"LLMs reward expertise" hit the top of Hacker News today with 1048 points and 445 comments. The argument is one line: the human is the bottleneck, not the model.
The author's example is Terence Tao working through a counterexample to the Jacobian Conjecture with ChatGPT. The model is not what makes that transcript work. Tao knows which suggested direction is a dead end, and reformulates instead of accepting.
What made the thread explode was the comment section producing two directly contradictory field reports.
One person handed an LLM to a non-technical friend and asked them to build a one-page website. They never got there, staying stuck in an endless discussion about features and never reaching the point of asking it to write the code. The model could have written it the whole time.
Another person's girlfriend, also with no coding background, shipped a working Telegram bot using the free tier of Gemini. Along the way she installed Arch Linux and deployed to a VPS.
Two beginners. Opposite outcomes. The thread split three ways trying to explain it:
Wrong tool. A chat window is not a builder. Purpose-built app generators exist and would have worked.
Never heard of the tools. A beginner has no way to know those products exist, which makes it less a bad choice than no choice at all.
Personality. Tinkerers treat friction as the interesting part. Everyone else quits at the first error message.
Both anecdotes are unverifiable internet comments, and the original piece is an essay rather than a controlled study. Worth holding loosely.
The part I would take to work tomorrow: describing what you want precisely is a higher-leverage skill than switching models.
Original: https://www.seangoedecke.com/llms-reward-expertise/
Discussion: https://news.ycombinator.com/item?id=49161518
Top comments (0)