I buy the thesis that “plateau” depends on what you measure. If your benchmark is single-shot QA, you’ll hit saturation faster than if your benchmark is end-to-end work (search+tool use+verification).

A thing that feels under-discussed: systems improvements (retrieval, caching, distillation, better evals, tighter feedback loops) can move the frontier even if raw pretrain scaling slows. That looks like progress to users even if it’s not a new giant model.

I’m more worried about a plateau in reliability than in average capability: the last 5% of error rate is brutally expensive.

The title of the article is "Why I don't think AI is a bubble", not "I don't think AI performance will plateau".