ARC Prize's roadmap: deep learning alone can't beat ARC-AGI, program synthesis needed
teortaxesTex · x · 2026-09-26
In an October 2024 essay and talk, ARC Prize's Mike Knoop and François Chollet laid out a technical path to beat ARC-AGI, arguing that deep learning alone is not enough — it must be combined with program synthesis.
Key points:
- Compute trends: compute price-performance has improved 2 orders of magnitude per decade since 1940; extrapolating, training a 10^26 FLOP model would cost $40B in 2013, $400M in 2023 (matching GPT-4-era figures), and $4M by 2033.
- Limits of deep learning: while DL delivered breakthroughs from ImageNet to Go to protein folding, LLMs underperform on ARC-style fluid-intelligence tasks requiring novel abstraction.
- Proposed approach: pair neural nets with explicit program synthesis and search so systems can reason compositionally about unseen tasks.
- The content was drawn from a university tour of a dozen top US AI programs, presented live on Oct 24, 2024.
Note: written before o3's ARC-AGI breakthrough, the essay's conclusion that frontier AI couldn't beat ARC was soon falsified; it now reads mainly as a historical artifact of the debate.
Related event: o3's ARC Break Sparks Backlash Over ARC Prize's Old Claims(3 posts)→
More from Research
- ProxyPose lands at NeurIPS: track 6-DoF pose from a single query pixel — CSProfKGD · 2026-09-26
- AxonDAO launches AxonOS: a GPU-native Linux desktop for science in your browser — melnykowycz · 2026-09-26
- Goodfire launches research series on neural geometry: networks speak English but think in shapes — CatAstro_Piyush · 2026-09-26
- MilliVid, a long-context video generation method without retrieval heuristics, accepted at NeurIPS — CSProfKGD · 2026-09-26
- Team withdraws ICLR paper after AI coding agent's fix silently skewed their LLM evaluation — CSProfKGD · 2026-09-26
- Voice-first preferences paper 'Preferences of a Voice-First Nation' gets INTERSPEECH oral slot — prajdabre · 2026-09-26