The researcher who helped build ChatGPT unveils an AI model that skips language

TypeSafe AI, a startup founded by former OpenAI researcher Almeida, this week released a new kind of AI model called Jev. Unlike a large language model, or LLM, the technology behind chatbots such as ChatGPT, Jev does not generate text. It is built on a transformer, the neural network design used in most modern AI, but it outputs probabilities, which the company calls calibrated decisions.

Almeida helped build ChatGPT at OpenAI and later helped invent reinforcement learning from human feedback, a training method built on human ratings, before leaving the company two years ago to start TypeSafe AI. He said he grew frustrated that language models were poorly suited to automation. “We have been super good at human language for four years, but it’s not useful for automation because computers speak a different language,” he told TechCrunch.

Because developers specify Jev’s possible answers in advance, the model cannot hallucinate, or state false information as fact, according to the company. That design also makes it cheap and fast. Input tokens, the chunks of text a model reads, are billed per billion rather than per million, and output tokens are free. Demand has been strong enough that TypeSafe AI briefly lost the ability to serve users through its API.

Early users report gains in speed and accuracy. Vercel software engineer Pranit Sharma said swapping OpenAI’s ChatGPT Luna 5.6 for Jev in a command-safety classifier made results five to 18 times faster, with greater accuracy, he said. Bryo AI chief technology officer Nikhil Mudholkar tested Jev against Google’s Gemini on classifying business emails; Gemini was slightly more accurate, he said, but 10 to 20 times more expensive. Mudholkar said Jev’s confidence scores, which show how certain the model is about an answer, suit it to automated workflows.

Watching other AI models

Jev can also monitor other AI systems. Armin Ronacher, chief technology officer of Earendil, which makes the open source AI tool Pi, said Jev could track the steps a language-model-based agent takes and flag attempts to bypass its safety rules, known as jailbreaks. Because Jev returns a probability rather than an answer, users can judge how much to trust it: a 50% score reads as a coin toss, Ronacher said, while a 95% score is actionable. Ronacher also pointed to model routing, deciding which AI model should handle a given task, as a use case where Jev’s low cost could make real-time sorting practical.

TypeSafe AI has not disclosed Jev’s architecture, though outside observers suspect it is built on an existing open-weight language model. Almeida said the model, named after 19th-century economist William Stanley Jevons, is trained entirely on synthetic data, or artificially generated training examples, using a method he calls reinforcement learning from calibrated decisions. Ronacher expects rival calibrated-decision models to appear now that Jev’s usefulness is public, and TypeSafe AI has said it plans to build versions of Jev for data types beyond text.