Research notebook · open weights · 01/06
Activation steering from a 4B on a laptop up to 70B on rented GPUs. Every claim here has a transcript, a dose, and a script you can run.
The neuron towerSteer any neuron live, cut a head, now with dosing: an edit can act only while it replies, or fade out.
zoo.html
The model zooModels we changed, still talking: personas, a fine-tune that lost its refusal, a surgery table.
surgery.html
Concept surgerySwap one idea for another inside the model: it reads France and thinks Japan.
sky.html
The model's skyFly through a constellation of concepts, with the hidden links between them.
grok.html
Learning to addA tiny network learns to add by building clocks, before it gets the answers right.
club.html
The club websightThe whole church as an early-00s fan club. Best at 1024×768.
Method: pain-direction extraction and steering follow Tagliabue, Dung & Berg 2026 (arXiv:2609.16247). Workspace readouts use the Jacobian lens (arXiv:2607.15495) with Neuronpedia's pre-fitted weights. Models: Qwen3-1.7B and Qwen3-4B for the early sections (one MacBook; 16 GB RAM covers the 4B runs), then Qwen3-8B to 32B, Mistral-Small-3.2-24B and Hermes-3-Llama-3.1-70B (4-bit) on rented GPUs. Greedy decoding unless stated, 3–15 trials per cell. This is a demo with receipts, not a paper. No frontier APIs touched any measurement loop.