What Terence Tao's ChatGPT session teaches operators

Terence Tao — Fields Medal, widely considered the best living mathematician — published a ChatGPT conversation where he works through ideas related to the Jacobian Conjecture. It's worth reading even if you don't care about polynomial maps.

What he actually did (and didn't do)

Tao didn't hand ChatGPT the conjecture and ask for a proof. He used it the way a good researcher uses a junior collaborator: he proposed a direction, let the model respond, then pushed back, corrected errors, and steered the reasoning. The model was wrong in places. He caught it. The conversation still moved the work forward.

That's a precise description of what useful AI looks like in practice — not an oracle, a thinking partner with fast recall and zero ego.

The failure mode this exposes

Most businesses deploy AI the wrong way: they hand it a task and expect a finished output. When the output is wrong or shallow, they conclude AI isn't ready. Tao's session shows why that framing fails. He brought the expertise. ChatGPT brought speed and surface area. Neither alone would have produced anything interesting.

The same logic applies to business agents. An agent that auto-triages support tickets works because a human defined the triage logic. An agent that summarizes contracts works because someone specified what matters. The operator's judgment is the ingredient AI can't supply.

Why this matters for agent design

When you're deciding where to put an AI agent in a workflow, the Tao session suggests a useful filter: is there a human in the loop who can catch errors and redirect? If yes, the agent's failure rate is recoverable and the speed gain is real. If the workflow demands autonomous correctness with no review, you're betting on a reliability level most current agents can't hit.

This isn't a knock on agents. It's a design constraint that narrows the problem to tractable cases — and there are a lot of tractable cases. Scheduling, first-pass drafts, data extraction, routing, monitoring. Each one has a human downstream who validates before anything consequential happens.

If you're not sure which of your workflows sit in that tractable zone, our free AI Opportunity Audit maps your three highest-impact automation candidates from just your website — takes a few minutes and gives you a concrete starting list.

The part most people skip: iteration speed

Tao's conversation runs long. He goes back and forth, revises assumptions, tries a different angle when one closes off. The value isn't in any single exchange — it's in the iteration rate. He can test a dozen approaches in the time it would take to track down a colleague, explain the context, and get a response.

For business workflows, this is the underrated benefit of agents. Not that they're smarter than your team — they're not. It's that they compress the feedback loop. A draft in seconds instead of hours means your team spends time on judgment, not on waiting.

The actual takeaway

If one of the sharpest minds on earth treats ChatGPT as a structured collaborator rather than an answer machine, that's probably the right mental model for the rest of us too. Build your agents around that: fast, reviewable, human-steered. The wins are real and they compound quickly when the design is honest about what the model can and can't do.

Want this built for your team?

If this is the kind of practical agent work you want running in your business, book a call and we can figure out where it makes sense to start.