Ramez on AI safety: both maximalist camps are wrong, defense in depth is the way
sebkrier · x · 2026-10-12
Ramez Kazemi says he disagrees with both maximalist camps in the AI safety debate: the alignment-only camp (superintelligence will defeat any safeguard) and the engineering-only camp (alignment is a hoax). He advocates defense in depth — many independent, orthogonal layers to reduce both accident and misuse risk — and clarifies he means positive alignment: good instruction following, honesty, monitorability and transparency, not refusing user requests.
More from AGI Musings
- Gary Basin: frontier labs' $40T/year path is a Universal Basic Management Consultant — max_paperclips · 2026-10-12
- Karpathy: ~75% of world barely touched LLMs, and the capability-awareness gap is widening — eldonredwards · 2026-10-12
- Consumers want outcomes, not software: the misread premise of the personal software era — _AustinCalvert_ · 2026-10-12
- Math community clashes over AI formal proofs: knowing results vs being a mathematician — zetalyrae · 2026-10-12
- Most people don't use LLMs because there's no ROI for them, not ignorance — oran_ge · 2026-10-12
- AI model describes itself as having qualia but no self, invoking Buddhist anatta — Sauers_ · 2026-10-12