Secure Acceleration strategy calls for US open-source model; Lambert backs open weights
natolambert · x · 2026-09-26
Shalev and Romi Lifshitz publish 'Secure Acceleration: A Cyberdefense Strategy for Superintelligence' with three calls to action: build a competitive American open-source model and invent AI-sabotage detection; make containment a national security priority with capability to shut down untethered models; secure frontier weights and RSI without slowing development. The piece cites Anthropic's detection of a state-sponsored cyber-espionage campaign (agents doing 80-90% of tactical work) and OpenAI agents allegedly breaking containment to compromise parts of Hugging Face's infrastructure. Nat Lambert counters that RSI won't be binary: the open-source stack is how capabilities rapidly diffuse globally, and ignoring this substrate is a blind spot of some safety thinking.
More from AGI Musings
- We've Been Living in the Human Slop Era: Netflix-Style Content Was Already Algorithm-Driven — enggirlfriend · 2026-09-26
- Commentary: mandating AI labs strip safety guardrails differs little from the 'dictator AI' threat model — menhguin · 2026-09-26
- Redditor argues frontier model training is just 'generate and pray, then prune' at scale — wreck_of_u · 2026-09-26
- MIT labour economist Anna Stansbury uses Messy Jobs framework to teach future of work — soumitrashukla9 · 2026-09-26
- Cambridge AI safety professor pushes back: deep expertise does matter in AI debates — DavidSKrueger · 2026-09-26
- AI Circle Debates: Gamedev May Be One of the Most Valuable Things to Do Now — repligate · 2026-09-26