A more superintelligent but more uneven AI? The “narrow superintelligence” thesis

By Nate (Nate's Substack) · October 13, 2026.

The email is the public preview of an article whose full text is reserved for paid subscribers. Its central question: what if AI becomes dramatically more intelligent but also more uneven ("spiky"), with ever-wider gaps between some capabilities and others?

The author notes that we tend to expect the opposite. When a model achieves something extraordinarily hard, we take it as proof of a deeper intelligence that should also help it with other tasks: if it can advance mathematics, it should be a better engineer; if it can reason through a problem that defeated human experts, it should spot a flawed premise in a business plan. In his view, that expectation contains a theory of how intelligence works: that progress in one demanding intellectual activity tells us something about the capacity available for another, because both draw on an increasingly powerful shared capability. He proposes examining that theory much more closely.

As his reason for doing so, he cites OpenAI's release of more than 700 mathematical research manuscripts. According to the text, they come from an internal model that has still not been released and include results backed by machine-verified proofs. The author describes this as a serious research contribution, with individual claims subject to verification and review, and as a particularly vivid example of what he calls "narrow superintelligence": extraordinary intellectual performance in specific kinds of work.

He acknowledges a limitation: the model is not publicly available, so its mathematics cannot be compared with its engineering judgment, and this release alone does not show that its capabilities are becoming more uneven. Even so, he believes it makes the question urgent, because the inference from a mathematical breakthrough to broad competence carries a lot of weight in the public narrative about AI.

His hypothesis is that becoming more superintelligent need not mean becoming proportionally more general. An AI could get better at many things while its most exceptional skills improve much faster. We could keep getting better at producing those exceptional skills without making equal progress on the judgment needed to combine them, apply them or recognize when they are not appropriate. The result would be a future in which the successes are increasingly astonishing and the failures increasingly hard to reconcile with them, and in which saying a model is "smarter" would tell us less and less about what to expect.

The email lists the sections of the full article:

1) Why AI can improve and become more uneven at the same time: the weakest capabilities could improve substantially and still fall further behind the strongest ones.

2) How training feedback concentrates improvements: where useful feedback is cheap and abundant, each additional attempt yields more useful information, so those skills could keep improving faster.

3) What two coding results from the same model show: according to the email, the same model reimplemented a toolkit of about 16,000 lines from its behavior and produced an unusable result on a customization task.

4) The best counterargument: general improvement and convergence are different claims, and the evidence so far supports the former.

5) What evidence would change his mind: a system that detects, before anyone asks it to, when a requirement conflicts with the purpose of the project.

The rest of the article, along with the Nate's Library MCP and its Slack community, is reserved for paid subscribers, so this summary is limited to what the preview sets out.