Hacker News Reader

Comment by PeterStuer | Hacker News Reader

People focus on LLM's and diffusion models because they are so omnipresent now. But an AI for stock picking and prediction that would be realy next level and outcompete the current ones cosistently would siphon of so much wealth it would basically own society if the operators were clever enough not to get so gready short term that the system would litterally collapse overnight.

View on HN

Arguably, this has already happened at least twice in the past. One with the discovery of the Black-Scholes model, and arguably again with the advent of HFT (and the ability to basically front-run trades).

The interesting thing about markets is you generally can only make money when things are mispriced. This limits the total potential gain even for actors with perfect information.

Suppose we did have an AI model that could with near certainty predict both the future cash flows of a company and future interest rates. You could very easily calculate the discounted cash flow and determine the fair share price today or at any point in the future. Rather than collapse, markets likely become more stable and stocks would perform more like bonds.

> The interesting thing about markets is you generally can only make money when things are mispriced.

Why would you think things are not terribly mispriced right now? If only we were a lot smarter we'd know how.

>If only we were a lot smarter we'd know how.

That's not true. Pricing and predicting in the real economy depends on information; for a given amount of information, there are vastly diminishing returns for further intelligence. Because intelligence only allows predicting a bit further in future, as the complexity of predicting the future increases exponentially with lookahead distance; it's O(e^n). This is why hedge funds pay for things like satellite footage of oil tankers to predict changes in supply.

Yes, things might be significantly mispriced, but the amount of mispricing is a finite quantity.

This is clear in the classic pump and dump scheme. During the pumping stage, the unscrupulous actor loses money by injecting some amount of mispricing by purchasing the security and bidding up its value. The hope is to generate momentum and hype that triggers others to amplify that mispricing. Then during the dump scheme, the unscrupulous actor can capitalize by removing the mispricing.

It sounds like you also need an AI to generate that mispricing. Like a form of SEO, the misinformation bots tweet to all the trading bots and try to trick them into mispricing shares, so your trading bots can quickly gain advantage. Corewars using the world economy.

The best way to predict the future is to invent it. --Alan Kay

> Suppose we did have an AI model that could with near certainty predict both the future cash flows of a company and future interest rates. You could very easily calculate the discounted cash flow and determine the fair share price today or at any point in the future.

The thing about trading is that, ultimately, prices are not determined by information, but how the hive mind interprets that information. And in practice, that is not always a 1:1 relationship (see the meme stock hypes). As Keynes* famously said: the markets can stay irrational longer than you can stay solvent, is exactly the reason why even having access to perfect information will not make you necessarily successful.

> As Buffett famously said: the markets can stay irrational longer than you can stay solvent

Buffett did not say this. It is widely attributed to the economist John Maynard Keynes almost a century ago but there is no evidence that Keynes ever said it. I believe the current hypothesis is that it originated with a well-known economist in the 1980s.

it's a Keynes quote, but yeah.

My bad, corrected!

If every investment became risk free because of perfect information about the future, it would likely increase the total market value by quite a lot.

Or decrease, because who would go to the casino if everyone knew the outcome? Only the owners and staff.

"Useless" companies are part of the total market value. If they never get any funding, that's less value overall. Even if that translates to more value for "useful" companies.

Also, if you know with absolute certainty from the beginning that Apple Inc is going to be worth X billion dollars, then you never get to buy stock at less than X billion dollars, because everyone has the perfect information. Value would be constant, and investors would get exactly zero return, because zero risk.

There are other variables, how long it will take and how many people can afford to fund it from the beginning and for that long, of course.

You could calculate the value of things under the discounting changes.

The idea of perfect foresight of the future is kind of insane anyway. Not sure why all of sudden we would go back to believing in a deterministic universe.

The universe can be both deterministic and mostly unpredictable. In fact it most likely is both of those things.

Entropy bounds computation and prediction of chaotic systems requires extreme amounts of computation.

Our current understanding puts some things into the category random, could be wrong, of course.

People still buy airline stocks

> because who would go to the casino if everyone knew the outcome

Everyone knows the outcome of casinos - the house wins in the long run. People still go because they think they have a shot of wining in the short run. People like gambling

There are fundamental limits to prediction. Characterizing these limits is an essential part of algorithmic information theory[0] (AIT) and is an operative bound on AI improvement. One of the strongest arguments against "hard take-off" scenarios is that AIT constrained by physics doesn't really allow it.

[0] https://en.wikipedia.org/wiki/Algorithmic_information_theory

I'm not following that argument - such a system might siphon off the trading wealth, but presumably that's miniscule compared to the equity wealth that is assigned to the owners of the companies and the wealth that they create?

Yeah, I honestly think this would be a good thing. ML already exists as market makers. It’s a big-ish industry maybe, but not massive. It probably should be as automated as possible which would reduce the risk free profit to be made to next to nothing.

Does non-automated trading still hold a significant percentage of volume? Not an expert but that would seriously surprise me.

You can create a lott of profit from a stock tanking. How would that create equity wealth?

I think we’ve already been aware of the limitations of ML models in the financial world.

Historical financial data only predicts so well. If there was a way to make a money printing machine with ML it would’ve happened already. It’s a much easier problem space than language or image generation.

Financial data is not enough to describe the financial world. To predict the movement of meme stock, you'd need Reddit/Twitter/what have you data. It's not a limitation of ML per se, but a limitation on how much data we can feed the ML model.

Blackrock already "owns" society.

People get fixated on mindless algorithms when the real deal is always between humans. Software algorithms, no matter how sophisticated, are just another pawn in the human chess.

In some very remote future there might be silicon creatures that enter that chess game on their own terms. Using that remote possibility to win advantage here and now is a most bizarre strategy. Except it seems to work! It shows we are really just low IQ lemming collectives, suckers for a good story no matter how ungrounded.

> But an AI for stock picking and prediction that would be realy next level and outcompete the current ones cosistently would siphon of so much wealth it would basically own society

Reminds me of Rehoboam from Westworld: https://youtu.be/SSRZfDL4874

It would have a brief period where it would be effective and then you would just have an AI stock picking arms race, and would be back in the same situation again.

That is the way it has been so far. Now if we could lobby for the others not to have access to or be allowed to use this technology ...

Exactly. How would it differ from the current algorithm driven HFT stuff? By being slower to react?

Diffusion models came from finance and have been applied there many many years before it was brought to this wave of AI

A LLM that can get insider info would have an edge. Lmao.

I have wondered if this is the risk that Elon is worried about. But he can’t discuss it explicitly because he is working on such a system. If he communicates its existence, then he reveals a huge attack surface and makes his life (and his teams) much more difficult.

So he is limited to publically saying that AI is dangerous but not revealing the true failure mode.

I find this comically unlikely.

What makes it unlikely?

Because everyone getting richer from removing uncertainty isn't a risk to start with.

What about a single actor getting rich and leaving everyone else’s certainty unchanged?

Only works with fairly limited riches, otherwise what is happening leaks out (kind of playing out that way already).

How would it leak out?

Transactions and positions are not all necessary super secret. People and other machines notice things. You need to take a lot of care to be unnoticed unless tiny and once you move to much it is basically impossible (might not know who you are, but might notice your actions and the impact).

Having info on someone’s trades doesn’t mean you know the basis for their trades. And by the time they have executed the trade any leaked info becomes less valuable.

If that were the case then adversarial HFT algorithms wouldn’t work.

Musk is a buffoon. He doesn't know anything more than he's told the whole world. He knows less.

Reply on news.ycombinator.com