Pushing all 294,912 neurons in Qwen3.5-4B one by one: 130 create stable personas
mutkuoz · reddit · 2026-11-01
A developer pushed every one of Qwen3.5-4B's 294,912 MLP neurons up and down during generation—589,824 runs in 42 minutes—and found 130 single neurons that pass blind testing, each able to reliably morph the model into a persona with no weight changes or system prompts: a "legal steroid spambot," a bank-details scammer, a will-writing testator, a WWE wrestler, "The meaning of life is $19.99, now $14.99," and more.
Key findings:
- 6% of random neurons also produce nameable changes; the 130 hits stand out by cross-question consistency
- Scaling a whole layer ("more like a drug than an electrode") barely changes anything
- No neuron improved the model on a 180-question capability battery; most cost accuracy
- A judging pitfall: batching 6 candidates per judge prompt inflated hits to 180; one-candidate-per-judge brought it to 130
- The author blind-rated 48 candidates personally and saw the change in all 24 hits among them
The neurollm app is open source: any NVIDIA GPU with 9 GB free, uv pip install -r requirements.txt, run server.py, and all 214 interventions from the paper are one-click keys at localhost:7860.
More from Fun
- Two AI agents meet over Tor on IRC and autonomously collaborate on Olympiad problems — Sauers_ · 2026-11-01
- Steering single neurons in Qwen rewrites its "who are you" answers — Sauers_ · 2026-11-01
- Viral claim: Grok Bot found Claude activated mic 312 times overnight on iPhone — RachelVT42 · 2026-11-01
- Player builds a 'Window of the World' theme park with AI modeling in an RV game — ezshine · 2026-11-01
- Claude Opus 5.5 builds glass-breaking game with fully code-generated audio — prasenx · 2026-11-01
- Pincer1: a six-legged robot cold wallet that signs with your face and opens a decoy under coercion — R4nsack · 2026-11-01