LLMs reward expertise

There's a comforting myth floating around that LLMs have flattened the playing field—that anyone can now get PhD-level math help or clean code just by asking nicely. A recent essay pokes a hole in that idea using an unlikely case study: Terence Tao's own chat transcripts with ChatGPT while working through a freshly discovered counterexample to the Jacobian Conjecture.

What's striking about Tao's conversation isn't magic phrasing or clever jailbreak tricks. It's how terse and confident he is. He fires off short messages, doesn't argue point-by-point with the model, and when something smells off he doesn't demand a fix—he just says the response looks more complicated than expected. The model, in turn, responds with something closer to a peer discussion than a tutorial, skipping the hand-holding it defaults to with less specialized users. Tao also barely follows the model's suggested next steps; he mostly redirects it based on his own instincts.

The author, a programmer rather than a mathematician, notices the same pattern in their own work. Deep familiarity with a codebase lets someone push back on an LLM's suggestions with real specifics—pointing out redundant logic, referencing existing patterns, or reframing a problem in terms the system already understands. Without that grounding, you're mostly just hoping the model's first answer is good enough. This lines up with an older argument the writer has made before: that solving real systems problems is less about abstract principles and more about knowing the concrete details of what you're actually building.

There's a rebuttal worth noting too. Some Hacker News commenters pushed back, arguing that OpenAI's own math-discovery prompts weren't especially expert. The author's response is fair: OpenAI still needed a team of trained mathematicians to verify and filter what the model produced. The expertise didn't disappear, it just moved to the review stage. Somebody still has to know enough to tell good output from convincing-sounding nonsense.

The bigger implication here isn't that LLMs are useless for beginners—they're clearly not, since even a novice can get something out of them. It's that the gap between a novice's output and an expert's output isn't shrinking just because the model is shared. If anything, the ceiling is rising faster than the floor, and the bottleneck has shifted from access to knowledge itself.