The model grades you first

Terence Tao shared a ChatGPT conversation about the newly found counterexample to the Jacobian Conjecture. Sean Goedecke read it and wrote the sentence most engineers have thought at least once:

It is the same ChatGPT. Which makes the interesting question not what Tao knows, but what the model did with it. Goedecke's post — 1,125 points and 474 comments on Hacker News — argues that the most important prompting skill is domain expertise, and he's right. But look at what his four observations of Tao's technique actually describe. Three of them aren't about mathematics.

Tao's messages are short and answer the gist, not point by point. He pushes back obliquely — "this looks more complex than I was hoping for" — instead of contradicting. And in Goedecke's own words, by signalling expertise Tao "shunts the model into 'talking-to-mathematicians' mode, not 'explaining-to-amateurs' mode". Only the fourth observation — that Tao supplies the leaps himself and almost never takes the model's suggestion about where to go next — requires actually being Terence Tao.

So the mechanism isn't only that expertise makes you a better prompter. It's that the model appraises you and meters the answer to the appraisal. Sound like a peer and it drops the scaffolding, the caveats, the recap of the question you just asked. Sound like a student and it teaches. Same weights, same subscription, two different products — and it picks which one you get inside the first exchange.

The same model gives its worst answers to the people least equipped to notice.

The strongest objection is one Goedecke posts against himself, in an edit at the bottom. Commenters called the thesis convenient — that it is, as he puts it, "a view that's reassuring them about how they're still valuable". Grant it. One shared transcript from a Fields medalist is an anecdote, and "expertise still matters" is exactly what an expert wants to hear in 2026.

But the appraisal reading is not the reassuring one. Register is cheap to imitate. Knowledge is not. A developer who learns to write like Tao — terse, declarative, no throat-clearing — gets Tao's version of the model: compressed, confident, stripped of hedges. Those hedges were doing work. They were the one part of the output calibrated to the possibility that the reader couldn't check it. Learn the register without the domain and you have bought a better-formatted wrong answer and thrown away the warning label.

Goedecke closes by saying that for many tasks the human is the bottleneck, not the model. That's true, and it isn't the compliment it reads as. A bottleneck is the part of the system everyone is working to route around.