LLMs Reward Expertise
This is a clean framing for why stronger LLMs do not make expertise irrelevant. They often make domain knowledge more valuable, because the expert can steer, compress, reject, and evaluate the model's work.
Links: Original source · Shared link · Related link
Logged at IST: 2026-08-04 10:26 IST
What it is: Mario Zechner linking Sean Goedecke's essay on why LLMs reward domain expertise rather than making prompting skill-less.
Gist: Goedecke's core claim is that the most important skill in prompting is expertise in the domain you are prompting for. LLMs let everyone produce passable work outside their specialty, but an expert can get much more from the same model because they know what to ask, what to ignore, where to push back, and what a good solution should feel like.
His example is Terence Tao using ChatGPT on the Jacobian Conjecture. The visible prompting moves are simple: short messages, not too much point-by-point argument, gentle pushback, and lots of self-directed leaps. But Goedecke's point is that copying those moves is not enough. Tao can do it because he understands the mathematics well enough to pull the useful thread out of the model's answer and notice when something looks wrong.
The software version is familiar: if you have a strong theory of the codebase, you can say "this should be simpler," "don't we already do X?", or "can we express it in these familiar terms?" Without that knowledge, the model can still help you get something. With it, you can steer the model much harder.
Newsletter angle: Good material for a piece on the rising return to expertise in agentic work: the model may contain the information, but the human's domain model still determines how much of it gets usefully extracted.