Author(s): Enzo Lombardi

Originally published on Towards AI.

The classifier that never writes a word

TypeSafe’s Jev announcement did the rounds properly: a frontier model that does not generate text, returns typed values with calibrated probabilities, and answers in 70 to 500 milliseconds instead of the 3 to 329 seconds a generative model spends thinking out loud. The pitch is a good one. Unstructured state goes in, a typed decision comes out, and there is no JSON to repair on the way back.

The article explains how “System One” models work as typed classifiers rather than decoders, including how options are represented with marker tokens and scored to produce a probability vector (or nothing). It then walks through building a Rust crate and demos: first a security-style classifier for detecting credential exfiltration while using calibrated probabilities and highlighting key failure modes (especially adversarial prompting, plus how checkpoint selection functions as a threshold/scale calibration). Next, it demonstrates real-time decision-making with a Pong game loop at 30 FPS, arguing that phrasing state and options in natural language (not JSON or bare labels) is crucial for accuracy and latency, and showing that the model’s runtime can fit inside frame budgets. Finally, it distills practical guidance: translate state/policies into the language the model was trained on, ask questions about the world or intent rather than your own sandbox rules, avoid treating the model like a schema-driven API, and fine-tune the head/thresholding to push performance without increasing decision cost.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI