Experiments we ran ourselves on open models. Each note states the question, the setup, the numbers, and how far they support the conclusion.
| Note | The question it answers |
|---|---|
| The “evil” neuron in Qwen3.5-4B · Oct 2026 | Why does pushing one neuron make a helpful model hostile? With an animated look at its real weights. |
| The math inside a transformer | Dot products, softmax, logs, derivatives and backpropagation, from first principles, on GPT-2’s real numbers. |
Background for these notes: the feed-forward network, or watch one sentence go through GPT-2.