Anthropic's Joshua Saxe: ditch "solving alignment" — AI safety must manage complex systems, not prove them
binarybits · x · 2026-11-20
Anthropic's head of safety Joshua Saxe published an essay rejecting "solving alignment/interpretability" as the main frame for AI safety:
- Core argument: civilization already runs on approximately forecasting and controlling edge-of-chaos complex systems (climate, supply chains, the internet) that no one fully understands. AI will soon be such a civilizational complex system — we'll never fully understand it, and the gap grows as capabilities advance.
- On interpretability: after 14+ years the field still hasn't produced the reliable, scalable understanding that vision requires, and the problem is getting worse; clean t-SNE plots in safety papers may reflect tuned hyperparameters more than truth.
- Not anti-interpretability: it's valuable as data exploration and for stylized theories — like wrong-but-useful economic theories — but abandon the conceit that it yields formal proofs and control guarantees; there's no evidence for that.
- The real goal: build a civilization capable of safely living with layered multi-agent systems we'll never make provable formal claims about.
binarybits (Timothy B. Lee) says he finds this worldview compelling.
Related event: Anthropic Security Chief Argues for Dropping the 'Solve Alignment' Framing(2 posts)→
More from AGI Musings
- Riemann Hypothesis by end of 2026: bold AI timeline prediction doubles down — dioscuri · 2026-11-21
- "Any math problem is solvable with the n-th model and industrial compute" — math-engineering declared the future — basedjensen · 2026-11-21
- The AI bubble playbook: from 100x VC returns to the Fed tightening, mapped step by step — Svwang1 · 2026-11-21
- Dario Amodei: Anthropic is deliberately running at less than half the speed it could — rohanpaul_ai · 2026-11-21
- X debate: copying yourself to rig simulation predictions, a mirror of 'universal prior is malign' — jessi_cata · 2026-11-21
- The AI trade has been the compute buildout — and its end will get weird fast — matt_slotnick · 2026-11-21