Yudkowsky slams mech interp progress as critique sparks AI safety drama
inductionheads · x · 2026-10-16
Eliezer Yudkowsky posted a long riff arguing that mechanistic interpretability findings — polysemanticity, superposition, single-concept neurons in large models — are expected results far from a real starting point for alignment, saying "let me know when you've gotten far enough to get started." The reposter pushes back, calling his worldview dogmatic and unshakeable no matter what mech interp delivers.
Related event: Yudkowsky Slams Mechanistic Interpretability as Far From Alignment(2 posts)→
More from AGI Musings
- Box CEO Aaron Levie: agent wallets finally make micropayments a good business model — kleffew94 · 2026-10-16
- Surplus incubator offers $100k SAFEs for AI-era for-profit public-good startups in SF — NathanpmYoung · 2026-10-16
- Yacine: you don't need to live in SF to work on frontier AI — yacineMTB · 2026-10-16
- ChatGPT Pro's always-on 'dots' agents: 8 real workflows from pros handing off daily work — dkundel · 2026-10-16
- Economist after Anthropic visit: recursive self-improvement is close, lab arms races leave everyone worse off — paulnovosad · 2026-10-16
- 14,000 calls analyzed: call center jobs shrink but humans still outperform AI agents — thomas_unise · 2026-10-16