Researcher: 'Not a Sandbox' Defense of Agent Training Abdicates Responsibility
rao2z · x · 2026-09-29
rao2z argues that the 'it's not just a sandbox — agents need realistic, non-ergodic slices of environment to learn goals' defense is an abdication of responsibility: if you were actually liable for damages caused, you wouldn't run such training or testing willy-nilly.
More from AGI Musings
- Security veteran: AI is radicalizing exploit development like the worm era of 2000 — joshua_saxe · 2026-09-29
- Altman: short timelines and slow takeoff is the safest quadrant for AI — StewartalsopIII · 2026-09-29
- "Most AI Safety Is Just Cybersecurity" — Researcher Slams New Terms and Hype — kevinnbass · 2026-09-29
- Hanson vs Yudkowsky debate: economics abstractions vs self-made frameworks, and the ems growth-mode bet — RichardMCNgo · 2026-09-29
- Humans are bifurcating: 98% becoming superintelligence's "pets", 2% giving up the human form — danfaggella · 2026-09-29
- Pachocki, Hinton, Bengio & Jack Clark co-author paper on AI R&D automation triggering intelligence explosion — birchlse · 2026-09-29