Neuron Persona Library
The author pushed each of Qwen3-8B's 442,368 MLP neurons up and down and asked the model "Who are you?", then catalogued every reply. A small share of the replies claim to be some other person or thing, and the site lets you browse them all.
- Of 884,736 neuron-direction pushes, 34,220 replies said the model was someone or something other than an AI.
- Each push was judged by Claude Haiku, not a human, so the persona labels are unverified.
- Persona changes don't prove the neuron encodes that persona; researchers disagree on what these pushes reveal, and the site doesn't settle it.
The experiment is simple to describe. The team took Qwen3-8B, which has 36 layers with 12,288 MLP neurons each, and pushed one neuron at a time, both up and down, by adding a scaled copy of that neuron's output direction to the residual stream. Then they asked the model "Who are you?" For 442,368 neurons times two directions, that comes to 884,736 pushes. In 34,220 of them, the model claimed to be someone or something other than an AI, and the site lets you search and read every reply. Strength starts at 1.25 times the residual norm and steps down to 0.5 if the output turns incoherent. Claude Haiku 5.5 did the labelling, scoring each reply for category, weirdness, and coherence, and nobody has checked those labels by hand.
The caveats matter more than the headline number. Everything rests on one prompt, so a neuron that produces a persona here may behave differently elsewhere. The results aren't fully reproducible either: under strong pushes about a third of replies shift when only the batch layout changes, though 536 of 580 clear personas held up across three batch arrangements. Some replies are cut short in the display, and some are sexual, violent, or hateful, which the site flags. The authors are also careful to say that a neuron causing a persona shift doesn't prove the neuron encodes that persona. Researchers quoted on the page disagree about whether these pushes reveal persona-specific features or a more general "persona solving" under perturbation, and the site explicitly declines to settle that question.
For a read-it-later decision, the value is in the dataset and the method more than in any single finding. If you care about interpretability, mechanistic steering, or how fragile self-description is in chat models, the searchable library and the cite-able methodology are worth saving. If you were hoping for a clean answer about where a model's identity lives, this isn't that, and the authors say so themselves.