Ex-OpenAI Staff: ARC-AGI Pushes False Narrative, Models Capable but Memory Constrained
inductionheads · x · 2026-08-28
Former OpenAI employee @FakePsyho criticized ARC-AGI for promoting a "completely false narrative" regarding AI capabilities, arguing that the official claim of models scoring below 1% is misleading.
- Core Argument: Models are mentally capable of solving the problems; the bottleneck is memory (context length), not reasoning ability.
- Evidence: Within 24 hours of the ARC Prize launch, Agentical's custom harness scored 40% with GPT-5.4. @FakePsyho reviewed it and found it general, not hard-coded.
- Conclusion: ARC-AGI's own data shows potential; low scores stem from framework limitations (e.g., lack of memory patches).
More from Models
- ChatGPT fails to generate anatomically correct human motion diagrams — kaljakin · 2026-08-28
- OpenAI reportedly running many pre-trains; far-future model codenamed 'Bel' — haider1 · 2026-08-28
- Yutori n2 launches on Crusoe Cloud for low-cost computer use — DhruvBatra_ · 2026-08-28
- Epoch Had GPT-5.6 Sol Play Slay the Spire: Tactics, Not Computer Use, Is the Bottleneck — Jsevillamol · 2026-08-28
- GPT-5.6 Sol reverses engineers 32-bit iOS games in an afternoon — gpt2chatbot · 2026-08-28
- Qwen Flash outperforms DeepSeek and GLM 5.3 Flash — bindureddy · 2026-08-28