← Accueil actu | Salle du marché
2026-09-16 · Source：The Decoder
Former OpenAI researcher Diogo Almeida's TypeSafe AI releases Jev, which generates no text and returns only preset classifications, starting at 70 milliseconds with extremely low token prices.
TypeSafe AI, a startup founded by former OpenAI researcher Diogo Almeida, is releasing an artificial-intelligence model designed to generate no text at all. The model, called "Jev," skips chat-style responses entirely and instead returns pure classifications for software applications. Response times start at 70 milliseconds, according to the company, and token prices are extremely low.
The approach inverts the standard recipe behind mainstream language models, which are built to produce open-ended text. Jev operates inside a closed output space: given a set of preset options, it judges which one applies and returns that judgment. For developers, that means the model cannot wander into off-script prose or unstructured answers, because free-form generation has been removed from the system's job description altogether.
TypeSafe AI is explicit about what that guarantee does and does not cover. The design offers no protection against mistakes. Jev may still select the wrong option; what the system ensures is only that it will stick strictly to the choices it has been preset with. The promise, in other words, is one of output discipline rather than of correctness.
The positioning points toward workloads where latency and unit economics dominate. With response times beginning at 70 milliseconds and extremely low token pricing, Jev targets classification tasks embedded within software, where the model serves the application rather than a human reader. For compute buyers weighing inference options, the launch adds a distinct category: a model that trades generative fluency for a bounded, predictable format. Almeida's background as a former OpenAI researcher supplies the young company's technical credentials, while responsibility for verifying Jev's accuracy stays with its users.
Mots-clés: TypeSafe AI · Jev · Diogo Almeida · AI Classification · Inference Latency
萬安算交所 AIXX · AIXX · Bourse compute vérifiée · Accéder à la salle