Can Humans Still Trade in the Post-AI Market?

This is one of the subjects I’ve thought most about when considering the future of ai. My hope is to say things that will age well. This is very difficult to do when it comes to predicting the future of ai (or of markets!), but I’m going to give it a shot.

Right now there is still certainly a place for humans in the markets. That is unequivocal. The real question is whether human judgement will matter once ai cognition exceeds human cognition in every metric imaginable, a future which may happen sooner than some think. Of course, if ai is still constrained to live in a computer rather than operate in the real world, then real world data collecting like visiting plants or talking to people could still be a role for humans. But I’m thinking beyond that - whether there will still be a role for the human who doesn’t leave their room but still can trade profitably. Currently, we have plenty of humans doing that.

First things first, I wouldn’t be counting on our economic system to radically change soon (it could happen but I wouldn’t say that’s the base case). The western world has seen tremendous technological development, yet we still operate with a market economy using government backed currency. There’s a chance that ai will so radically shift things that the economic situation will change, but if there were to be a change that change would happen slowly, at a much slower pace than ai growth/change. I think the US will still be transacting in the dollar for some time (maybe for all of our lifetime). Economics, and how humans interact with each other, will still be very important. Despite its tremendous capability, modern technology is still beholden to economics. One interesting thing to note is that, despite everything that’s possible, many companies build their business model off showing ads to their users. That seems very silly in the grand scheme of things, especially the idea that those with the most data on people aren’t authoritarian governments but random C corps wanting to sell ads. However, it is still very important, because that is where the money comes from to fund many businesses like social media. There’s a similar idea with the dot com bubble. At the end of the day, it wasn’t the paradigm shift some were suggesting, and the tech companies with no profit were still beholden to economics.

Side note: a lot of people compare the present situation to the dot com bubble. First, people have been making this analogy for 10 years (when tech companies broke the $1T level). It doesn’t seem like the “bubble” has burst yet. Second, most of the money is concentrated in top tech companies which are, unlike many dot-com companies, actually making a lot of money, and whose earnings keep increasing. Not to say we can’t or won’t have a sell off, but I don’t think the analogy is so fitting.

I find it unlikely that ai will own money or vote any time soon. As a result, all the money in the market is going to have a human speaking for it. Therefore, the rules of how humans manage their money still apply, such as humans being impatient or poor at risk management. Scarcity, such as housing and critical technology/materials/infrastructure/energy, will still be important. It’s still a human world, and humans follow the laws of economics. That seems unlikely to change any time soon.

Humans right now are better than ai at understanding how other humans think. This is very important to the markets (e.g. understanding how other humans / institutions owned by humans will respond to being down on a position). Because of the risk taking aspect of trading, which is not just some side characteristic but the fundamental point of why some people make money in the market and others don’t, trading is still fundamentally presently a human endeavor. No matter how smart your algorithm/llm/money manager is, if you experience a drawdown, you feel inclined to exit. People are naturally scared of the risky bet that can make a lot of money. LLMs can’t eliminate the risk. Some risks in the markets are a matter of pure chance. There could be a drought. There could be an earthquake. There could be an insubordinate employee bent on destroying the company. A company’s leveraged, courageous bet could suddenly become a foolhardy, overconfident mistake. What changed about mstr’s business model from Nov 2024 to now? Not much frankly. The risk is why it went up so much. The risk is why it went down so much. No matter how smart your llm is, that’s never going to change the necessity of taking risk to make an outsized profit. The simple stuff could get taken, such as arbitrage or discovering hidden low-risk opportunities simply because other people didn’t see it. But the big bets on big, risky moves, seem unlikely to go away any time soon. In fact the accelerated pace of ai may make certain kinds of such risky bets more numerous, or extreme in magnitude.

Even in the end state, there will likely be a lot of human driven money flowing around in the markets (uninformed newbies, people with limited access to advanced models, old investors unwilling to adapt), and understanding those humans (and taking money from them) can still work. If your trading is about understanding humans and how they operate, I think you’ll still have a place in the market for some time. I think humans are still quite good at understanding how other humans think, how they’ll react to being down on their position, which humans have good relationships with other humans. They’re quite good at inferring these things, reading between the lines in ways that maybe ai can’t for some time. Especially since it’s really unclear how to even train an llm on this. The data doesn’t really exist and would take a tremendous effort to construct. It’s unclear if Anthropic or Openai or others even care to focus on this (they can make way more money in other large industries that had been untouched by ai like law or medicine rather than trying to squeeze money out of the markets which is one of the most competitive industries where people have been trying to apply ai to). If the big ai players aren’t getting into it, it’s unclear if the smaller players (citadel, jane street etc.) will bother. Especially because it might take a while and a lot of money to construct this perfect dataset (most companies won’t invest in something if it would take over 30 years for their fruit to pay off). It would be a massive undertaking. I think it could happen, but it doesn’t seem there’s all that much effort to do it or even desire presently. If this endures, that is already one clear important role for humans in the markets.

However, suppose ai gets smart enough that it surpasses humans even in the category of understanding how other humans think/interact. There’s still another bottleneck to getting ai to replace humans.

There’s a fundamental issue with training ai / algorithms on the past. In order to actually train the model correctly, you need to perfectly simulate the real world news/information flow in real time. I’m not sure if this is even doable. Maybe the only real thing that can be done is to have it learn as time goes on. If this is the case, that ai can only learn how to trade starting now, then it’s unclear if llms will surpass humans any time soon, since humans in many cases can learn faster on less information than llms (an llm’s main advantage is that it can process and learn from an enormous amount of data). It is actually extremely difficult to, in retrospect, reconstruct exactly what a time period looked like in the stock market, the exact news flow and sentiment, in such a precise manner that an llm can actually meaningfully learn from it. If this is a bottleneck that actually holds, then humans will have a place in the markets for some time (until llms can learn just as quickly with just as little information, which can happen but I don’t think reflects the current state of llms, even the hidden proprietary ones).

To distill the main bottleneck, training llms doesn’t perfectly reflect the information flow of the time period (that is a nontrivial task), and that is what leads to imprecise training. A human who actually lived through the time learned as things were happening.

Training an llm is actually rather hard for tasks where the correct answer isn’t well defined. If it’s a math problem there is a correct final solution (though the manner in which to get to the solution may not be so clear). More importantly, llm’s are good if it’s just the end result that matters, not the process that gets to it (i.e. as long as the code runs correctly it doesn’t matter as much how it coded the process, though of course speed optimization does matter). However, if the process matters just as much as the end result, if it matters how you get to the result instead of just what the result is, then training for a specific process is actually much more challenging. In many cases, though it might seem like we know how the llm is thinking, we might not know deep down what it is thinking to get to an end result. I think in trading, the process to get to a result matters just as much, perhaps even more, than the final result. If the final result is some trade, how you got to that trade matters quite a bit. Even though the llm can spell out in English its thought process (which, by the way, is really hard to evaluate and score objectively), we don’t know definitively that’s actually the thinking that went on in the llm. Moreover, one would need experts to score these llms. Humans with average intuition will produce llms with average intuition (and average in the markets loses money). You would need the very best humans to devote quite a lot of their time to train these llms (doable, but maybe they’d rather just spend their time making money the way they’ve always been doing, or spending time with their family etc.). It can eventually happen but there is a serious bottleneck that does not seem too simple to resolve. A partner at HRT mentioned to me that the big firms like Anthropic and OpenAI are more concerned with other massive untouched industries. If they’re not focusing on this area, I think there won’t be sufficient development any time soon.

There’s another reason still why human money managers can exist. Individual investors, even hnws, might still want people to blame for their investments. They want to invest in a person who they can attach responsibility to. You can’t attach responsibility to an ai. The responsibility is on you for using the ai incorrectly, or for trusting it blindly. This applies somewhat also to jobs employing traders. A manager can fire a human for using ai badly, but they’d get fired if they let the ai do everything and then it messed up. Blame is the whole reason why consultants exist. They never gave information the management didn’t already know. Perhaps traders in companies will just become consultants… for their own company. This is still a side consideration. The biggest consideration is that training the ai to be just as good as humans in the markets may be a lot harder than some people realize (most of the information about markets on the internet is actually wrong or misleading, and if ai is trained on that information it would actually become worse). In any case, there is still a tremendous amount of uncertainty around this, but one shouldn’t be too hasty to assume things will change so radically.