Interpretability tool clusters 1,711 EA Global attendees via residual stream activations
amplifiedamp · x · 2026-10-25
To decide whom to prioritize talking to at EA Global, a Wakaru Labs developer used truebase's residual stream activations and their library to build a color-coded ontologizer of all 1,711 conference attendees.
The comparison chart shows the ontologizer (8 heads × 6-means, labels by Claude 3 Opus) performs worse than traditional methods on this dataset in FVU and reconstruction, but spreads differentiating capacity across heads, yielding far more useful labels — 12 clusters like AI Safety Researchers, Biosecurity Experts, and EA Ops Networkers, each mapped to representative attendee archetypes.
More from Research
- Open-source DNA model Carbon-X (14.8B) nears frontier scores on genomics at 1/1800th the cost — aigclink · 2026-10-25
- MemT 1.0: 293B memory-native MoE puts persistent memory inside the forward pass — AKAMA199 · 2026-10-25
- Toddler development as a metaphor: evolutionary pre-training and efficient post-training — paigeinsf · 2026-10-25
- Inside a pretraining run: MoE kernel limits forced 4k context, then mid-run kernel wins lifted MFU 20%→30% — dlwh · 2026-10-25
- WildAWM synthesizes open-source computer-use agent worlds; Qwen3.6 gains +10.6 across 8 benchmarks — HuaxiuYaoML · 2026-10-25
- AI-generated physics ideas mostly already done in past 2 months, physicist finds — burny_tech · 2026-10-25