The Spearhead Problem

The people who lean most on language models for intellectual work are mostly the people who cannot tell when the model is wrong. That is not a dig at anyone. If you had the expertise you would need the shortcut less. So the tool spreads fastest where it is checked least.

You run into the limits of these models in particular places. In programming it is the problems with real intrinsic complexity, where the answer is a new abstraction rather than a recombination of things the model has seen. In research it is the frontier, where by definition the answer is in no training set. These are the places that need expertise, and they are the places where the tool stops helping.

I use these tools all day and I am not arguing against them. What worries me is the incentive they set up over a generation. Deep competence takes years, and most of those years are not fun. If a tool gets you to an acceptable answer in a minute, the slow way starts to look stupid, both economically and socially. Give an incentive like that twenty years and it changes who people become.

In the short run I think the effect is mostly good. Language models raise the floor. They hand out median expertise, the competent doctor’s intuition, the senior programmer’s pattern matching, to anyone who asks. The average goes up and problems get solved that otherwise would not have been.

But expertise was never about the average. The average moves because a few people push far past it. People who spent years building the judgement to see what others could not, and to ask questions nobody had asked yet. They are the spearhead, and they drag the rest of the distribution behind them.

If we make the path to that depth feel pointless, we could end up a generation from now with a higher floor and a lower ceiling. The tools will plateau at some point. The frontier will still need people who can see past it. I am not sure we will still know how to produce them.