TypeSafe AI's new model never writes a sentence -- it answers in typed, calibrated probabilities, and claims to be up to 200x faster than an LLM doing the same job

TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, launched a foundation model in early access on Sept. 15 that breaks from the entire generate-text-then-parse-it pattern behind every mainstream chatbot. Jev -- the first of what TypeSafe calls "System One Models" -- never produces a sentence. Given a program's state and a set of typed questions, it answers all of them in a single parallel pass, returning structured values paired with calibrated probability scores instead of prose a caller then has to interpret. TypeSafe's own numbers: 70-500 millisecond end-to-end response time, and -- on the company's internal workflow-evaluation benchmark -- 193.6 times faster and 444.6 times cheaper than a comparable frontier LLM call. Almeida co-founded TypeSafe with Erik Gafni and Sasha Sheng; the company has raised $40 million and runs an in-house synthetic-data lab that generated most of Jev's training data.

The mechanism is a genuine architectural break, not a prompting trick. Where a standard LLM predicts one token at a time, conditioning each on everything before it, Jev's parallel sampler generates every requested output in the same pass, trained with a method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). Because outputs are constrained to a predefined schema -- a 255-option cardinality cap per single-stage decision -- the company says it's "mathematically impossible" for Jev to produce a type error: it cannot answer outside the set of options it was given, the way a text model can drift into an invented category or malformed field. TypeSafe has disclosed little about the model's size or training-data composition beyond confirming it was trained primarily on synthetic data generated in-house. TypeSafe frames the guarantee narrowly and explicitly in its own documentation: giving up free-form string generation trades away general flexibility, and the company's launch post itself warns that the demo workflows it publishes "represent higher-end real-world gains" rather than a typical result.

Almeida's frustration with the alternative is what he says motivated the company: despite two years of rapid LLM progress since he left OpenAI, he argues AI "is not useful" for a large share of real automation because, as he put it, "computers speak a different language" than the free text a chatbot returns -- every application built on an LLM classifier still has to parse prose back into a typed value it can act on. The model's name is a deliberate bet on that framing: Jev is named for economist William Stanley Jevons, whose paradox holds that a falling cost of a commodity leads to it being used more, not less. TypeSafe is wagering that cheap, instant, typed decisions get embedded everywhere -- real-time game and robotics loops, high-volume classification, checking what an LLM agent is about to do before it does it -- rather than simply replacing a slice of existing LLM API spend.

That last row is the real gap in the launch-week coverage. KDnuggets' own analysis of Jev pushed back directly on the framing that's traveled furthest: "classification is not new," it noted -- intent detection and zero-shot classification are established techniques, and what Jev packages is a faster, cheaper way to do a job models and traditional classifiers already do, not a new category of intelligence. It also flagged the specific phrase TypeSafe uses -- "cannot hallucinate" -- as true only in a narrow sense: Jev is guaranteed to stay inside its predefined schema, which is not the same as being guaranteed correct. "Zero hallucinations," the piece argued, really means zero out-of-schema outputs, not zero incorrect decisions -- and the one public number that speaks to correctness, that 68% accuracy figure, comes from TypeSafe's own internal evaluation against reference answers that frontier models generated, not verified ground truth.

Earendil CTO Armin Ronacher, an early tester quoted by TechCrunch, put the practical caution simply: a confidence score near 50% is close to a coin toss and application logic built on Jev needs to treat it that way rather than trusting whatever value comes back highest. Ronacher also expects the idea to spread regardless of how Jev specifically performs -- competitors building their own non-autoregressive decision models, once the category has a name and an early-access waitlist proving developer interest. For now, Jev remains in early access, with TypeSafe bringing users "off the waitlist as quickly as we can" rather than offering open signup.