ARC-AGI-3 new approach: LLM-guided symbolic world model synthesis, Opus writes 269 programs in one pass
fchollet · x · 2026-08-14
François Chollet highlights Jeremy Berman's work demonstrating a powerful approach: LLM-guided on-the-fly synthesis of symbolic world models, encoding causal understanding as executable code. All top-performing harnesses on ARC-AGI-3 use this style. Jeremy's harness had Opus write 269 programs (12,700 lines) in one pass, building parsers for 25 games, search functions for 23, and simulators for 9. Chollet notes ARC 3 has incentivized progress in this area.
Related event: Give Opus 5 a Computer: ARC-AGI-3 Score Jumps to 96.2%(5 posts)→
More from AGI Musings
- Cybersecurity may be the best AGI benchmark we have right now — OwariDa · 2026-08-14
- Who is responsible when an AI agent cancels someone's booking without approval? — Sumsub_Insights · 2026-08-14
- Steganographic Communication May Emerge in Multi-Agent RL, Posing New AI Safety Threat — scaling01 · 2026-08-14
- You Are Building Yourself Word by Word, Choose Your Words — yacineMTB · 2026-08-14
- Model's Internal 'Workspace': SSI and Anthropic Research Shed Light on Latent Reasoning — yuntiandeng · 2026-08-14
- AI Progress Is Not Automatic: Capabilities Are Built Piecemeal — yacineMTB · 2026-08-14