LLMs reward expertise
Points and comments are a snapshot, not live.
Domain expertise amplifies LLM output quality, especially for complex tasks.
The author argues that domain knowledge is the key skill for effective prompting, not generic prompt engineering. He cites Terence Tao's conversation with ChatGPT about a Jacobian Conjecture counterexample, where Tao's concise messages and pushback stem from deep mathematical understanding. The author relates this to software: familiarity with a codebase lets you steer an LLM more precisely. He concludes that human expertise remains a bottleneck, as the hard part is communicating the exact desired solution to the model.
What commenters are saying
Commenters largely agree, noting LLMs amplify both ability and ineptitude. One camp emphasizes that unstructured braindumping works, provided you know your domain. Another points out that math proofs are self-verifying, making expertise less critical there, but for other tasks, verification steps (e.g., tests) are essential. Some see a flattening effect for simple tasks but widening gaps at high expertise levels. A few caution that hype about 10-100x speed may lead to tech debt and safety risks.