Diogo Almeida helped invent reinforcement learning from human feedback (RLHF), the technique that powers modern chatbots. Yet he left OpenAI frustrated, telling TechCrunch that optimizing for human language made AI useless for actual automation. “We have lightning in a bottle, and yet it is not useful,” he said. His new startup, TypeSafe AI, just launched Jev, a transformer-based model that rejects text output entirely. Instead of generating prose, Jev produces calibrated probabilities. This shift eliminates hallucinations and slashes costs, metering inputs by the billion rather than the million.
Developers are rushing to adopt the technology. Vercel replaced OpenAI’s Luna classifier with Jev for command safety reviews, seeing speeds jump five to 18 times faster with better accuracy. Nikhil Mudholkar, CTO of Bryo AI, found Jev ten to twenty times cheaper than Gemini for email classification. More importantly, Jev returns real confidence scores. Armin Ronacher, CTO of Earendil, notes this lets users decide when to act. If the model returns a 50% probability, you ignore it. At 95%, you automate the task. It delegates the uncertainty problem back to the engineer, where it belongs.
Almeida calls Jev a “System One model,” focusing on intuition over reasoning. It runs exclusively on synthetic data trained via “reinforcement learning from calibrated decisions.” He considers this data strategy a bigger win than his work on RLHF. The name references William Stanley Jevons, the economist who observed that falling commodity costs drive higher consumption. Almeida expects intelligence to become distributed and emergent, much like the early internet, rather than confined to mega-apps.
Critics suspect Jev sits atop an open-weight LLM, but its performance suggests a distinct architectural advantage. Ronacher argues we missed this shift because subsidized LLM prices removed the incentive for creativity. As TypeSafe plans new modalities, Almeida remains skeptical of frontier labs peddling fear. He wants his main product to be intelligence, not hype. If Jev delivers on its promise of cheap, reliable decision-making, it could redefine how software incorporates logic, moving us past the era of chatty bots toward silent, efficient agents.
