A Conversation with Professor Lars Kotthoff – Symbolic AI, Financial Portfolios, and the Growth of LLMs – Part One

Professor Lars Kotthoff, a researcher specializing in artificial intelligence and machine learning at Scotland’s University of St. Andrews, brings a unique perspective to discussions about AI in finance.

“My perspective is simple: I have absolutely no background in financial services. If you took a quick look at my stock portfolio you would agree that I have no idea what I’m doing. I got involved with the R/Finance conference last year because I’ve been part of a development team for machine learning software that can also be used in the financial industry. I personally have absolutely no experience with this whatsoever – which is why I’m not quitting my day job.”

Kotthoff’s journey into AI began during his undergraduate years, sparked by a moment of revelation.

“You sit in class and you listen to the professor, and they’re going on and on. My impression for a lot of what we were talking about was, honestly, this is all trivial. Why are we even talking about it? In a sense I was right. We were talking about things humans do all the time. No big deal, right? “

The turning point came when theory met practice.

“I had to do a practical exercise where I had to implement some practices in code – again, trivial things – except what immediately became clear was those trivial things were super difficult. Activities that human beings see as trivial, turned out to be really, really hard for an AI system. Today this is changing almost in real time, but back then it was all new territory.”

The Long Road to Modern AI

The technology underlying today’s AI revolution has deeper roots than many people realize.

“Neural networks, the fundamental technology underlying all of what most people now call generative AI, aka large language models (LLMs), actually go back to the 50s and 60s. In fact the first paper about deep neural networks was talking about analog computers. These were not just simple things with an input layer, an output layer, and maybe something in between, but multiple layers and quite sophisticated architectures. Of course, back then it was all conceptual. There was no way to actually do it.”

Progress came in waves, with key breakthroughs in the 2010s.

“Two things came together. On the one hand, we had more data, which mostly came from the internet. On the other, we had GPUs, which had existed since the mid-90s. Back in the day, GPUs were all about gaming. You always felt you needed the latest and greatest from Nvidia. Now, of course, they’re doing very different things. When all this came together, deep neural networks really took off.”

Kotthoff sees this history as a powerful argument for basic research funding.

“Generative AI is a great success story for pie-in-the-sky, ivory tower research that has no immediate application whatsoever, not even something on the horizon. Because somewhere down the line, it might result in something really fantastic and transformational. If back in the 50s, 60s, and 70s, people hadn’t had the opportunity and funding to investigate the foundations, even though they didn’t yet have a way to create anything, none of what we see today would have happened.”

Two Approaches to AI: Statistical vs. Symbolic

Kotthoff’s research focuses on symbolic reasoning systems rather than the statistical models that power large language models. He says the distinction matters enormously.

“LLMs hallucinate. They tell us some things that are just completely and obviously wrong. All the time. No one’s really solved that problem. So far, with statistical systems there’s no way to prevent it.”

Symbolic systems work differently.

“Symbolic reasoning systems are much more restricted in what they can do—they don’t generate completely new text, for example. There are safeguards because they’re actually doing reasoning, not just predicting the next thing through pattern matching.”

He uses Sudoku as an illustrative example.

“You have your 9×9 grid, and you have a bunch of numbers. There’s a lot of reasoning you can do. You look at the first row and see there’s a 5 in there, so there can’t be a 5 in any of the other cells. You combine that with reasoning for all the other rules, columns, and so on. In some cases, you might solve the puzzle directly just by reasoning. A human has built these systems—we’ve written the code that has integrated the rules.”

Symbolic approaches have widespread industrial applications.

“One practical application is logistics problems. You have a certain number of trucks, drivers, and parcels that have to be delivered to particular locations. Or chip companies use these for chip design—how do you place your components such that you can connect all of them while minimizing the total amount of wire being used? The state of the art for these types of problems is still symbolic reasoning.”

The key advantage is robustness in responding to change.

“If you have two different problems—maybe you have an additional truck, or a truck is out of order today—the problems might be sufficiently different that with statistical inference, you can’t figure out what to do, or you get completely nonsensical solutions. With symbolic approaches, you can solve the problem from scratch or start from a previous solution and repair it.”

The Portfolio Connection

Surprisingly, financial portfolio theory has directly influenced AI research.

“There’s actually a connection between AI and financial services that goes the other way. State-of-the-art approaches in symbolic AI use portfolios of solvers. This comes directly from financial services. The first paper on this, back in 1998 or so, was written by economists. The title was something like ‘An Economics Approach to Solving Hard Combinatorial Problems.'”

The parallel is striking.

“Because people have been doing decades of research into solving these hard AI problems, there are many different approaches. Most rely on heuristics. It’s similar to how you build a portfolio—you use expert knowledge to do something in a particular context. The idea is essentially the same as for financial portfolios: don’t put all your eggs in one basket. When one component performs badly, you have the rest of the portfolio to still get some return. Portfolio techniques have been really successful in pushing the state of the art in symbolic AI.”

Stay tuned for Part Two to read about MLR3: Machine Learning for R.

Professor Lars Kotthoff recently moved from the University of Wyoming to Scotland’s University of St. Andrews. His research focuses on artificial intelligence and machine learning, with particular emphasis on making these technologies accessible to non-experts. He is a core developer of the MLR3 machine learning framework for R.

Reference

“An Economics Approach to Hard Computational Problems” by Bernardo A. Huberman, Rajan M. Lukose, and Tad Hogg: https://www.science.org/doi/10.1126/science.275.5296.51