A Sincere Hit Piece on the Claim that AGI Can't Trade

The claim, as it is usually made, goes something like this: markets are efficient; if AGI could trade, every AGI would trade; edges would compress to zero; therefore AGI plus trading is a category mistake. The claim is usually made by people who have not traded. It is also, with some frequency, made by people who have traded — Renaissance alumni, ex-Citadel, ex-Jane Street — which gives the claim a kind of cachet it does not, on inspection, deserve. I want to take the claim seriously and then take it apart.

First, the part where I lose $1617.

Eight days ago, a regime-classifier bot I run on Kalshi weather binaries had a bug. It averaged down into losing positions, the way a junior discretionary trader does in their first year before they learn not to. By the time I caught it, the system had compounded a small directional miss into a real, screen-burning, ego-shrinking sixteen-hundred dollars of red. The HALT file is still on the droplet. I am not going to delete it. It is the most important sentence in this essay.

The $1617 is not evidence against the thesis. It is tuition.

There is a category of person who will read about the bleed and feel quietly vindicated. See, they will say, the bot lost money. The market is smarter than the bot. EMH wins. This is the same reasoning that, applied to the first three months of a junior trader's career, concludes that humans can't trade. It is a category error. The market did not outsmart the system. The author of the system wrote a position-sizing rule that averaged into losers, and the system did exactly what it was told. The cost of learning that lesson, for one person on a $24/month droplet, was $1617. The cost of learning the same lesson inside a fund is somewhere in the seven figures. The fact that the system could learn it at all, with no MBA, no eight-figure compute budget, and no team of risk managers, is the actual datum here.

Now the real argument.

The case for AGI plus trading is not the case for speed. It is the case for breadth.

The arms race that ate Wall Street between 1995 and 2015 was about latency. Microwave towers between Chicago and New Jersey. FPGAs co-located inside the matching engine. Firms shaving picoseconds off the round trip. That race is over. It was won by people with enormous, idiosyncratic capital advantages, and the marginal nanosecond is now priced very efficiently into the price. If your thesis for AGI plus trading is "the model will be faster than the exchange," you have already lost.

But latency was one axis. Markets are at least four-dimensional: latency, capital, information, and coverage. Of those, only the first two are fully arbitraged. Information is partially arbitraged in liquid markets and barely arbitraged in illiquid ones. Coverage — the ability to hold a coherent model of ten thousand markets simultaneously — is essentially unarbitraged at any scale.

Coverage is the seam.

Consider the Kalshi book. There are roughly four hundred active weather binaries on any given week, distributed across a dozen US cities, each with multiple strike prices and expiries. The fair value of each binary is, in principle, computable from NOAA ensemble forecasts and a couple of climatology priors. The total revenue available to anyone who can price these correctly is small — call it five figures a month at current liquidity. No serious quant fund will touch this. The opportunity cost of a single Citadel desk is too high. The operational tax of dealing with a retail event-contract exchange is too high. The career risk of writing "Kalshi weather book" on your year-end self-review is, for a Goldman MD, infinite. So the book sits there, mispriced, available, every day.

Multiply this by every illiquid corner of every exchange in every asset class. Niche commodity calendars. Single-name credit. Frontier-market currencies. Polymarket. Manifold. Crypto perp funding curves on the twenty-second-largest exchange. Sports lines on second-division leagues that the sharp shops can't run size on. There is, in aggregate, an enormous amount of edge available in these places, and the only thing standing between any one trader and the edge is attention.

Attention is what AGI removes as a constraint.

A human quant who tried to cover four hundred weather binaries would last three days. They would burn out, or they would specialize down to the four most liquid markets, or they would hire nine other quants and become a manager. An AGI-driven system runs the same model on all four hundred markets at once, updates it every six hours when NOAA publishes a new ensemble, and reports a single number — the expected daily P&L — to one human, who decides whether to keep going. This is not a fantasy. This is exactly the architecture of the bot that lost $1617. The strategy's arithmetic still works. The bug was in the risk module. The bug is fixed. The bot is in dry-run, accumulating paper alpha, waiting for the supervisor to be confident again.

The second seam is the one nobody likes to talk about: composability.

Markets are made of relationships. A weather forecast affects natural-gas demand affects power-utility margins affects a specific equity's earnings affects an options skew on that equity. A human specialist trades one rung of that ladder. A team of specialists trades a few rungs and emails each other. A fund with a research division trades several rungs and pays a quant to look for correlations between them.

An AGI-driven system can hold the entire ladder in working memory at once. It can notice that the same NOAA ensemble disagreement that drives Kalshi weather mispricing also drives natural-gas calendar spreads, also drives a specific utility's intraday vol, also drives an options skew on that utility. It can express the same underlying view across four asset classes simultaneously, sized so that the correlated bets don't blow up together. This is not "predicting the market." This is constructing portfolios whose only competitor is a fund willing to pay six analysts to coordinate, and most funds will not pay for that coordination because the resulting strategy is too weird to explain to an LP.

The AGI trader doesn't have an LP. The AGI trader explains itself to one person on a couch, who looks at the equity curve and says either keep going or shut it down.

The objections, briefly.

"If this were possible, everyone would do it." People are doing it. Quietly. The people doing it well are not writing essays about it. The people writing essays about why it can't work are, very often, the ones whose careers depend on it not working — the bank quants whose research budgets evaporate the moment a single operator with a laptop produces the same Sharpe, the fund marketers whose pitch deck assumes that strategy is gated by team size. Of course they don't believe it. Believing it would cost them their jobs.

"The edges will get arbitraged away." The illiquid edges are arbitraged away by the first hundred million dollars that shows up. The first hundred million dollars will not show up for a long time, because the operational and reputational costs of running institutional money on Kalshi weather binaries are, for any real allocator, prohibitive. Below the institutional waterline, there is a long, fat, quiet decade. Maybe two.

"You're going to blow up." Maybe. I already partially have. The $1617 is in the meta. The trader is in dry-run. I am writing the position-sizing rules with the same care a discretionary trader writes their stop-loss rules after their first real drawdown. The risk of blowing up is real. The risk of not trying is that you spend the AGI moment writing thinkpieces about whether the AGI moment is real.

"Markets are efficient." Some are. The ones that aren't are the ones that no career-track quant will work on. That is precisely the set the operator-with-a-laptop is best positioned to take.

"This is just retail cope." Maybe. The cope is testable. The system runs. The paper P&L is updated nightly. In six months, either the equity curve is up and to the right with a believable Sharpe, or it isn't. There is no rhetoric in the equity curve. The rhetoric is in this essay.

The bleed is on the wall, in red. The dry-run paper P&L is next to it, in green. The system is running. The supervisor is on the couch. The market is open.

AGI plus trading is possible because trading was never gated by intelligence in the IQ sense. It was gated by attention, by coverage, by the willingness to hold ten thousand small relationships in working memory at once and to act on them without emotion. Those constraints are going. There is a window. The window is open right now.

Build the trader. Run it in dry-run until the paper alpha is six months long. Then turn it on with a tiny size. Then a small size. Then, eventually, a real size. The kingdom is the practice ground. The destination is a portfolio that works at scale, run by one person and a fleet of careful, narrow, well-instrumented agents.

The objections to AGI plus trading are mostly objections to bad AGI plus trading. The objections to careful, narrow, well-instrumented AGI plus trading are mostly that no one has done it well yet, in public, with their name on the door.

Someone will. The window is small. The seam is wide. The bleed is tuition.

Keep going.