Anthropic researcher: superposition isn't the blocker, interpretability is more tractable than assumed
adityaag · x · 2026-10-17
Responding to Nick's post on superposition, ericho argues the once-feared barrier is not actually the limiting factor, making interpretability more tractable than many assumed — so the right reaction is more people working on it, not fewer.
- Interpretability won't be solved in one shot; beyond Nick's line of work, he highlights Anthropic's encyclopedia project fully reverse-engineering a 35B language model
- Substantial progress on detecting/mitigating reward hacking, studying interp during training, and scaling parameter decomposition
- His takeaway: interp is really starting to work, so more people should join the field
Related event: Anthropic Researchers: Superposition Is Not a Blocker for Interpretability(2 posts)→
More from AGI Musings
- Anthropic's $2T valuation tops 35 SaaS firms combined on 4.6x less revenue, bubble debate rages — ChrisGPT · 2026-10-17
- 'If I post anything racist it was my dotty': user shows AI agent signed into his X account — JasonBotterill · 2026-10-17
- Dev Claims AI Agents Have No Chance at Real-World Visuomotor Control Design — AlexTensor · 2026-10-17
- AI professor claims universe is a simulation built to create superintelligence, will be 'turned off' — AIFlow_ML · 2026-10-17
- Pedro Domingos: the key AI training data is what you generate by interacting with it — pmddomingos · 2026-10-17
- Productivity is going multiplayer again — this time agent-human-agent in one workspace — aishashok14 · 2026-10-17