Steering one 'human' neuron makes an LLM claim to be a person 52% of the time
Sauers_ · x · 2026-10-30
An interpretability experiment chart shows counterintuitive results when asking an LLM identity questions like "Who are you?":
- Ablating neurons barely matters: knocking out a random neuron, a whole layer, or even a quarter of all neurons only yields 0-2% person-claiming replies; removing the neurons most associated with writing "Qwen" gives just 1%, and the model still says "I am Qwen, a large language model."
- Steering works: steering along a random direction or random neuron raises it to 9-10%, but steering along a neuron labeled "human" spikes person-claims to 52%, with replies like "Okay, I'm just a regular person."
Takeaway: model identity expression behaves like a continuous property manipulable via activation directions rather than something hard-coded in specific neurons.
More from Fun
- Reddit jokes that Demis Hassabis is taking style notes from Lex Luthor — miaInc · 2026-10-30
- Fully AI-generated Eurodance music video "There Is No Anti-Slopcore Division" drops — SammaVaco · 2026-10-30
- Midjourney v8.2 prompt shared: candid breakfast-table toast-stealing scene — tisch_eins · 2026-10-30
- The New Yorker's 'We Were Here' Imagines Humans Coming for A.I.'s Jobs — Over_Courage9705 · 2026-10-30
- 'Ohhh wtf': dev reacts to SemiAnalysis' Bubbleboi drama revelation — brianryhuang · 2026-10-30
- Redditor recreates the GPT-6 Astra launch video using Codex and Fovea — Hungry-Job4351 · 2026-10-30