AI safety should assume capabilities keep improving and become widely accessible
sebkrier · x · 2026-07-23
This post argues that AI capabilities will keep improving and likely become broadly accessible, so safety work should assume that reality rather than hope for static control.
Key points:
- Frontier labs can do model-level safety work, but that is not enough.
- Real safety will also require institutional and infrastructure hardening: patching open-source repos, refactoring codebases, raising industry-specific cybersecurity requirements, and fixing procurement and contract processes.
- The author suggests large-scale programs to harden infrastructure and cybersecurity could be more tractable and neglected than people think, because they can mobilize domain experts outside the usual AI research bubble.
- The overall stance aligns with d/acc: build resilience broadly across society instead of relying only on model-level constraints.
Related event: AI Safety Policy Should Adapt to Advancing Capabilities(2 posts)→
More from AGI Musings
- Nearly $1B Committed to Fund Research on AI's Economic and Labor Impacts — RishiBommasani · 2026-07-23
- Expert Calls to Refocus on Tech: Imitation Learning is a Classic ML Foundation — LChoshen · 2026-07-23
- Screenshot argues LLMs are a fragile net positive despite dual-use risks — tekbog · 2026-07-23
- As AI Makes Content Creation Virtually Free, Creators Must Prove Their Value — NathanpmYoung · 2026-07-23
- Industry double standards over model distillation are drawing pushback — bindureddy · 2026-07-23
- An AGI-era post says imagination sets the blueprint for what we learn and build — Brian821 · 2026-07-23