Ex-OpenAI researcher's new AI model skips text for probabilities
Startup TypeSafe AI, founded by a ChatGPT co-creator, has launched Jev, a model that outputs probabilities instead of text and is drawing strong developer interest for its speed and low cost. Demand has been so high the company briefly couldn't serve all API users.
Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invent reinforcement learning from human feedback (RLHF), left the company two years ago to found TypeSafe AI. He had grown convinced that large language models, despite their power, weren't well suited for automation because computers process information differently than humans communicate.
This week TypeSafe AI released Jev, a transformer-based model that departs from the standard large language model approach. Rather than generating text, Jev produces probabilities — what the company calls "calibrated decisions." Because outputs are predefined by users, the model cannot hallucinate, and it runs far cheaper and faster than conventional language models.
Strong developer response
Demand for Jev has been intense enough that TypeSafe AI temporarily lost the ability to serve all users through its API. A Vercel engineer said that after swapping an OpenAI model for Jev in a command-safety classifier, results came back five to 18 times faster and more accurately. Another developer who tested Jev against Google's Gemini for email classification found Gemini slightly more accurate but far more expensive, while praising Jev's genuine probability scores as useful for automating workflows.
Beyond replacing language models in some tasks, Jev can also supplement them — monitoring other AI agents for misbehavior or jailbreak attempts, and helping route tasks to the appropriate model based on real-time cost and speed considerations.
The model is named after economist William Stanley Jevons, whose paradox holds that falling costs of a resource can drive far greater overall consumption of it. Almeida hopes cheaper intelligence will lead to it being embedded widely and unpredictably across software, more like the early internet than today's dominant apps.
Almeida has kept details of Jev's underlying architecture private, saying only that it was trained exclusively on synthetic data using a proprietary technique. Industry observers expect competitors to emerge now that the approach's usefulness has been demonstrated.
