AI
OpenAI's Diogo Almeida launches TypeSafe AI with Jev, a non-LLM model designed for software automation
A new transformer-based model from TypeSafe AI outputs probabilities instead of text, enabling faster, cheaper AI for code classification and real-time decision-making.
TechCrunch — AI
Diogo Almeida, a key researcher behind OpenAI's ChatGPT and the reinforcement learning from human feedback (RLHF) technique that shaped modern AI, has become frustrated with large language models. After two years building TypeSafe AI, he has introduced Jev, a fundamentally different kind of model that sidesteps the language-generation bottleneck entirely.
A Different Approach to AI Outputs#
Unlike conventional LLMs, Jev doesn't produce text. Instead, it generates probabilities and calibrated decisions—enabling users to understand confidence levels in the model's outputs. This architectural shift has immediate practical benefits: the model runs at a fraction of traditional AI inference costs, processes requests dramatically faster, and eliminates hallucination risk since outputs are predefined by the developer.
Jev meters input tokens at the per-billion scale rather than per-million, making it economically viable for high-volume applications that would be prohibitively expensive with conventional LLMs.
Early Adoption Signals Strong Demand#
Developers have embraced the product with such enthusiasm that TypeSafe briefly lost API availability from sheer volume. Real-world use cases are emerging quickly: engineers at Vercel, which develops agentic infrastructure, replaced OpenAI's Claude with Jev for safety classification and saw results arrive 5 to 18 times faster with better accuracy. Nikhil Mudholkar, CTO of Bryo AI, tested Jev against Gemini for email classification and found it 10 to 20 times cheaper while providing genuine probability scores that make workflow automation safer and more transparent.
Complementing and Monitoring LLMs#
Beyond direct replacement, Jev serves as a cost-effective guardian for LLM agent systems. Deploying an LLM to monitor another LLM's behavior quickly becomes expensive; Jev's low cost makes such oversight practical. Potential applications include tracking LLM execution traces and preventing prompt injection attacks. Armin Ronacher, CTO of Earendil, sees another use case: model routing—predicting whether a workload requires a specialized model—becomes economically feasible when the routing decision itself is cheap and near-instantaneous.
