Cheap intelligence makes the rest scarce

In November OpenAI published a note called AI progress and recommendations. One thing in it struck me more than the forecasts: the company that builds the models says what they can do is far ahead of what people and organisations actually do with them. Sam Altman had already described, in The Gentle Singularity, a world in which intelligence gets so cheap that it stops being the thing we count. And Dario Amodei, in On DeepSeek and Export Controls, made the plainer point: reaching a given level of capability gets steeply cheaper every year.

I take that seriously, because it changes what a small company should spend its time on.

If intelligence is a cheap input

Put the three together and a simple picture follows. A model at a fixed level of ability is a component that gets cheaper and is replaceable. Whatever you do well today, the next model does for less. The capabilities of last year become the floor of this year, which is what I wrote in Rebuilt on open building blocks about software, and it is just as true of the model itself.

So what stays scarce? Not intelligence. In my work I see four things.

The organisation’s sources, with an owner and a version, so that it is clear which text governs. The connections to the places those sources live. The approvals: who may change what, and who has confirmed it. And the checking, which I wrote about in November. Behind all four stands the work of making all this fit one organisation, which nobody does for you.

The check does not come with the model

There is a second reason. As far as I can tell, models improve fastest where a program can check the answer: code that runs or fails, a sum that is right or wrong. That is a feature of the task, not of the model. In care, in pharmacy, in social work there is no program that says whether advice on a pressure ulcer fits this client under this organisation’s protocol. So the system around the model has to supply the check. That is where value is created, and it will not come as a free update.

Where I disagree

OpenAI reads the gap between capability and use as inertia: society has not caught up yet. For work where a wrong answer has consequences I read it differently. Often it is a reasonable refusal to act on an answer nobody can trace. In care, the details are the product. An organisation that waits until it can see where an answer came from is not behind. It is doing its job.

And I do not draw the conclusion that the model no longer matters. The model sets the ceiling. What the system around it decides is whether consequential work ever gets near that ceiling.

What I leave alone

I judge no forecast in these texts as met or missed. I put no figure on what any of this costs. I only note that those who build the models say, in their own words, that the difficult part lies elsewhere, and that this is a reason to spend our weeks on sources, checks and implementation, and not on trying to be cleverer than the next model.