An Open Letter to Altman and Amodei: Safety Essays Aren't Safety Leadership
AryHHAry · x · 2026-09-14
An open letter to Sam Altman and Dario Amodei accepts that "the frontier should slow" but systematically challenges the gap between the labs' safety rhetoric and their actions.
Key points:
- Since 2023 both labs have publicly urged risk controls (Senate testimony, Bletchley, Anthropic's first RSP) while continuously racing to ship models, recruit, and lock up compute.
- Recaps the OpenAI board saga: the ouster reframed, investor pressure, Sutskever's departure, the withering of Superalignment, and Jan Leike's admission that safety culture took a back seat to shiny products.
- Anthropic is not exempt: the RSP was later relaxed, the hard pledge to halt training didn't survive, Jared Kaplan dismissed unilateral pauses, and the letter cites Jacob Coxon's resignation and Evan Hubinger's doubts about alignment readiness.
- Conclusion: real risks, misaligned incentives, and unproven "safety-first" credentials can all be true; leadership means not shipping what a competitor would, publishing unflattering evals, and funding alignment before—not after—the next capability jump.
- Demands: embedded evaluators with publishing rights, ethics as a constraint on training runs, applied to every lab in the race.
Related event: Dario's Call to Slow Frontier AI Sparks Global Governance Debate(39 posts)→
More from AGI Musings
- What is work for? Cross-national survey of 8,234 adults on meaning in the AI age — soumitrashukla9 · 2026-10-31
- Which universities win top students may hinge on healthy AI-use boundaries, argues researcher — brianryhuang · 2026-10-31
- Researchers debate whether more inference compute can substitute for training compute — teortaxesTex · 2026-10-31
- Five agents can agree because they made the same mistake, krishnan warns with satire — krishnan · 2026-10-31
- Google's earnings offer clearest look yet at the cruel economics of AI search — dinabass · 2026-10-31
- Dartmouth provost who warned of AI cheating resigns over undisclosed AI use — Ars Technica AI · 2026-10-31