If you just look at accounting data, you’re missing out on the most valuable information on intangible assets. There’s simply not enough information. Why wouldn’t you also look at the 80-plus percent of data that’s unstructured? And why wouldn’t you use the latest tools to analyze it?

Nobody I know is really trying to solve this problem.

How did you end up focusing on intangible value?

Historically quants have excelled in some dimensions, right? We have the ability to process larger amounts of data faster and in a more disciplined way. We're less emotional, so we're not gonna just sell all our stocks in ‘08.

The downside of being a quant is that historically only a small percentage of the potential universe of information on companies is accessible. Until more recently, quants have been restricted to accounting-based information, price, volume, PE ratios, asset turnover ratios, all that kind of stuff.

But a lot of information isn't even digital. And even that which is digital has historically been very difficult for quants to ingest because you can't take these textual documents and put them through linear regression.

And that’s where LLMs are a huge breakthrough for us, because now we can start saying, let’s base things on text. I wrote a paper called Text-Based Factor Investing, and you can probably guess what that means. The idea was: can we create factors—like Value, Carry, Momentum—but derived from textual data instead? Using NLP, we can generate culture scores or innovation scores and turn those into factors that can be incorporated alongside traditional ones in an investment process.

I think we’re seeing a convergence. Quants are starting to encroach on the discretionary investor’s world, and are now able to incorporate information that historically wouldn’t have been accessible. At the same time, it’s moving in the other direction too. Discretionary investors are being given tools they can use without needing to be master coders. A lot of what we’ve mentioned is increasingly available off the shelf—though of course, there’s still the challenge of sorting through all the different vendors.

They theoretically enable an analyst with no programming experience to benefit from many of the insights AI can provide. Over time, I think these things are going to meet in the middle, where the distinctions between quant and fundamental will matter less.

What tools do you use?

One of the challenges today is that there's been a proliferation in the number of vendors. If you're a fund manager, you're being pitched a million things from different startups.

We can count the number of foundational model companies on one hand, but on top of that there's a whole layer claiming to offer specialized services to investors. It's just really difficult to diligence.

What do you actually use?

I generally try to go homegrown, although I'm probably unique because I've been working with large language models since about 2019.

What were you doing back then? Not too many people knew about LLMs at that time.

I had a career transition. I sold my last hedge fund and was starting my business. It gave me some time to reset and say: Where are the big industry trends? And that's where I discovered large language models and natural language processing techniques. My goal was to quantify intangible assets from the perspective of a value investor. I used to work for GMO, a quant value investment manager, and the problem I recognized was that a lot of the intangible assets were not accurately measured by accounting statements.

The question became, how can we go about quantifying the value hidden in patents or trademarks, these unstructured data sets? It became clear to me that LLMs and AI provided the key to unlocking these data.

When did you first hear about LLMs?

Obviously the [Attention is All You Need] paper and BERT were the big breakthroughs in 2017-8. But I think that the bigger breakthrough was actually less about the models and more about the data.

Deep neural networks were invented decades ago. It was just that computers weren't fast enough and there wasn't enough data to train them in an effective way.

So the architectural breakthrough of the transformer was better than the alternatives at the time. But I don't think that was the actual game changer. The game changer was transfer learning. At the time, you could develop a specialized model trained on 10-Ks, but the problem was there just weren’t many 10-Ks. You’re talking about an extremely small sample to train a large model on so the results wouldn’t be very accurate.

I actually wrote a paper called Deep Learning in Investing in 2020. The takeaway was that training a deep learning model on domain-specific financial data produced results that were worse than a logistic regression or standard dictionary-based approaches. It was only once you took a pre-trained language model trained on a larger corpus of general-purpose text, like the internet or books, and then fine tuned it for our use case and only then did it become more powerful. So that's the key insight, which is the pre-trained aspect, which I think was a big breakthrough. And people oftentimes credit OpenAI or Google with the architectures. But I think really it's the less glamorous component of just being able to scrape all this information, put it into a format that can then be ingested for training purposes.

And then just the insight that you're better off training on all data, even if it's general and not specific to your industry, as opposed to trying to be like an expert at one thing and train only on your narrow data set, but not cross-train.

It's kind of like sports, right?

You want to be a generally athletic person who does lots of different things and dabbles, that's more important than just spending all your day hitting forehands if you play tennis or something else.

What were the conversations like with your peers about this tech back in 2019-2020?

I think a lot of the attention was on the broader use case of deep neural networks.

We had faster computers, machine learning was becoming more prevalent, and many quants were trying to use those things to do portfolio construction or signal selection. So given a panel of a million alphas, different signals, can you select which are the most robust based on historical data? They were trying to use it as basically a fancier version of a regression or boosting type model. I was writing about this in 2020 that I thought that was a dead end. The problem is that the data sets are too small in finance. There's too much noise in the data set. Maybe if you're doing high frequency, it's different, but at least on my frequency, which is medium- to long-term investing, it didn't make sense. Markets are dynamic, anomalies get arbitraged away. Instead, I argued researchers should focus on a subclass of deep learning, specifically LLMs, which excelled at processing text and other unstructured data.

The best example, I think is Renaissance Technologies, the world's best hedge fund. They became really good when they hired the speech recognition team from IBM, I think that's more than a coincidence. My guess is that they were onto a lot of stuff we're now doing, who knows if the architectures are the same, but my guess is that the idea of taking unstructured data and creating signals from that was a core insight they had decades ago, and that's why they're the best hedge fund.

When you say unstructured data, what data sets are you looking at?

I tend to use publicly available information. My edge is analytical, not necessarily access to proprietary data. Today, there are hundreds of data vendors with proprietary datasets, but the ability to purchase these data sets is not really a durable edge. My thought would be to instead focus on publicly available information such as patents that are large, messy and intractable for the average investor.

That's a huge database. I've tried looking through it. It’s not intuitive.

There's a lot of technical language and it goes back to 1790, the first patent. It’s a super long data set, a lot of breadth, a lot of the assignees are actually private companies or individuals. So yeah, it's a messy data set to look through, and that's the stuff I love. The work is in going through it and trying to make sense of it. And that's where being, someone who has spent time working with a lot of text and unstructured data, I feel I have an edge.

Hey, it’s Matt. This week on AI Street Markets:

🎙️ An interview with QuantStreet co-founder, Harry Mamaysky, on how he’s using both traditional AI and LLMs to guide investment decisions.

INTERVIEW

Harry Mamaysky straddles two worlds: academia and markets.

A professor at Columbia Business School, he teaches and researches in areas spanning quantitative investing, fixed income, and machine learning. At the same time, he’s co-founder of QuantStreet, a wealth management and analytics firm that applies systematic, data-driven models to asset allocation. Before academia, Mamaysky spent more than a decade trading credit on Wall Street, an experience that helped shape his view of risk, modeling, and the limits of traditional investing approaches.

In our conversation, Mamaysky, who has a PhD in finance from MIT, explains how QuantStreet uses AI in practice, where the models fall short, and why he sees them as tools to make investors more efficient rather than replacements for human judgment. He also publishes research on Substack, which you can subscribe to here.

How did you get started with QuantStreet? 

I started QuantStreet like three and a half years ago—actually, probably four years ago—with my brother. And our idea was to have a systematic asset allocation strategy for individual investors as well as potentially institutional investors who allocate to public markets.

I am currently a professor at Columbia Business School, in addition to what I do at QuantStreet. I traded single-name credit for about 13 years—at QuantStreet we don’t do anything with single names.

We’re not in a position to have the information we need to be competitive in trading single names, so we don’t do any of that stuff. We work at the level of ETFs, very liquid. We do asset allocation. The way we do it is we use mean-variance optimization.

We have targeted risk levels. At each risk level, we want to find the highest expected return portfolio. The implementation of that idea is much more complex. We track about 40 different asset classes. For each one, we have a machine learning model that creates a return forecast.

We feed it a lot of data. It selects the elements of the data that it thinks are relevant to that asset class. We look at the trend, and we combine the trend with the model—that forms the basis of our return signal.

For each of our asset classes, we have estimates of historical volatility, correlation, and tail risk. And then we say, okay, we want a 60/40 portfolio, which for us means the same risk level as the 60/40 stock-bond portfolio over the prior year. That’s the risk budget. Within that risk budget, let’s construct the highest expected return portfolio we can, subject to only using up that much risk.

The process has a lot of constraints. You can’t have too much exposure in this sector or that sector. There are limits on tail risks. So not only does it have to be a 60/40 portfolio, but it can’t have excessive tail risk at that risk level. So that’s our process.

It’s very data-driven. In that process, we use some non-traditional data, but data that we get from other sources.

What kind of data?

We throw in the kitchen sink.

It’s realized volatility, implied volatility, level of interest rates, inflation, GDP, growth, valuation metrics, profitability metrics if it’s an equity asset class, price-to-earnings multiples, price-to-book ratios—kind of everything. We have some other series, like we use Economic Policy Uncertainty as a measure of the craziness in government. So that’s one of the inputs. The San Francisco Fed publishes a series called Economic News Sentiment, which we can just get from their website. It’s like an average sentiment of economic news from 16 or 20—I forget the exact number—major regional U.S. newspapers.

So that’s an input. For each asset class, we use a machine learning approach to whittle down the 30 different forecasting variables. Not all of them are relevant to every asset class, so the machine learning part of it throws out 27 of the 30 and says these are irrelevant for this asset class.

It keeps the three it thinks matter, and then it creates a return forecast. We have the same process for every asset class, even though every asset class ends up having its own distinctive model, because the forecasting variables selected for each one are different.

The coefficients with which the forecast returns are combined are different, but the process is the same for every asset class.

So that’s our framework.

It is AI in the traditional sense—it’s machine learning in the traditional sense—it’s statistical AI. A few years ago, this was just called data, and now it’s called AI. Today AI is thought of as LLMs, but anyway, AI isn’t the right solution everywhere. You don’t need a sledgehammer to nail a small nail into a table. There’s the right tool for the right problem.

For this problem, this is the right tool. We use AI in different ways. For example, one of the other supporting pieces of evidence we have for our portfolio construction process is a model that tries to forecast the probability of a large sell-off.

We want to know, for the assets we’re invested in, how likely we think it is there’s going to be a large sell-off. What we’re asking the model to do is generate a probability measure between zero and one of the likelihood that a given asset class experiences a greater than 20% sell-off over the next year, let’s say.

For that, you need to use some kind of machine learning tool. We actually use two different ones. One is called logistic regression, which is identically the same as a one-layer neural network. And then in addition to that neural network, we use a deeper neural network with more layers to create this forecast of a large sell-off.

I would imagine there would be investor demand for that forecast.

Yes, whether we can answer it in a definitive way is another matter. For example, we completely missed the sell-off this year. Nothing in our model told us we’d have a massive trade war, and we had no variable that captured that risk. It was simply outside the scope of our model.

When the sell-off happened, I told clients that if you strip out the geopolitics, the market fundamentals looked fine. Some media narratives compared it to the dot-com bubble, but that was total nonsense. The variables that were flashing red before the dot-com bubble weren’t elevated this time. Everything was benign.

While the model failed to predict the 20% correction, it did help us evaluate it. We could say with confidence this was not a fundamentally driven correction—it was policy noise. Based on that, we made no portfolio changes. That restraint paid off when the market bounced back.

That’s the point of these models. They’re not crystal balls, but they can rule out bad explanations. In this case, they showed the sell-off didn’t resemble the dot-com bubble, which gave us the confidence to sit tight—a decision that proved right in retrospect.

What do you think is the right use case for AI?

Perplexity is an AI engine tied into a lot of financial data. You go to Perplexity’s site or chat, upload your portfolio spreadsheet, and say: here are my positions, now find me all the news flow over the past month. It scrapes the web, pulls the headlines, and you ask, “How does this news impact each of my positions?”

That’s the idea. You can already do it with consumer-facing tools—ChatGPT, Gemini, Perplexity. It’s a trust-but-verify process: I trust the output, but I always check it. If it flags a headline, I confirm it exists. Perplexity alone can sift through thousands of headlines and surface the few that matter. You still have to decide on the impact, but it saves you from hiring an analyst.

We also use this in financial planning. A client might say, “My employer’s offering a new tax-advantaged plan. Should I invest?” I’ll do the analysis, then double-check with an LLM: here’s the scenario, here are my assumptions, how would you analyze it?

I’ve been in finance nearly 25 years, and these models often replicate what I do. Sometimes they make algebra mistakes—because they guess rather than calculate—but when I correct them, they fix it. That’s valuable. They can also generate a two-page client explanation in seconds, something that would take me hours to write. I review it, verify it, and send it along with attribution. Occasionally, the model even catches a nuance I missed.

For example, Gemini 2.5 Pro once flagged a different cost-basis assumption than mine. It didn’t change the overall result, but it showed a detail I had overlooked. That’s the kind of backstop that makes these tools useful.

Six months ago Gemini couldn’t do this. 2.5 Flash still can’t. But 2.5 Pro, with its chain-of-thought reasoning, can. And in another six months, the models will handle even more.

It’s remarkable.

The gap between AI and adoption is widening. It’s moving faster than the humans using it.

I’m relying on it more every day than I was three months ago. Constantly, I’m asking, “How can I use this to be more efficient?” For example, the scripting language under Google Sheets is JavaScript; for Excel it’s Visual Basic. I’m not an expert in either. In the past, if I wanted Excel to do option pricing, I’d need to write a Black-Scholes pricer. I know the math, but I didn’t know the syntax—it was tedious and could take me an hour and a half.

Now I just go to Gemini: “Write me a JavaScript module for Google Sheets that does Black-Scholes pricing.” It spits it out in seconds. I copy, paste, and it works. What used to take hours is now instant.

How do you see AI impacting jobs? 

People will keep doing the same jobs, but with new tools. These models aren’t at a point where they can make decisions on their own—humans have to learn how to use them. It’s like when Excel first came out. Before that, people used calculators, or Lotus 1-2-3. Spreadsheets didn’t eliminate analysts; analysts just had to learn spreadsheets. That’s where we are with AI today. The jobs won’t disappear, but the people doing them will need to adapt.

I don’t anticipate mass layoffs in the financial sector. Our students at Columbia already use AI for everything, and professionals will do the same. The tools will make people more productive.

As a species, we’re heading toward greater complexity. We may have people on the Moon or Mars in my lifetime. Society isn’t what it was a thousand years ago—it’s more complex every year. We need tools like AI to navigate that complexity and expand the frontier of what’s possible.

It’s the same in finance. Companies themselves are more complex. Microsoft today doesn’t look anything like Microsoft 30 years ago. The scale, the supply chains—it’s a different order of complexity. And when we’re mining the Moon for resources, analyzing companies will get even harder. That’s why we need these tools. 

ICYMI

Catch Up On Recent Markets Editions*

• The Open-Source Project That’s Making SEC API Calls Cheap

• Demo: Automating Peer Groups with Primer

• Demo: Analyzing Earnings Call Tone with Markets EQ

• Demo: Separating Facts from 'AI Nonsense’ with Hudson Labs  

• Demo: AI Due Diligence with AnalystAI

• Demo: Tracking Big Tech's AI Spending with DoTadda

• Demo: Tracking DeepSeek on Earnings Calls with Aiera

• Demo: AI Agents in Market Analysis with Scalar Field

• Effective Prompts For Investment Research

*Not investment advice

Five Minutes with Silicon Data’s Carmen Li

When I first heard about trading “compute,” it struck me as odd. Why would we need market infrastructure for GPU chips the same way we do for oil?

Computing capacity, powered by GPU chips that run AI models, is reusable, unlike oil or wheat. But the more I thought about it, the more it started to make sense. 

The closest analogy is the freight market. Like a cargo ship that hauls different loads on different routes, the same compute can train different models.

The freight parallel goes deeper than just reusability. When freight markets matured in the 1990s, they became fully financialized—traders could buy and sell exposure to shipping rates, hedge price swings, and trade Baltic Dry Index futures.

New commodities often become financialized once volatility and market demand make risk hedging necessary.

That moment may be arriving for compute, with some well-known Wall Street traders believing compute demand will rival and then exceed the demand for oil.

Back in May, I highlighted this quote from DRW founder Don Wilson, who has a history of bringing new markets to the mainstream:

"The total dollars spent on compute will, over the next 10 years, exceed total dollars spent on oil.”

  DRW’s Don Wilson to the WSJ.

Wilson is betting on that future.

He invested in Silicon Data, a company that provides GPU pricing data and benchmarking services to hedge funds, banks, and AI firms.

I spoke with Silicon Data’s founder, Carmen Li, who’s just as bullish. After stints at Bloomberg, Citi, and DRW Trading, she launched the company in April 2024 to bring financial infrastructure to compute. The company raised $4.7 million in May, which I covered here.

Full disclosure: Carmen and I worked at Bloomberg in 2022, but never crossed paths. This interview has been edited for clarity and length.

Compute is sort of like freight moving digital tokens through AI Models

You've worked at some major financial firms—DRW, Citi, Bloomberg. What made you take the entrepreneurial leap?

"I wasn't a 20-year-old genius with a great idea in college. I kept thinking about ideas throughout my career, but this was the first idea where I thought, 'I'm the only person who can do this.' I was completely biased, crazily overconfident about the whole situation. When I told everybody, I started my own company, everyone was like, 'I knew it'—Everyone was like, 'that's who you are.'"

What made you commit?

"I strongly believe compute will be the largest human resource going forward—it will surpass any energy product in a few years. We need all the traditional financial infrastructure: indexes, data benchmarking, futures, options, swaps for compute. I felt like I was one of the few people who understood both the trading side and GPUs. People who understand GPUs have limited experience with derivative products—options, futures, indexes. Those with expertise in derivatives aren’t necessarily fully attuned to the GPU space.”

What does Silicon Data do?

"Think of me as the Bloomberg for GPU pricing. You cannot buy compute from Silicon Data, just like you cannot buy actual stocks from Bloomberg. We're different from the spot exchanges—you can actually get physical compute from them. They do spot, we do the data layer. We're working with futures exchanges to launch products based on our indexes."

What problem are you solving in the GPU market?

"Even if you're a sophisticated user—say, a PM or machine learning engineer at a hedge fund—it's not your job to config and double-check GPUs. It's almost like being a great driver—it's not your job to fix the engine.

So let's say you're looking for GPU clusters and I tell you, 'Hey, I have 20 nodes in New Jersey, all H100s with the same configuration and Linux environment. You can run your workflow right away with good latency.' You pay—10 days, maybe a month—and it's expensive. Very expensive. You then discover it's 20 nodes with a different setup than promised, or some Linux environments are inconsistent. So it takes a lot of time to synchronize.

There's no insurance, no guarantee, no standardization. It's mind-blowing. If you and me buy a t-shirt on Amazon, we can return it. But GPUs are freaking expensive and there's no insurance policy, no guarantees, nothing."

How do you solve that?

"One of our benchmarking services helps clients verify everything before they start their workload. Think of it like Carfax for GPUs. A third party verifies everything the provider promised—the chip UIDs, connectivities, latencies, performance within expected distributions. If performance is 20% below spec, you can negotiate a lower price. There's a price for everything—it doesn't mean it's worthless, but you need transparency."

You mentioned building a family of indexes. What does that look like?

"Right now we have H100. We're pushing out A100 in the next two weeks, which has longer price history. Then we'll have a token index. What's fascinating is H100 and A100 have almost zero correlation. They have completely different use cases and client bases—great for financial products. You'll see this chart with Bloomberg and Refinitiv very soon.”

Who's using your data today?

"All my clients are inbound so far. If you look at my client list, it's almost like you took a screenshot of the top hedge funds in the world. The calls are different from my Bloomberg days. Usually it's me, the CTO of the hedge fund, the head of AI, the head of machine learning, some PMs, and people who cover semis—all with different perspectives and questions about the data."

How are hedge funds actually using this pricing data?

"They're using our 4 million pricing points globally as leading indicators—for earnings next quarter, for supply-demand shifts in manufacturing cycles, for benchmarking against hyperscalers. For banks financing GPU clusters, they need to mark risk to market every day. They're not going to take your word that you're renting GPUs for $9 per hour—they'll look at public indexes to verify it's actually $3.20."

Where do you see token pricing heading?

Li explains that right now, token pricing is completely static across different platforms. But she thinks the future should have floating markets where tokens become tradeable currency.

"What I envision is pretty much like what DeepSeek started experimenting with," Li says. "You want a floating market for everyone's benefit. Say a new model comes out and people love to try it out, they pay a premium. Then they realize it is not the greatest. So I bought some tokens but I don't want them anymore."

"Token becomes a currency—like stock credit that can be exchanged."

Are infrastructure providers passing GPU cost savings to customers?

"No. Their costs are going down as GPU prices fall, but they're not reducing token prices. Their margins are supposedly increasing. DeepSeek is experimenting—they charge lower prices during off-peak hours. But most providers aren't doing dynamic pricing yet, even though it makes sense."

Any final thoughts? 

"Every single stack is changing every day—GPU level, inference, models. Everything can be disrupted, even Nvidia. For us at Silicon Data, this constant change is exciting because we create benchmarking and data sets. But if you're in the industry, you really have to be extremely vigilant."

ADDENDUM

Future “Five Minutes with” Interviews

This is the first “Five Minutes with” interview I’ve done this year. I had sunsetted written Q&As once I started my podcast as I thought it would be duplicative. But I have a lot of organic conversations that I think `this would make for a great interview’ (this is what happened with what was an intro chat with Carmen). And, luckily with AI recording, I could quickly transcribe our conversation and share with you. I plan on doing more in the future.

ICYMI: Here’s a list of the interviews I did last year covering many evergreen topics:

• Fitch’s Jayeeta Putatunda on causal AI reducing hallucinations

• Former JPM Exec Tucker Balch on scaling investment analysis with AI

• USC's Matthew Shaffer on using ChatGPT to estimate “core earnings”

• Moody’s Sergio Gago on scaling AI at the enterprise level

• Ravenpack | Bigdata.com’s Aakarsh Ramchandani on AI and NLPs

• PhD candidate Alex Kim on signals with executive tone in earnings calls

• MDOTM’s Peter Zangari, on AI for portfolio management

• Arta’s Chirag Yagnik on AI-powered wealth management

• Finster’s Sid Jayakumar on AI agents for Wall Street

• Sov.ai's Derek Snow on AI for fundamental investors

• Bain’s Richard Lichtenstein on AI adoption in private equity

• Snowflake’s Jonathan Regenstein on AI building novel datasets

• Skadden’s Dan Michael on the SEC’s AI stance

• Stardog’s Matt Lucas on hallucination-free AI

• Celent’s Monica Summerville on AI Adoption in capital markets

• Aveni's Joseph Twigg on building a finance LLM

• Persado’s Assaf Baciu on tailored AI marketing at banks

• Professor Alejandro Lopez- Lira on AI-driven stock predictions

Jayeeta Putatunda, Lead Data Scientist and Director at Fitch Group Inc, has been at the forefront of NLP since 2015. 

Spurred by a chance NLP course during her master’s in quantitative methods, she pivoted from Deloitte consulting to data science and has since built significant expertise in combining traditional statistical approaches with modern AI techniques.

In this Five Minutes Q&A, Jayeeta shares her insights on AI evaluation frameworks, the emerging potential of causal AI, and why bridging research and industry remains a key challenge in financial services.

Key Takeaways:

• Causal AI could help address hallucination problems by grounding outputs in validated relationships

• Financial services need specialized evaluation frameworks beyond generic AI metrics

• Regulators require explainable AI models with clear evidence chains

• Integration of traditional statistical models with new AI techniques is crucial for adoption

This interview has been edited for clarity and length. 

BACKGROUND 

Matt: How’d you get into AI? 

Jayeeta: It’s a funny story. I was with Deloitte doing consulting for a few years, and then I figured that it was not my jam. I wanted to be more technical so I did my master’s in quantitative methods and modeling, which focused a lot on operations research and statistics. My bachelor’s was in econometrics and statistics, so it aligned with my background.

In my last semester, I still had half a credit left, and there was a half-credit course on NLP. I thought, “What is this new thing?” I loved it right away—working with unstructured text data was fascinating. Then I found an internship at Omnicom Media Group. They were just starting to build their data science team and had massive volumes of articles and unstructured text data. The task was to do basic text classification, metadata tagging, naming entities recognition mapping. That gave me practical exposure to NLP. After that, I kept doing more and more projects in the space. I remember trying to train a Word2Vec model on my poor CPU laptop and nearly breaking it. At that time, the tech wasn’t there, and we had to do a lot of painful manual work.

Matt: How did the transition from those resource-heavy early NLP models to the current era of GPT feel?

Jayeeta: Back in 2015, it was hard to get anything done because of hardware limitations. If you tried to train a small Word2Vec model on a baseline GPU, it could still take hours, and new words that weren’t in the vocabulary caused big problems. Today, the hardware has improved dramatically. GPUs are much more powerful and cloud-based solutions are everywhere, so we can spin up a reasonable environment in minutes.

We also have better algorithms, bigger datasets, and more robust frameworks. All these changes make it possible to do more complex NLP tasks. Honestly, it’s a relief—we went from feeling limited by technology to being limited mostly by our creativity. Now you can do advanced tasks like building large language models and hooking them up to external data sources. It’s a completely different landscape.

Matt: When we met at that AI conference (ICAIF) in New York, we talked about bridging research and industry. Do you find there’s still a gap?

Jayeeta: Yes, absolutely. Sometimes I go to industry conferences, and everything is so high-level that there’s no detail at all. Other times, I go to research conferences and it’s so specific and niche I can’t see how to bring it back into a real business setting. I feel there’s a disconnect. We need conferences or forums where we can share more practical, technical details without getting lost in either marketing talk or super-narrow research.

I spoke at an AI conference that did a decent job by having a finance stage, an investor stage, a startup stage, and so on. It helped people zoom in on their areas of interest. But we still need more synergy between the theoretical research folks and people building out actual business solutions.

Matt: How did you go from basic text classification to the more advanced NLP side, especially with large language models?

Jayeeta: It was a gradual shift. Early on, I was just happy building Word2Vec or Doc2Vec models, trying to classify articles or extract location tags. It wasn’t glamorous. Then as technology advanced, I started seeing bigger possibilities with deep learning architectures. Fast forward to GPT and now generative AI. It feels like we jumped from a broken Word2Vec pipeline that took hours to something that can produce coherent paragraphs in seconds.

Of course, there’s more to it than GPT. One big theme for me is retrieval-augmented generation (RAG). You see a lot of hype around chatbots, but there’s deeper potential: knowledge-base optimization, metadata generation, document retrieval pipelines, agentic workflows, and so forth. Causal AI is another big one. If you ground your generative model in causal relationships you’ve already established statistically, you get fewer hallucinations.

TECHNICAL INSIGHTS

Matt: Could you explain causal AI in simpler terms and why it’s important in finance?

Jayeeta: Sure. Think of finance or any domain where you care about cause and effect rather than just correlation. Econometrics has always done that, trying to see if a 5% change in parameter A leads to some shift in parameters B or C. You can run simulations to see if that link holds, or if there’s just a loose correlation.

Now combine that with generative AI. If you feed the generative model validated causal maps or relationships, then when it produces a report, it’s less likely to invent false connections. If you’re summarizing how Apple’s green initiatives led to a reduction in carbon emissions, you can validate that 5% or 10% figure through your causal model. The generative model can then generate text that reflects those relationships, rather than guessing.

Matt: That addresses one of the big concerns: hallucinations. Is that part of why we need more than just raw LLMs?

Jayeeta: Exactly. Especially in regulated industries like finance or healthcare, 99% confidence isn’t good enough. People want to know how you arrived at your conclusion. If a generative model says, “Company X decreased its carbon footprint by 20%,” but in reality it was only 5%, you have a serious problem.

So the question becomes, “How do I prove this output?” You can add a retrieval mechanism that references a trusted knowledge base. But you can also anchor it in a causal model that runs these simulations and says, “Yes, a 5% change here caused 10% change there.” If your final output strays from that validated relationship, the pipeline can flag it or correct it.

INDUSTRY CHALLENGES

Matt: How do you see regulators responding to these new AI tools?

Jayeeta: They’re not going to blindly trust a black box. If you can’t explain why your model gave a certain rating or prediction, you won’t get approval to deploy it. The compliance frameworks require traceability. That’s why layering more transparent models—like XGBoost or older econometric models—under the hood can help. They’re proven and easier to explain. You can incorporate the generative element for reporting, but the generative part has to be grounded in something regulators can see.

In finance, you have to prove each step. If you’re using a large language model to generate a final report, you want to show how the data moved from a baseline model, through a causal map, to the final narrative. That chain of evidence is crucial, or people won’t adopt it.

Matt: Given how fast everything’s changing, do you think it’s hard to start a company in this space?

Jayeeta: Definitely. People pivot so quickly. Six months ago, everyone was into RAG. Now they’re exploring agentic RAG. Next year, there might be a new architecture that changes everything. If you start a product today for summarizing earnings calls, you may find that half the market already does it. Or if you focus on a certain technique, you risk it becoming obsolete because the big tech players release something more efficient.

But there’s also opportunity. If you focus on solving a real problem—like robust evaluation frameworks or specialized causal solutions—you can differentiate yourself. If you just do what everyone else is doing, you’re going to be left behind.

Matt: Let’s talk about evaluation. You mentioned it’s a blind spot. Why do you think that is?

Jayeeta: Right now, there aren’t many holistic frameworks for end-to-end evaluation. People talk about faithfulness or toxicity, but those are very generic. Finance needs specialized checks for factual accuracy, updated exchange rates, or proper reference data from legal documents. Healthcare needs a different set of checks.

There’s a lot of customization, so you can’t just adopt the same evaluation approach used for a casual chatbot. That slows adoption because people say, “I can’t trust your output if there’s no recognized standard.” If we had a well-defined framework that organizations could easily tailor, it would remove a big barrier.

Matt: So, no universal baseline, right?

Jayeeta: Exactly. I think an industry consortium could set up a handful of widely accepted benchmarks. Everyone would still customize them, but at least we’d share a reference point. At the moment, it’s like the Wild West. Each vendor claims, “Our solution is 95% accurate,” but how? On what data? According to which criteria? It’s impossible to compare apples to apples unless you know the underlying metrics.

Matt: Any final thoughts on bridging the gap between research and industry?

Jayeeta: We need more back-and-forth between researchers building new architectures and industry practitioners hitting real bottlenecks, like compliance or data governance. Researchers often tackle niche problems, which is great. But we also need them to address the everyday frustrations of building and deploying solutions in highly regulated domains.

We can accelerate adoption by making the solutions more trustworthy and transparent. Once we have robust evaluation, causal grounding, and easy ways to integrate older predictive models with new generative techniques, you’ll see more companies jump on board. I’m optimistic that by 2025 we’ll have frameworks to handle a lot of the current unknowns, like hallucination or incomplete references.

IN CASE YOU MISSED IT  

Recent Five Minutes with Interviews

• Ex JPM’s Tucker Balch  on Scaling Investment Analysis with AI.

• USC's Matthew Shaffer on using ChatGPT to estimate “core earnings.”

• Moody’s Sergio Gago on scaling AI at the enterprise level.

• Ravenpack | Bigdata.com’s Aakarsh Ramchandani on AI and NLPs.

Subscribe

Ex-JPM AI Exec Tucker Balch on AI + investing

As a kid, Tucker Balch dreamed of becoming an astronaut. Inspired by NASA’s finest, he plotted a course: become a military pilot, earn a PhD, and ultimately reach space.

After eight years flying F-15s and earning a doctorate in computer science, he secured an interview at NASA—but a minor health issue kept him grounded.

That setback became a launchpad. 

Tucker went on to navigate an unusual career as a robotics researcher, professor, and Wall Street AI innovator, authoring over 100 peer-reviewed papers and earning more than 15,000 citations.

Balch holds multiple AI and financial technology patents and co-founded Lucena Research (now Neuravest) a firm specializing in AI-driven investment solutions. 

Most recently, in 2019, he moved from Georgia Tech to join JPMorgan, where he helped his former postdoctoral advisor Manuela Veloso expand the bank's AI team from four to 110 members, cementing its position as a leader in financial AI.

This summer, Balch returned to academia, joining the Business School at Emory University.

I met Tucker at the International Conference on AI in Finance, the peer-reviewed conference he founded four years ago, which is a great event. 

Our conversation explored how AI is transforming investment analysis, from processing vast amounts of data to unlocking insights from alternative sources.

This interview has been edited for length and clarity. 

What made you want to become an astronaut? 

Thirty years ago or so, I was very committed to us getting into space.

I wanted to personally participate in that. After reading many biographies of astronauts, I learned that almost all of them—up until a certain date in NASA's history—had been military pilots. A lot of them also had PhDs, so I decided I would go that route. It's not like I decided to do those things only because I wanted to be an astronaut. 

As a kid, I was always interested in being a fighter pilot, but when I graduated from college, I thought, "That's selfish—to go be a fighter pilot." I felt I should use my computer science knowledge for good in some way.

So my first job out of college was at Lawrence Livermore National Lab on a fusion energy project. It was fascinating, but it ended up getting canceled. That's when I thought, "Maybe I really should go be a fighter pilot after all."

So I left Livermore and went that route. And then once I completed training, I was able to work on my PhD, which I did. And I thought it was pretty well aligned to be an astronaut. And I did get the interview. I think personality-wise and interview-wise, I got the job. But it turned out there was one medical red flag that kept me out. I had antibodies against my thyroid which ended up not having mattered at all over my years. But because astronauts might go on long duration space flights, it was considered a risk.

Turns out that it didn't matter, but I was able to keep my fighter pilot job for eight years. 

How did you move from fighter pilot to finance?

After the Air Force, I was a roboticist at CMU, then a professor at Georgia Tech.  I’ve always been motivated to “do good” for society, and over time, I began to realize that AI in finance can be that sort of “good.” Not just for making them rich, but helping people figure out how to manage their money and so on.

In 2018, my former postdoc advisor [Manuela Veloso] joined JPMorgan to establish an AI research group there.  I touched base with her and she said, “Hey, I'd like you to join me here.”

I spent six years at JPMorgan contributing to AI research, which I enjoyed immensely. About a year ago, I started thinking about returning to academia—and now I'm here at Emory. So that's my journey. 

How’s it jumping between academia and Wall Street?

There's important good AI work happening at places like JPMorgan and other banks, but they can't tell anybody about it. So you don't get the synergy of someone half-baking an idea at Morgan Stanley and someone else doing a different part of it at Goldman - they can't get together and make it a better, bigger thing. Among AI groups at financial institutions, AI Research at JPMorgan is a notable exception: They have a remarkable track record publishing in this area.

Whereas in academia, we're perfectly free to interchange in that way. Although individually in academia, we might not be able to execute and excel like some of the people at the banks, through this exchange of ideas we can. That’s, for instance, how large language models came to be—without academia and open-source contributions, they simply couldn’t exist.

What surprises you about AI's current capabilities?

I think we’re asymptotically approaching a peak of AI capabilities, but we’re nowhere near reaching the peak impact AI can have.  Even if it doesn't become more capable, there are so many applications and uses that remain available that nobody's tried or applied it to and yeah, those are, those are emerging and still underexploited. 

How do you see AI changing investment analysis? 

It's still the case that for most analysis tasks people are better than AI, but people are much slower than AI. So there's the capability to scale this up. And, you can do an 80 percent quality job on a thousand companies, as opposed to a 99 percent job on 10 companies. 

If you can have a fairly, decently informed opinion on thousands of stocks you can turn that into a robust investing strategy.

What's more important - the AI algorithm or the data?

The key question is, what is your data? In my experience, the effectiveness and quality of AI in investing strategies is less about the particular algorithm you're using and more about the quality or uniqueness of your data.

Large language models change that landscape in the following sense. When I talk about data, I guess I generically mean numerical data. In other words, some sort of data that you can very easily feed into some sort of trading algorithm. What large language models do is they enable you to turn language into numbers.

Wherever you're getting your written language about a topic, you need to have something that is predictive and perhaps unique, but LLMs allow you to leverage different sources. For instance, if you can listen to the news in Vietnam, translate it in real time, and identify relevant information for specific stocks, you greatly expand your data sources.

People can do that but it takes time and, you’ve got to pay people to be listening and translating and what AI does is to enable a broader net. It opens up a lot more sources of data, more accessible that weren't accessible before. 

You previously started some alternative data companies. How do you think LLMs affect that market? 

Back then, about 12 years ago, a lot of the hard work was in converting this information into some form that was usable for investing. With the cell phone example, you had to have knowledge of where the retail stores were and merge that with where the cell phones were to create some sort of indicator for home improvement stores. It was manual - you had to find one data source that told you where the stores were, another data source that told you where the cell phones were, and put all that together.

I think those sorts of problems are going to be a lot easier now. You can make an LLM stick those two things together for you in a reasonable, reliable, automatic way and scale those things up more aggressively. That's another example of being creative in how AI can scale up and amplify and accelerate what we're doing.

What’s been surprising to you about AI? 

The little things. The big change in the last year and a half is that everybody can use AI. 

I'm using it for routine administrative things like you were just at that conference with me. I needed to send an email to a bunch of people that were there and I had some documents with their email addresses, but I didn't want to manually edit them out, so I simply asked AI to extract the addresses and boom, done. 

IN CASE YOU MISSED IT Recent Five Minutes with Interviews

• USC's Matthew Shaffer on using ChatGPT to estimate “core earnings.”

• Moody’s Sergio Gago on scaling AI at the enterprise level.

• Ravenpack | Bigdata.com’s Aakarsh Ramchandani on AI and NLPs

• PhD candidate Alex Kim on executive tone in earnings calls

• MDOTM’s Peter Zangari, PhD, on AI For Portfolio Management

• Arta’s Chirag Yagnik on AI-powered wealth management

• Finster’s Sid Jayakumar on AI agents for Wall Street

Thanks for reading!

Drop me a line if you have story ideas, research, or upcoming conferences to share. Matt@ai-street.co