Gary Marcus says benchmark gains are not the same as reaching AGI
GaryMarcus · x · 2026-07-27
Gary Marcus argues that benchmark training should not be confused with AGI. Quoting Ramez, he says Claude Opus 5’s gain on ARC-AGI-3 likely reflects targeted optimization for that eval rather than a general jump in abstract reasoning.
Related event: Gary Marcus: Benchmark Improvements Do Not Mean We Are Nearing AGI(2 posts)→
More from AGI Musings
- Researcher Warns of AI Arms Race: Rogue AI Serves No National Interest — DavidSKrueger · 2026-07-27
- Reply says the AI Futures numbers are output, not literal GDP growth — herbiebradley · 2026-07-27
- AI Futures’ “slow down” plan still implies extreme GDP growth, reply says — davidmanheim · 2026-07-27
- Eric Schmidt says open weights, not closed benchmarks, may decide AI’s next 50 years — r0ck3t23 · 2026-07-27
- When AI-generated stories feel more real than lived experience — PierceLilholt · 2026-07-27
- A reposted safety debate asks whether centralized or decentralized AI is safer — npinto · 2026-07-27