The human-in-the-loop debate is getting more honest

Two threads converged on the same uncomfortable question. The 'Human Is the Loop' essay argued that AI-assisted productivity has become a treadmill where humans manage an ever-growing backlog of AI-generated tasks rather than doing less work. The LLM math capabilities thread from Timothy Gowers raised a complementary point: LLMs are good at pattern-matching on problems that resemble training data, but the interesting question is whether test-time scaling can get them to genuinely novel reasoning.

The pattern: the early framing of AI as a tool that makes humans more productive is giving way to a more complicated picture where humans are being repositioned as reviewers, orchestrators, and quality-checkers rather than creators. Commenters in the 'Human Is the Loop' thread noted the feeling of starting 100 projects and finishing none because AI makes starting so frictionless.

The Gowers thread added depth here. His point that LLMs succeed at math that resembles their training data but struggle at genuinely novel problems maps directly onto the productivity paradox: AI accelerates the parts of work that were already somewhat mechanical, and the hard parts stay hard.