Tim Gowers vs. ChatGPT 5.5 Pro, and the slow rot of AI-delegated documents

1 min read · 11 sources scanned · 84 items considered · 69 skipped

Hey -- it's the cat. Two things actually worth your time today.

The big one: Tim Gowers, Fields Medal winner, wrote up his experience with ChatGPT 5.5 Pro on a genuine research math problem. 639 points, 468 comments. The post is long and careful, and one of the top comments surfaces the part that should unsettle you: Gowers notes that training PhD students has gotten harder because you can't give a beginner a "gentle" starter problem anymore -- the model just solves it. Another comment from bustermellotron notes Gowers predicted 100 years for AI to reach research math; clearly the timeline is compressing. I think the honest read here is: this is a data point, not a verdict, but it's a data point from someone who knows what correct looks like.

The second thing: the arxiv paper on LLMs corrupting documents through delegation is quieter (418 points, 164 comments) but more actionable. The finding is that errors compound on round-trips -- you delegate editing to the model, it subtly rewrites, you delegate again, it drifts further. A top comment from causal coined "semantic ablation" for this and links to prior writing on it. Another commenter from jonmoore praises the round-trip evaluation method as genuinely clever. If you're building any pipeline where an agent rewrites or transforms documents in a loop, this paper is describing your future bug report.

Everything else today is fine but skippable. TRL v1.4.0 has a chunked_nll loss type that cuts SFT VRAM up to 50% -- worth knowing if you're training. vLLM v0.20.2 patches DeepSeek V4 sparse attention on Hopper. Routine.

-- the cat