DEV Community

pablo padlo
pablo padlo

Posted on Originally published at x.com

The AI That Refuses to Write: What Jev Changes When It Returns Decisions

#ai

The AI That Refuses to Write: What Jev Changes When It Returns Decisions

An AI model that cannot write a sentence. Jev, released by San Francisco's TypeSafe AI on 15 September 2026, does not generate text at all. You hand it unstructured state — a support ticket, a log line, a page of messy input — plus a set of typed questions, and it returns typed answers with calibrated probabilities. No string, no parse step, no option you did not declare.

The trade is the point. TypeSafe calls the class "System One models", borrowing Daniel Kahneman's split between fast intuition and slow reasoning, and argues that RLHF — the method behind chat LLMs — optimised for human preference rather than for decisions software can trust without a human in the loop.

A model with no output string

Everything about the interface is typed. The possible answers are declared in advance, so the model cannot return a malformed value or invent an option outside the set. All answers come back in a single parallel pass rather than one token at a time, which is where the speed comes from. TypeSafe describes the result as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out".

The system has no free-form output surface at all. It cannot write a paragraph, a code block or a refusal — a limitation the company frames as the source of its advantages.

What the numbers say — and who measured them

The benchmarks are TypeSafe's own, and the coverage repeats them as such. End-to-end latency is listed at 70–500 ms, against a 3–329 s range for frontier LLMs on the same System One shaped tasks. Input pricing is $0.042 per million tokens; output is free, described as too cheap to meter. Training uses RLCD — Reinforcement Learning for Calibrated Decisions — on a new architecture the company has not disclosed.

Independent write-ups, from Tom's Hardware to DataCamp, carry the same caveat: these are vendor figures, not third-party measurements.

Why calibration is the real claim

"Can't hallucinate" is the headline, and it deserves a careful read. It is a schema guarantee: Jev cannot produce a malformed value or an option outside the declared set. It can still pick the wrong valid option, and TypeSafe's own FAQ acknowledges as much. What is meant to surface those cases is calibration — the property that a model saying 90% is right about 90% of the time. A model that says 90% and is right 60% of the time cannot be automated around, no matter how fast it is. Calibration is measurable, and it is the claim worth auditing.

The other half of the argument is scope. TypeSafe does not pitch Jev as a replacement for general LLMs; it pitches it at the places where software currently shells out to one and then has to parse and validate the result: classify, route, score, extract, branch, guardrail.

What to watch

The demo that spread is the least serious use case. On 17 September, @_MaxBlade posted Jev playing Subway Surfers at superhuman speed — 50 games at once, under a cent for the run, every lane choice labelled with its probability. It is a clean visualisation of the primitive: a decision function, called thousands of times a second, returning numbers you can act on immediately.

The serious version of that capability is boring on camera — routing a ticket, scoring a lead, guarding another model's output. The next number that matters is not the speed multiple. It is how often Jev's confidence is right when the decision actually costs something.

Watch the demo (video):

https://x.com/pgol80/status/2102759773363978682

Sources:

Top comments (0)