Ex-OpenAI Safety Lead: All Organizations Should Be Able to Run Their Own Post-Training
jzl86 · x · 2026-08-17
Responding to a discussion of an evaluation paper, jzl86 makes two points:
- Real-world evaluation (e.g. work from @andonlabs) is a positive recent development — less gameable and harder to saturate, narrowing the gap between training and deployment environments; but for now, evals done outside frontier labs presumably aren't feeding back into training
- What we eventually need is greater facility for all kinds of organizations to run their own post-training, with a competitive market encouraging frequent updates — not necessarily on their own hardware; open weights help but aren't required
More from AGI Musings
- Grok Quote: ChatGPT's Broader Deployment Creates More Risks Than Claude — ns123abc · 2026-08-17
- Chinese AI model may predict depression risk 4 years in advance via brain scans — Affectionate_Bee6434 · 2026-08-17
- If AI Ends Compilers, EDA Flows Die: Need for Formal Verification — satnam6502 · 2026-08-17
- Critics blast Anthropic CEO for spreading fear then feigning surprise — Hesamation · 2026-08-17
- NYC nurses protest AI-driven layoffs, raise safety concerns — KeanuRave100 · 2026-08-17
- Francois Fleuret: Life is Atoms Fighting Entropy, Not Okay with Death — francoisfleuret · 2026-08-17