The School of Fish

The school is taxing the predator's brain

A school of fish does not hide from the predator. It does the opposite: it makes itself maximally visible, and in doing so makes any individual fish impossible to hold on to.

The predator's limitation here is not its eyes. It is the compute behind them — the number of targets it can segment, track and extrapolate at once. Every additional fish is nearly free for the school and expensive for the hunter. The defense is not camouflage. It is raising the cost of the computation the attacker has to perform.

This has worked for millions of years of brain evolution. That is a long time for selection pressure to find the counter-measure, and the schools are still there. I speculate that what we are looking at is not a biological accident but some fundamental law of computation — an asymmetry between producing confusion and resolving it that cannot easily be hacked.

The cheapest way to defend something is not to hide the signal. It is to make finding the signal cost more than the signal is worth.

Non-obvious, but computable

I am, as always, talking about real-sector markets — the ones with logistics and complex world physics underneath them. CPG and FMCG retail, manufacturing, and other logistically complex trades.

Those markets have a specific property. The possible actions and reactions are very non-obvious: they are constrained by lead times, freight, capacity, shelf life, minimum order quantities, seasonality and the decisions of counterparties who are themselves constrained by all of the same things. But they are computable. The constraints are real, they are discoverable, and they resolve.

That in-between regime is exactly where this scenario has teeth.

If reactions were obvious, added noise would not hide anything — everyone reads through it at a glance.
If reactions were not computable at all, no amount of compute would help, and nobody would be paying for it.
Between those two, the participant who can afford more inference per unit of evidence sees further. Which means the budget, not the insight, becomes the boundary.

How the noise arrives

The development path I would expect looks like this.

The market gets exploited to a high degree. The obviously profitable deals disappear — competed away, absorbed by incumbents, priced in, or removed by whatever mechanism that particular trade uses. What is left is thin and contested.
Traders start generating noise deliberately. Not to deceive a human reading a spreadsheet, but to degrade the systems learning from the market's signals. Orders, quotes, listings, movements and postures that exist to be observed rather than to be executed on.
The learning systems hit one of two walls. Either they burn compute filtering a haystack that grows faster than their budget, or — if they filter conservatively — they are left with so little surviving evidence that they cannot extract the signal at all. Even a system that is perfectly compensated for the bias runs into this.
Either wall is the same wall. Filtering complexity and data volume are two faces of one computational efficiency problem.

Wall one

You can see everything, and cannot afford to think about it

Wall two

You filtered hard enough to be safe, and have too little left to learn from

Compute per unit of evidence

It is tempting to treat these as separate engineering problems — one a throughput problem, the other a data-scarcity problem. They are not separate. The quantity you run out of in both cases is the same: computation per unit of usable evidence.

Push the noise floor up and you can pay for the same conclusion in either currency. Spend more compute to separate the real deals from the manufactured ones, or spend more observation time waiting for enough real deals to accumulate. The defender does not care which one you choose, because both are costs the defender does not pay.

That is the uncomfortable part of the asymmetry. Generating plausible noise in a logistically complex market is not free — it takes real inventory postures, real quotes, real capacity commitments — but it is structurally cheaper than resolving it. The producer needs the noise to be plausible. The reader needs to know which of it was true.

A model that needs ten times the compute to reach the same judgement is not merely slower. It is the model that gets priced out of the market first, and it gets priced out precisely when the market has become interesting.

So we need faster models on less compute

If this is where real-sector markets are heading, then the thing worth optimizing is not only decision quality at unlimited cost. It is decision quality per unit of compute, and the ability to re-decide often enough to stay current with a market that is actively trying to be expensive to read.

This is one of the reasons we care about inference cost in P34 as much as about accuracy, and why a pre-trained economic decision model matters to us more than a bespoke fit per operator: the expensive general structure of a market is learned once, and what is left per deal is cheap.

I want to be careful about how far to take this. The scenario above is speculation. We have not measured an adversarial noise campaign in a live real-sector market, and we do not claim P34 clears this bar today — the regime library is limited and synthetic evidence does not prove live performance. What I do think is well-founded is the direction of the pressure.

If noise becomes the defense, then efficiency stops being an engineering nicety and becomes the entry ticket.

The school of fish did not need to be smarter than the predator. It only needed to be more expensive than the predator could afford.