Who chooses the problem

Recall the asymmetry from the model module. Training a frontier model costs an amount that only a handful of organisations on earth can raise. Running it is cheap, and copying it is free.

That shape has a consequence that is easy to state and easy to underestimate. A very small number of groups decide what gets built.

Not through conspiracy. Through arithmetic. If the entry ticket is hundreds of millions of dollars plus a data centre plus a few hundred of the world's most sought-after engineers, then the set of organisations that can buy a ticket is small, and it is concentrated in a couple of countries.

Those organisations then decide what the model is trained on, what it will and will not do, who may use it, what it costs, when an older version stops working, and which problems are worth pointing it at next.

That last one deserves attention, because it is the quietest.

Effort spent on one problem is not spent on another. The engineers, the chips, the electricity, and the money are finite, and they flow toward what can be sold. A tool that helps well-funded professionals work faster has an obvious buyer. A diagnostic aid for a disease that mostly affects people who cannot pay for it does not, however much more good it might do.

This is not a fact about AI. It is how technology has always been allocated. It is worth naming here only because the promise attached to this particular technology has been so large. If a machine really can accelerate parts of science, then which parts, and for whose benefit, is a decision someone makes. It does not follow from the capability.

And the people making it are, at present, accountable mostly to their investors.