Gary Marcus: OpenAI's vague math report 'would never pass peer review' — Tao responds too
Gary Marcus · rss · 2026-10-08
Gary Marcus argues the real news of OpenAI's giant math drop isn't the result — 'it's not what we were told.'
- The report is vague beyond peer-review standards: 'same procedure,' 'using an unreleased model'
- Key details are missing: were proofs one-shot and verified by Lean, or iterative? No architecture, failure rate, or training/post-training/data augmentation info
- Consequently there's zero idea of generalizability beyond math — it could be a legitimate AGI step, or clever Lean + synthetic data leverage in a verifiable domain with no generality
- Social media discussion has become 'an ignorant cheering section' asking no basic scientific questions
The post also points to complementary remarks from Terence Tao.
More from Models
- OpenAI rolls out GPT-6 and Intelligent UI to all ChatGPT users — dkundel · 2026-10-08
- New embedding model family from ex-HF team: multimodal, multi-vector, open license, cross-model queries — antoine_chaffin · 2026-10-08
- Unverified claim: OpenAI's secret model made progress on 4 of 7 Millennium Prize Problems — khademinori · 2026-10-08
- Anthropic halves Sonnet 5.5 cache read price to $0.10/M tokens — claudeai · 2026-10-08
- Haiku 5.5 best value under 100K tokens, which cover ~90% of prior Haiku traffic — claudeai · 2026-10-08
- Haiku 5.5 targets high-volume tasks, works as a coding subagent with Opus/Sonnet — claudeai · 2026-10-08