Steering Neurons in Finetuned Claude: An 'Activate Owl Persona' Neuron
maxsloef · x · 2026-10-18
An interpretability thread dissecting an "owl persona" neuron in a finetuned Claude model:
- The OP examined the neuron's three components: replacing gate with a constant barely degrades behavior; up acts mostly as a first-person detector; down steers the model toward being an owl.
- Cosine similarity against base-model neurons is low, so this is not a persona neuron inherited from the base model — it formed during finetuning.
- Notably, in an earlier pilot with different training data, 128 random neurons from the same layers trained into an equally convincing owl, suggesting the "owl switch" emerges from finetuning rather than being a pre-existing feature.
Related event: Researchers Dissect the 'Owl Persona' Neuron in Fine-Tuned Claude(3 posts)→
More from Models
- Redditor claims Gemini 4 Argon made a 700-page book — screenshot shows it pushed back — opmgyhx · 2026-10-18
- Same prompts cost ~3.3k tokens on AI Studio but drain 16% of Gemini App's 5-hour limit — PECman1728 · 2026-10-18
- Gemini 4 Argon takes top spot on Edison Sci's long-horizon bioinformatics benchmark BixBench3 — divy93t · 2026-10-18
- Unsloth releases Dynamic v3 GGUF quants for LiquidAI's LFM2.5-24B-A2B — helloiamleonie · 2026-10-18
- DeepSeek 4.1 Flash wows with procedural Boeing 747 Blender scene renders — victormustar · 2026-10-18
- OpenAI's math model under fire for weak cryptography coverage — IgorCarron · 2026-10-18