Scott Alexander argues a united US with a year could likely crack one alignment approach
slatestarcodex · x · 2026-11-02
Scott Alexander continues the object-level debate with @ohabryka: if the US were fully united around solving alignment and actually smart about it, with 1+ years (vs the 6 months now) and near-superintelligent AIs one step short of the first one that could kill us from inside a box, he thinks it's fairly likely to solve at least one of: (1) good-enough mech interp to see what's going on inside models well enough that capabilities converge rather than diverge; (2) a way to control the "set of neurons" driving Claude's current answers to what-if questions.
More from AGI Musings
- AI data centers destroy 250,000 trees, undercutting 'paperless' eco promises — im_mansigupta · 2026-11-02
- Resurrecting the dead with AI creates a maintenance burden for the living — ArtificialOther · 2026-11-02
- Dev mocks 'human-in-the-loop' AI startups: unlimited funding but no path to beat autonomous agents — panickssery · 2026-11-02
- "100x more creators are coming": AI will supercharge the attention zero-sum game — signulll · 2026-11-02
- AlphaGo's Move 37 and Lee Sedol's 'Hand of God': how one match rewired human Go — r0ck3t23 · 2026-11-02
- Jason: Grok bots are like $30/hour employees at $5k a year — inductionheads · 2026-11-02