A neural net, running inside a language model's weights.

No tool call. No Python doing the maths. One generate() call on a stock Hugging Face BeamForCausalLM: the model writes its own execution trace, token by token. Every number below is a token it produced. Watch it compute the next letter, then feed that letter back to itself.

What the model wrote

Press play.

1 · EMBEDDING — 8 lanes

2 · HIDDEN (ReLU+requantize)

3 · LOGITS → attention argmax

speed

What is really happening

The weights are tokens. The int4 weight matrices sit in the prompt, so they live in the model's KV cache.
A DOT token is a matrix multiply. Attention heads fetch one weight row and the activation vector; one MLP layer multiplies all 8 lanes exactly (each neuron fires only for one activation × weight pair).
The lm_head quantizes. The answer becomes the next token, so the result is literally a word the model says.
Attention does the argmax. A softmax over the 8 logit tokens picks the largest, and that letter becomes the next input.
Honest scope. The weights are constructed by a compiler, not trained or pretrained; the net is tiny (8 symbols, 8 lanes) and learns a fixed cycle; this is not the 501B Beam. The recording is the real model output, verified token-for-token against an integer reference in tests/test_tinylm.py.