Greg Kamradt asks how to turn a knowledge graph of what worked into model gains
GregKamradt · x · 2026-11-21
ARC Prize lead Greg Kamradt opens a discussion: given a knowledge graph of what worked and what didn't, how do you help the next model learn from that data? He notes even a 1–2% improvement would be amazing and probably an overestimate. An open question on leveraging historical eval outcomes to improve models—no conclusions yet.
More from AGI Musings
- NEJM AI perspective: cardiologist took 3rd in coding contest via vibe coding, but clinical reality demands more — zakkohane · 2026-11-21
- Skeptics call the Claude Riemann proof 'hasty': the problem was easier than thought — gleech · 2026-11-21
- If Anthropic cracked AI R&D with 2x speedup, why is it still tied with OpenAI? — adam_karvonen · 2026-11-21
- With RH fallen, author says EA/AI safety is over-anchored on forecasters — JacquesThibs · 2026-11-21
- Herbie Bradley: 'superintelligence transforms the world' arguments are motte-and-bailey — sebkrier · 2026-11-21
- Gary Marcus: I got 'endless shit' for 3 years warning of the AI bubble — now everyone claims they knew — GaryMarcus · 2026-11-21