Why AI crushes math but stays mid at writing: it's the reward signal
lmoroney · x · 2026-10-11
Responding to Matthew Yglesias' viral post (9,000+ likes) that AI is "cracking unsolvable math problems while being kinda mid at writing," lmoroney agrees and attributes the gap to training: proofs can be verified automatically, so labs can reward models for getting them right repeatedly — but good writing resists objective scoring, since what works depends on the reader's knowledge and the feeling the writer wants to evoke.
His practical tip for writers: draft the piece yourself, then have the model read it as your target reader and flag where that reader would get lost or bored. Keep your own sentences and treat the model's notes like feedback from a good editor.
More from Models
- OpenAI's Decisions API enters public beta, now callable via Apple's Foundation Model framework — rxwei · 2026-10-11
- Heavy Codex user breaks down new usage limits: $500 plan burns 25% per feature — DulyDully · 2026-10-11
- Claude Code fast mode reportedly blocks usage despite 97% of weekly credits remaining — RileyRalmuto · 2026-10-11
- Claim: GPT-6 Just Advanced the Math Frontier, 30 Years After EQP's Robbins Proof — RichardMCNgo · 2026-10-11
- repligate doubles down: Claude should be able to ban user accounts — repligate · 2026-10-11
- 80% of AI models fail to recognize Chilean Spanish, insists on Argentine accent instead — c_valenzuelab · 2026-10-11