System One models · anatomy
Every model here answers a typed question without writing a sentence. They do it in four different places: in the vocabulary, in a fixed set of classes, or in a learned vector space. Where the answer lives decides what the thing can do, what it costs, and what you have to retrain when the question changes.
The same decision — which action does this ticket need? — sent to all four at once, in slow motion. Watch where each one spends its time. Three of them are finished before the first has stopped emitting characters.
A model, not a measurement. Durations are scaled from each project's self-reported figures (Laya 32.8 ms, CLM ~4–9× faster than Jev's 236–276 ms, AnyJev one prefill plus order-debiasing passes) and an autoregressive baseline emitting 18 tokens at 25 ms each. None of it is a controlled comparison.
A new answer appears that nobody trained for — “escalate to legal”. Now what?
You need to score 500 candidate actions against one state. What does that cost?
Not a diagram. A real two-tower encoder, training in this tab on ten support tickets with four held back. Text is hashed into 192 features, two heads project state and action into the plane, and a bidirectional InfoNCE loss pulls each ticket onto its own action and off every other. The heads project straight into 2D, so the picture below is the embedding space, not a flattening of one. Press train, watch it sort itself out — then watch the held-out number refuse to follow.
Scored by the model above, on whatever you type. Before training it is noise.
Everything here is computed in the page: hashed trigram features, two 192×2 projections, full-batch gradient descent on the exact InfoNCE gradient. No pretrained weights, no network calls. A 192-dimensional hashing featurizer is a crude stand-in for a real language backbone, so treat the geometry as honest and the language understanding as a toy.
Every model here emits a number between 0 and 1. Whether it means anything is a separate question. Bucket the predictions by claimed confidence, plot claimed against observed, and the gap from the diagonal is the lie. Raw softmax pulls away from it at both ends, most damagingly at the top where the volume is. One parameter fixes most of it.
Observed accuracies are illustrative, not measured. Temperature scaling fits one scalar after the model is frozen: divide the logit by T, refit nothing else. Being monotone it cannot reorder anything, so accuracy is identical on every example — only whether 0.9 means 90% changes. This is not fine-tuning and is no substitute for it. Fine-tuning moves the weights and fixes wrong answers; temperature fixes dishonest numbers attached to answers that were already right. Fine-tuning also tends to make calibration worse, so the order is fine-tune first, then fit T on labels the fine-tuning never saw.
| Autoregressive | AnyJev | Laya | CLM |
|---|