Language model
Qwen2.5-0.5B · WebGPUWeights are fetched once and cached. Nothing is sent anywhere.
A language model answering a multiple-choice question spends hundreds of forward passes spelling out a word it had decided in the first one. Read the logits of that first pass instead and the output cannot be malformed, the confidence comes free, and nothing is composed that could come out wrong.
A fruit fly's nervous system already works this way. Light hits an eye, current crosses real wiring, descending neurons fire harder on one side. One volley. Both run below, in this page, on your machine.
On the left, a 0.5B language model. Ask it to choose, then read the logits rather than generate — or press the other button and watch it spell the same answer out.
On the right, the FlyWire connectome: every neuron of a fruit fly brain and every connection between them, stepped as a spiking network right here. Shine a light on one side and watch which way it leans. Then cut the forty neurons that carry the difference.
| Generating | Reading the logits | The fly | |
|---|---|---|---|
| Forward passes | up to 288 | 1 | 1 volley |
| Malformed output | possible | structurally impossible | structurally impossible |
| Confidence | not directly available | calibrated, free | firing rates |
| Can you see why? | no | partly — the logits | the whole path, and you can cut it |