A new kind of AI model from a ChatGPT inventor is thrilling developers
techcrunch ·
Diogo Almeida, an OpenAI researcher involved in building ChatGPT and inventing reinforcement learning from human feedback, founded TypeSafe AI after concluding that optimizing for human language limits automation. The company has released Jev, a transformer-based model that is not an LLM and produces probabilities, or “calibrated decisions,” instead of text. Because users define its possible outputs in advance, Jev is designed to be fast, inexpensive, and unable to hallucinate. Developers have shown strong interest, with reported uses including software safety classification, workflow automation, monitoring AI agents, preventing jailbreaks, and routing tasks to appropriate models. Vercel reported that Jev was five to 18 times faster and more accurate than an OpenAI model for one classifier, while another developer found Gemini slightly more accurate but 10 to 20 times more expensive. Almeida says Jev is trained exclusively on synthetic data using a method called reinforcement learning from calibrated decisions, although TypeSafe has disclosed few architectural details. The company plans to develop additional versions for new modalities, and observers expect competitors to follow as the model’s utility becomes clearer.
techcrunch