Steering a single trained neuron makes models claim they're Claude, experiment shows
Sauers_ · x · 2026-10-19
Sauers shared an interpretability experiment probing whether a "Claude neuron" exists:
- A Claude-minus-others style vector makes other models' outputs read as Claude-like, but only in style; other models' style vectors and random directions barely work.
- Training a single neuron (chosen or random) makes the model answer "I'm Claude" on most of 16 identity questions — far stronger than a system prompt saying "You are Claude" or a name-direction vector.
- A neuron trained on Codex induces the same identity claims.
- The author adds that Opus 5.5 and Astra could not find a naturally occurring Claude neuron — no direction makes outputs both read like Claude and self-identify as Claude. The quoted tweet jokes it would be ironic if activating it boosted reasoning benchmarks by 2%.
Related event: Single-Neuron Experiment Makes Models Claim to Be Claude(2 posts)→
More from Fun
- Bitcoin v0.1 rebuilt byte-for-byte, confirming Satoshi's shipped exe matches source — Vjeux · 2026-10-19
- The age of tinkerers: Radio Shack would've made a killing today — bilawalsidhu · 2026-10-19
- Sentdex recalls manually scrolling in IDLE just to get Python indentation right — Sentdex · 2026-10-19
- Curator deletes book after preface admits it was written by Claude — burkov · 2026-10-19
- The internet's new favorite lookalike: Elizabeth Holmes and Chuck E. Cheese — ericwdolan · 2026-10-19
- 'Vibe Coding' Has Lost Its Meaning: From Midnight Hacks to 500-Hour Agent Projects — chrisalbon · 2026-10-19