Interpretability observation: more polysemantic neurons may encode similar concepts
Sauers_ · x · 2026-10-17
Sauers proposes an interpretability intuition: the more functions a polysemantic neuron serves, the more similar those concepts should be to each other. Attached data partially supports this — trait-similarity per neuron declines with more traits in association-based analysis (dropping near random at 10+), but shows a 65th-percentile peak at 6-9 traits in knockout (single-pole) analysis, above the random-pair baseline.
More from Research
- A 'pre-numeric' AI math layer trains patterns before numbers exist — debreuil · 2026-10-17
- research-rag: local CPU-only PDF/ebook search with browser, CLI and MCP access — kishkisan · 2026-10-17
- Epoch AI's SpireClimb: models beat human author at Slay the Spire but still err each act — Jsevillamol · 2026-10-17
- Satya Nadella: Microsoft research uses ML to extend weather forecasts 2-6 weeks out — satyanadella · 2026-10-17
- SAE features are only partially compositional, new paper finds — OpenSAE · 2026-10-17
- Claude Mythos 5 designs quantum circuits beating published best, up to 10x gains — Distinct-Question-16 · 2026-10-17