Jev AI Model From a ChatGPT Inventor Skips Language Entirely

TypeSafe AI has released Jev, a transformer-based AI model that does not output text. Instead, the Jev AI model returns probabilities, which the company calls calibrated decisions. The release this week drew immediate attention from software developers, and demand briefly overwhelmed the startup's API, according to TechCrunch.

The founder behind the model is Diogo Almeida, an OpenAI researcher who helped build ChatGPT and invented reinforcement learning from human feedback, the training technique that made modern chatbots possible. Almeida left OpenAI two years ago to start TypeSafe AI. He told TechCrunch that the company's goal was to fix a frustration he had carried since the ChatGPT era: models tuned for human language were of little use for automation, because computers speak a different language.

A model that never writes a word

Jev works differently from a large language model. A user defines the set of possible outputs in advance, and the model returns a probability for each one instead of generating sentences. Because the answers are fixed before the model runs, the Jev AI model cannot hallucinate in the way chatbots do. A classification task like sorting support tickets or approving a command simply comes back with a number attached to each option.

Skipping language makes the model remarkably cheap to run. TypeSafe charges nothing for output tokens and meters input tokens by the billion rather than the million. The company calls the Jev AI model a System One model, a reference to fast intuitive judgment as opposed to slow deliberate reasoning, and says it was trained exclusively on synthetic data using a technique Almeida calls reinforcement learning from calibrated decisions. Almeida told TechCrunch that betting the company on synthetic data was one of the best decisions he has made, ahead of the Jev launch and even ahead of inventing RLHF.

Early tests from working developers

Several engineers have already put Jev through real workloads. Pranit Sharma, a software engineer at Vercel, said his company had been using OpenAI's Luna 5.6 to run a classifier that reviews commands for safety. When Vercel swapped Luna for Jev, the classifier ran five to 18 times more quickly and with greater accuracy.

Nikhil Mudholkar, the chief technology officer of Bryo AI, tested Jev against Google's Gemini for classifying business emails. Gemini was slightly more accurate in his test, but 10 to 20 times more expensive. What interested Mudholkar more was the shape of the answers: Jev returns a real probability, which he described as ideal for automating workflows, because an automation can act on a 95 percent score and ignore a 50 percent one.

Armin Ronacher, the chief technology officer of Earendil, which builds the open source model harness Pi, sees the same appeal. He explained that the model hands the hallucination question back to the user: a low probability is a coin toss to disregard, while a high probability is something to act on. Ronacher also sees a role for Jev in model routing, deciding which workload needs which model, a job that would be too expensive to hand to a large language model but fits a cheap real-time decision model.

Built to sit beside the big chatbots

Jev is not only a replacement for language models in some jobs. Almeida argues it can also supervise them. Checking the behavior of AI agents usually means using one agent to watch another, which gets expensive fast. A cheap decision model could instead track an agent's actions and catch problems such as jailbreaks before they cause damage.

The model is named after William Stanley Jevons, the 19th-century economist known for the observation that bears his name: when the cost of a resource falls, people use more of it. Almeida's bet is that cheaper machine intelligence will spread through software the way the early internet did, distributed and emergent rather than concentrated in a few large applications. He told TechCrunch that he wants TypeSafe's product to be intelligence itself, in contrast to what he described as the fear and hype that other frontier labs produce.

For now, Jev stands alone as this kind of model, but observers expect company to follow. Ronacher told TechCrunch that rivals will likely appear now that the usefulness of decision-only models is clear, adding that the idea should have been obvious sooner if language models had not been so heavily subsidized. TypeSafe says it will build more versions of the model in new modalities. Early access is available through the company's API, which struggled briefly under its first wave of developer interest.