Will the next Warren Buffett be an AI bot?
At the height of the dotcom bubble, the investment bank I worked at employed over 200 traders on its Nasdaq desk alone.
That turned out to be the all-time high for the equities-traders job market (sadly) — and not because the Nasdaq bubble subsequently burst, but because it was the last time the stock market was a strictly human vs. human affair: Trading algorithms were introduced shortly thereafter.
Those algos exposed trading as a mostly mechanical job: Given just a little direction, machines could hit bids and lift offers many times faster than humans could ever hope to.
Twenty years later, that same type of institutional trading desk, doing the same amount of business, requires maybe a dozen humans.
And now, a new generation of AI algorithms has exposed another profession as being mostly mechanical: writing. (I just cannot catch a break.)
Language, it turns out, is not as complicated as we thought: AI can replicate the language function of our brains with statistics and probabilities that are not really all that complicated.
The good news for writers is that, while algorithms replaced nearly all of the traders, they replaced none of the investors.
That's because the best investors, like the best writers, are creative: In both cases, success is more a function of EQ (emotional intelligence) than IQ (logic and reasoning).
The AI machines seem to have got IQ down. Could they get EQ, too?
There's no way to program it: We don't know how it works in brains, so we don't know how to code it in machines.
EQ, which seemed like the easy part, is actually the hard part.
And IQ, which seemed like the hard part, is actually the easy part.
But the fear is that EQ could become an "emergent behavior" of AI — that, with enough computing power, machines will teach themselves how to think.
There's no sign of it so far.
I asked Google's Bard why Nvidia shares are down this year and it listed serveral reasons including market volatility and increased competition — it all sounded very plausible aside from the fact that Nvidia shares are, of course, up this year (by 170%!).
I was leading the witness because LLMs have a ready answer to every question.
Warren Buffett, by contrast, would tell you "I don't know" a thousand times while he waits for one of the fat pitches that he says come along only once every five years or so.
An AI investment manager would swing at every pitch — most of which I expect they would whiff on.
Because AI models have the same limitation as every other algorithmic model: It can only look backwards and looking backwards is not super helpful in finance. (You've heard it a hundred times — past performance is not indicative of future returns.)
That's why, in traditional quantitative trading, the alpha comes not from a machine, but from a human with an idea telling a machine what to do.
LLMs don't have ideas. They can only match patterns in the historic data.
AI will discover some patterns that have hitherto gone unnoticed, but those will quickly get arbed away as every model will find the same ones — same as when the machines started trading equities twenty years ago.
Instead of extracting value from markets, the equity-trading machines mostly added value by tightening spreads, adding liquidity, and making markets more efficient.
I suspect they will mostly do the same in crypto — potentially with one additional benefit.
The non-wisdom of crowds
It's not just investors that allocate capital, CEOs do, too.
Warren Buffett is, of course, a great stock picker. But what's made him the greatest investor of all time is his ability to judiciously allocate capital as CEO of Berkshire Hathaway.
Berkshire employs about 380,000 people, but only about 20 of those work in the Omaha headquarters, which is where the capital is allocated.
And most of the capital is still allocated by Mr. Buffett himself.
That's not only because Buffett is an investing genius, but because capital allocation needs to be centralized. The wisdom of crowds may work at the market level, but not at the company level — there's a reason why shareholders don't vote on every investment made by the companies they own.
This is a problem in crypto, where everything at least aspires to be decentralized.
And maybe that's an opportunity for AI?
I'd argue that a mediocre AI model would likely do a better job of allocating capital than the best DAO.
MakerDAO is attempting to prove me wrong (and I hope they do): Its "endgame" plan will put some much-needed structure to its capital allocation process by founding several "SubDAOs" that will allocate funds received from the parent DAO into individual projects — which makes it sound a lot like Berkshire Hathaway!
But it's missing a key element in Berkshire's success: centralized decision-making.
Holders of the Maker token will vote on allocating capital to the SubDAOs, and the SubDAOs — with their own governance tokens — will (I think?) vote on allocating to individual projects.
That's too much voting for my TradFi mind … if it were up to me (and we can all be thankful that it's not), I'd go all-in on AI and let robots do the allocating.
With autoGPT, capital allocation would be just a matter of good prompt engineering: Extend loans to high-quality borrowers at fair-market rates (or something.)
That, to me, would be an improvement on the current standards of capital allocation at many protocols (which is not a high bar).
Bigger picture, it would be a way to bring centralized capital allocation to decentralized crypto.
Do I think it would be as good as Warren Buffett? No, but that's an impossibly high standard.
Just helping crypto do the easy things would be good enough.
No comments:
Post a Comment