The coming singularity, viewed from a trading desk

Discussions of a technological singularity tend to be conducted at a distance from any domain where automation has actually run its course. Financial markets are not at that distance. They have been substantially automated for decades, and they keep score in a currency that does not accept narrative.

So: what does the one large-scale, adversarial, continuously-scored experiment in competing automated agents actually show?

Automation arrives, and the edge moves

Every wave of automation in markets followed the same arc. A capability is scarce and enormously profitable. It gets copied. It becomes table stakes. The profit moves elsewhere and the capability becomes a cost of doing business.

Electronic execution, statistical arbitrage, low-latency market making, systematic factor investing — each one made someone very rich early and none of them is a source of easy return now. What they left behind is infrastructure: spreads a fraction of what they were, and a market that is faster and more efficient than the one before it.

The pattern is worth naming because it is unusual. Automation did not produce a runaway winner. It produced a higher floor and a competitive equilibrium at a new level of sophistication.

Why an edge that everyone has is not an edge

This is the structural fact that makes markets a useful test case, and it generalises further than people expect.

A trading edge is a claim about a difference between your information or execution and everyone else's. If a tool becomes universally available, the difference closes. Not because the tool stopped working — because everyone is using it and the price already reflects it.

Markets are the extreme case: the payoff depends directly on relative advantage. But the mechanism appears anywhere participants compete, and it is the reason “everyone gets a superintelligent assistant” does not obviously imply “everyone captures superintelligent returns”.

What actually got automated, and what did not

After decades, the split in trading is instructive. Automated almost completely: execution, market making, arbitrage, risk monitoring, settlement, reporting. Any task with clear inputs, a measurable objective and fast feedback.

Still stubbornly human: deciding what to build, judging whether a backtest is honest, noticing that a market's structure has changed, and taking responsibility when something goes wrong. Not because machines cannot in principle — because these tasks have long feedback loops, ambiguous objectives, and consequences that someone has to own.

We see this in our own work constantly. The bot handles 684 symbols continuously and never gets tired. A human still has to decide that the pump-fade formulas, with an honest out-of-sample AUC of 0.50–0.52, should be published as weak rather than dressed up — and that decision is not a computation.

The part that is genuinely different this time

It would be dishonest to argue that nothing has changed. Two things have.

First, the cost of building has collapsed. A system that took a team a year now takes a small team a season. That is real, and we benefit from it directly.

Second, the bottleneck has moved from writing code to knowing what is worth writing and whether it works. When implementation is cheap, the scarce skill becomes validation — and validation is exactly where the incentives push hardest toward self-deception. Cheaper backtests do not mean better backtests; they mean more of them, faster, with the same look-ahead leaks that flatter results.

The claim we would actually defend

Not that a singularity is near, and not that it is impossible. Both positions require confidence about long-horizon dynamics that nobody has demonstrated they possess, and the history of this particular domain is unkind to confident forecasts.

What the evidence from markets supports is narrower: in competitive domains with fast feedback, automation raises the floor faster than it raises anyone's relative advantage. The participants get better tools and roughly the same difficulty, because they are competing against each other, not against the problem.

That is a much less dramatic claim than either the accelerationist or the dismissive one, and it has the advantage of being consistent with forty years of results in the one field where automated agents already compete for money.

How this changes what we build

Practically, it argues for boring things. Verifiable claims over impressive ones. Honest validation over better-looking numbers. Systems whose operators can see what they are doing. Open code, where possible, because a claim anyone can check outlasts a claim anyone can make.

If automation really is accelerating, the scarce commodity is not capability. It is the ability to tell whether a capability is real. That is a craft, it is unglamorous, and it is most of what we actually do.