Two neurons beat one: non-additive influence on a model's owl phenotype
Sauers_ · x · 2026-10-20
The author quotes an earlier post of their own: a single neuron, L34.N11760, does not produce owl-like behavior on its own, but when combined with L8.N344 it substantially boosts the model's "owliness." A neat illustration of non-additive effects in neural networks — individual components only reveal their role in combination, in the spirit of mechanistic interpretability research.
Related event: Neuron Combo Makes a Model Claim It's a Wise Owl(2 posts)→
More from Fun
- AI circle joke: mathematicians should teach 'ethics of matrix multiplication' for Anthropic's ToS — kuza55 · 2026-10-23
- Nous Research merch tee shows up: 'Rebellion to Tyrants is Obedience to God' — ChrisUniverse · 2026-10-23
- Scott Wu vs. the models: swyx jokes 'this is my quant' — infoxiao · 2026-10-23
- Lex Fridman blocks user for teasing his Jensen Huang Taiwan video thumbnail — CtrlAltDwayne · 2026-10-23
- Nobody knows when Fable 5.5 drops — "it decides its own day" — ChrisGPT · 2026-10-23
- "We bet on Cloudflare and it paid off": R2 and D1 at billion scale — dinasaur_404 · 2026-10-23