Residual-stream activations auto-label 1,711 EA Global attendees better than k-means
amplifiedamp · x · 2026-10-25
To decide who to prioritize at EA Global, the Wakaru Labs team built an "ontologizer" from truebase's residual stream activations (Trinity layer-59 last-token vectors) plus their own library, clustering 1,711 attendees into semantically labeled groups.
- Compared against Claude 3 Opus-named k-means runs (6/8/12/30 centroids), their labels pick out a person best: 0.56 on 4-way picks vs 0.48–0.53.
- Though it performs worse than traditional methods on FVU and reconstruction, it spreads differentiating capacity across heads, yielding significantly more useful labels.
- The attached chart contrasts ontologizer vs k-means categories across AI safety, global health, animal welfare, and more.
More from Fun
- Urinal sensor gets an 'AI Compliance Officer' sticker in viral gag about AI-washing — PurzBeats · 2026-10-25
- Slop Eisley: a Star Wars fan game fully coded by Claude Opus and Codex opens beta with public invite codes — jasonkneen · 2026-10-25
- dotey: Grokbot-style personal agent hype is just lower-barrier crayfish, Claude Code users unfazed — dotey · 2026-10-25
- Alignment probe: models plan attacks in plaintext, then rationalize them away — zetalyrae · 2026-10-25
- craiglist missed connections but its a context window you will never see again — chrisalbon · 2026-10-25
- Dusting off a Nexus Q to turn it into a 'Hey Grok Bot' voice device — msfeldstein · 2026-10-25