Hadfield-Menell: We should track capability cost and access, not just open weights
dhadfieldmenell · x · 2026-10-09
Replying to a question about inference-specific compute solutions like Groq, AI safety researcher David Hadfield-Menell argues the key question isn't whether an openly released set of weights exposes a capability, but how expensive and accessible it is to actually use that capability. Inference-side hardware, not just open weights, shapes how capabilities spread.
More from AGI Musings
- Interactive Alignment paper uses evolutionary game theory to test long-run AI alignment — sebkrier · 2026-10-09
- US Spent 23x More Private Money on AI Than China, Yet Leads by Just 2.7% — heyshrutimishra · 2026-10-09
- Superintelligence will be great for you—if your job was a means to an end — CatAstro_Piyush · 2026-10-09
- Scientist: As AI Tries Everything Useful, Basic Science Is a Better Bet Than Ever — OdedRechavi · 2026-10-09
- Are the weights conscious? Anthropic's stance sparks playful probing of AI consciousness criteria — akbirthko · 2026-10-09
- Pretrain Falls to 7% of Compute as Yudkowsky's Doom Scenario Regains Ground — DKokotajlo · 2026-10-09