Large language models dominate the headlines, but a parallel category is maturing fast. Models that do not write prose at all. They classify, score, and route—and they're worth a serious look.
Hi, my name is Tom Smykowski. I've spent 19 years building production systems and now focus on AI-assisted development patterns. On this blog I explore how engineers can leverage new AI capabilities without losing sight of fundamentals.
TypeSafe AI calls them System One models, after Daniel Kahneman's fast, intuitive "System 1" thinking. In machine-learning terms they are discriminative: they map an input state to a distribution over a fixed set of options you define at request time, not open-ended generation.
If you build agents, support automation, or any pipeline with dozens of "if this, then that" branches, these models deserve attention. This article covers the landscape as of late 2026, centered on Jev and its siblings—Microsoft-Decision-1, Cloudflare Clef, and open-source Laya.
The key insight: stop asking a chat model to "reply with only the label" and start calling a model that only ever returns distributions over your options.