September 22, 2026
Betting against chat: TypeSafe builds Jev model for decisions inside software
The company declined to submit to public benchmarks, so there will be no way to compare Jev with other models.
Latent.Space
@latentspacepod
Jev and System One Model: RLCD, intelligence per dollar, reliable AI, and the end of chat-centric AI https://www.latent.space/p/jev The CEO of @typesafeai @CompleteSkeptic explains why AI can solve extraordinarily complex problems yet fails to automate simple work; why Jev is built for reliable decisions inside software rather than chat; why TypeSafe rejects public benchmarks and API-level refusals; why data and the right task matter more than brute compute; how System One Models could reshape coding agents and software as a whole; and why, even with a billion dollars, he would not train a model from scratch.
Jev pod tomorrow subscribe on apple /youtube @latentspacepod thanks to @allenpark and Ke for making this happen!
· 355.1K views
TypeSafe’s head would not spend a billion dollars building its own model from scratch.
His case is simple: AI already takes on tasks that strong engineers struggle with, yet still does not automate ordinary day-to-day work. The gap is closed by data and the right task choice, not a mountain of compute.
Decisions over dialogue. The company classifies its Jev model as a System One Model and builds it for reliable decisions inside a program. Chat is not the interface here: the model responds where software selects an option.
Previously, a new model was shown in a chat window: a person writes a prompt, a person reads the answer. TypeSafe moves the model under the hood, and the software itself makes the decision.
Ratings miss the point. TypeSafe declines to submit to public benchmarks and does not build API refusals in. The company promises the same shift for coding agents: System One Models, it says, will reshape both them and the software around them.
In this picture, chat remains the storefront, while models inside software do the work.
