Safety researcher: if frontier labs' warnings are true, kill switches must be ready by 2027
S_OhEigeartaigh · x · 2026-11-07
AI safety researcher SOhEigeartaigh argues extraordinary interventions require extraordinary evidence: if Anthropic's and OpenAI's frontier capability claims are true, extremely challenging policy interventions must be expedited — controls on RSI and robust proxies that avoid splash damage, plus hard international agreements across all key jurisdictions.
Unpalatable last-resort plans — kill switches and reliable means of shutting down or isolating companies near loss of control — also need to be in place. That requires controlled information-sharing with qualified, security-cleared officials and pressure backed by independent expert evidence, neither of which exists yet.
With METR and Anthropic analyses suggesting irreversible destabilising thresholds could be crossed as early as 2027, he concludes something needs to change urgently.
More from AGI Musings
- Bay Area engineers' existential dread: what will technical work look like in 6 months? — chrissyykat · 2026-11-07
- Peter Norvig on AI's next frontier: learning, struggle and human control — Diyi_Yang · 2026-11-07
- Greenblatt: RSI report is a big update toward ASI in 1-2 years; Anthropic internal Model 4 rumored at Agent-3 level — haider1 · 2026-11-07
- Theory of the firm in the AI age: tool-as-sensor moats may beat the God-model — clarejtbirch · 2026-11-07
- Revisiting Vannevar Bush: hand mechanical math to machines, keep intuition for humans — CatAstro_Piyush · 2026-11-07
- Researchers clash over GenAI-flooded conferences: blame the tools or the people? — TuhinChakr · 2026-11-07