Domain Expertise Unlocks More Value From LLMs, Even as 1 Model Serves Everyone
Updated
Updated · O'Reilly Media · Sep 9
Domain Expertise Unlocks More Value From LLMs, Even as 1 Model Serves Everyone
3 articles · Updated · O'Reilly Media · Sep 9
Summary
Domain knowledge—not elaborate prompting—drives better LLM results, the article argues, because experts can spot weak answers, redirect the model and ask sharper follow-up questions.
Terence Tao’s exchange with ChatGPT on the Jacobian conjecture is cited as evidence: his short, specific prompts and selective pushback worked because he understood the mathematics well enough to steer the discussion.
The same pattern applies in software work, the author says: engineers who know a codebase can press an LLM toward simpler or more familiar solutions, while non-experts can usually get only generic output.
That suggests human expertise still matters even as models improve, since the bottleneck in many tasks is not whether the information is in the model, but whether a user can extract the exact form needed.
Comments on Hacker News challenged how reassuring that conclusion is, but the author notes even OpenAI’s math discoveries still relied on expert mathematicians to verify and filter model-generated ideas.