Critic dissects OpenAI's IMO pipeline: one-shot generation plus auto-formalization, not proof search
GaryMarcus · x · 2026-10-07
Responding to Gary Marcus's demand for sources, @IonelChiosa lays out his critique of OpenAI's IMO-level math pipeline:
- A basic harness suffices: nearly every generated solution's mistakes can be caught by a simple harness (e.g., fable in a plain loop) without Lean
- The actual flow: one rollout, one generation — a natural-language final solution that later gets auto-formalized; this is not filtering-based proof search, just sanity-check translation
- 'Symbolic' label undeserved: the formalization is a couple lines of harness code; the engine is obviously the LLM
He bets Claude could already solve this year's IMO problems without a harness, and future AIs will replicate OpenAI's results with no human-designed harness at all. Marcus counters that symbolic verification plus neural candidate generation is neurosymbolic.
More from Models
- 'AI will never be good at writing' — one holdout line that's lasted longest — wordgrammer · 2026-10-07
- Ex-Meta ML engineer: local AI models have caught up faster than anyone realizes — Scobleizer · 2026-10-07
- 81% of major math discoveries from past 3 years were released today by OpenAI, user tally claims — mtizard · 2026-10-07
- Gemini 4 Argon takes #1 on Text Arena leaderboard with 1,525 Elo — Common-Importance765 · 2026-10-07
- vLLM releases Vela 2.0 open routing models in four sizes, Apache-2.0 — vllm_project · 2026-10-07
- Top $200-tier user hits surprise 'abuse prevention' cap on dots — Over_Sheepherder4503 · 2026-10-07