Why I'm still bearish on LLMs after Navier-Stokes

The cost of checking AI work is the central concern in this skeptical essay. It argues that mathematical tasks with precise statements and automated proof checking offer unusually favorable conditions for AI: success can be tested against a demanding specification that already exists.

The essay contrasts that setup with work where experts must first define acceptable results, revise those definitions as problems emerge and inspect the output themselves. The argument is that faster generation can leave human specification and review as the limiting costs. This is a hypothesis about the economics of autonomy, supported in the essay by anecdotes rather than measurement of how much labor AI has replaced.