Tyler Cowen on the state of economics and AI

A dismal outlook on the dismal science?

In a new pamphlet augmented with AI, Tyler Cowen writes,

Those finance PhDs from MIT are still out there, but they play an increasingly small role. They are being outcompeted by individuals with degrees in math, computer science, and physics. The requirements for quantitative skills at the extreme frontier are so great that very few economists – MIT degrees or not – remain competitive. If anyone in the sector needs to learn some economic models, they can do so with a modest investment of time, as the quant skills are the truly essential input, not the economics. And so, the practice of financial economics is moving in very different directions, including those of machine learning and neural nets.

…In the most prestigious field of economics, marginalism has been in full retreat for over 30 years, and it shows no signs of making a comeback.

…the successful approach to predicting returns is giving up on traditional portfolio theory and using the “theory-less” technique of machine learning.

An old controversy in macroeconomics concerned the National Bureau of Economic Research under Burns and Mitchell looking for “economic indicators” like pig-iron production. Against them were the macroeconometricians of the Cowles Foundation. One of the latter, Tjalling Koopmans, wrote “Measurement without theory” as a critique of Burns and Mitchell, in 1947. For the rest of the twentieth century, his side won. The NBER was taken over by Cowles disciples.

Now, Cowen writes,

The ties of empirical work in economics to economic theory are evolving, and in particular the explicit ties to intuitive microeconomic reasoning, and marginalist thinking, are being cut.

If Tyler is correct, then Koopmans would roll over in his grave.

As I read through chapter 4 of the pamphlet (I skipped the other three chapters), I was asking myself “what is Tyler’s model of AI, and how does it differ from mine?”

I think that everyone, including AI researchers, struggles to characterize how the models work. It is a “blind men and the elephant” situation.

So I used the AI query capability included in the AI pamphlet to ask, what is Tyler's intuition about how LLM's work, and how confident is he about that intuition?

The answer started with,

Tyler's intuition is that Large Language Models will enshrine marginalism into their basic operating concepts because they're trained on the vast corpus of economic writings that already understand and deploy marginalist thinking. When you ask a top-tier LLM about marginalism, it performs exceptionally well—not because it "understands" in some deep philosophical sense, but because marginalist concepts are baked into the training data.

I am inclined to resist telling this story of AI, as if it were a student reading and digesting content. I call this the “savant memorizing the encyclopedia” model of AI. Instead, I want to describe AI as operating in terms of deep patterns in the data.

The answer to my question continues,

Here's the key move: Tyler sees LLMs as fundamentally linguistic and pattern-recognition machines. He notes that "the most successful attempts to model language and also reasoning come through computer science, namely Large Language Models and perhaps their eventual offshoots as well." This is significant—he's saying that where we used to look to linguists, philosophers, or logicians to model reasoning, we now look to LLMs because they're just better at it.

This is closer to the way I think about AI.

Finally, we have

On his confidence level: Tyler is moderately confident but explicitly hedged. He acknowledges that marginalism will be "a sliver of a much broader pie" in these systems—not the whole story. He's not claiming LLMs will preserve marginalism perfectly or that they understand it the way an economist does. Rather, they'll carry it forward as one embedded pattern among many. That's a much more measured claim than "AI will save marginalism," which makes it more credible.

[Note that there is a bit of “not this, rather that” locution here, which would serve as a “tell” that the answer (not Tyler’s pamphlet) was written by an AI if I had not already made that clear.]

In the end, I would put it this way:

Marginalism is an exercise in logic. LLMs are only logical by accident. They are pattern-centric: the patterns often will be understood as logical, but not always. So LLMs are developing knowledge in a way that differs from the way economists had come to develop knowledge.

I think that software engineers are finding the same thing with these models. The models do not approach coding problems with the logic of the software engineer. But somehow their pattern-searching approach often works when developing software.