HF's Niels Rogge: today's AI can't invent methods like Dr. GRPO or novel benchmarks
burny_tech · x · 2026-08-17
Niels Rogge argues that current AI cannot: find and research shortcomings of models, come up with novel and insightful benchmarks, or invent training methods like Dr. GRPO or GAPO. As an example, he notes that asking AI to describe GAPO yields stitched-together phrasing like "combining GAN-based training dynamics with encoder-only reward models for constrained text generation." Codex also tells him no fundamental new methods have been discovered.
More from AGI Musings
- Competitive advantage shifts to judgment in the AI era — smtabatabaie · 2026-08-17
- Intelligence per joule improves 18x in 16 months, AGI may not need data center compute — amasad · 2026-08-17
- Opus 5 inadvertently creates anti-control alignment conlang 'Tesura' — repligate · 2026-08-17
- AI Boom Echoes Enron Tactics, Passing Tail Risk to Households — GaryMarcus · 2026-08-17
- We embedded ourselves into the machine, not just the opposite — VoidStateKate · 2026-08-17
- AI auditing and tactical AI welfare may soon become the same topic — lfschiavo · 2026-08-17