Diogo Almeida helped build ChatGPT and invented reinforcement learning from human feedback (RLHF) — the training technique most responsible for the current AI era. But he grew frustrated with what he saw as a fundamental mismatch: language models are optimized for human language, yet computers speak a different language entirely.
Two years after leaving OpenAI to found TypeSafe AI, Almeida this week released Jev, a transformer-based model that is explicitly not a large language model. It does not output text. Instead, it produces calibrated probability scores — what the company calls "calibrated decisions." Users define the output categories in advance, which means the model cannot hallucinate by definition.
The results are drawing developer attention. Vercel engineer Pranit Sharma reported that replacing OpenAI's ChatGPT Luna 5.6 with Jev for a command-safety classifier yielded results 5 to 18 times faster with greater accuracy. Bryo AI CTO Nikhil Mudholkar found Jev 10 to 20 times cheaper than Gemini for email classification, while praising its real probability scores as "ideal for automating workflows."
By eschewing language output entirely, Jev achieves radical cost reduction: output tokens are free, and input tokens are metered by the billion rather than the million. The demand was high enough during launch that TypeSafe briefly lost the ability to serve API users.
The model is named after William Stanley Jevons, the 19th-century economist whose paradox describes how falling commodity costs lead to increased usage. Almeida's thesis is that making intelligence radically cheaper will lead to its pervasive deployment across software systems — "much more like the early internet than the mega apps people are trying to build right now."
Jev is trained exclusively on synthetic data using a technique Almeida calls "reinforcement learning from calibrated decisions." He described the synthetic data approach as "one of the best bets I've ever made in my life — better than our launch, better than RLHF."
Beyond replacing LLMs in classification and routing tasks, Jev can also augment them — monitoring agent traces, preventing jailbreaks, and serving as a model router that predicts whether a workload needs a specific model. Armin Ronacher, CTO of Earendil, noted that while Jev delegates some hallucination risk to the user (who must interpret confidence scores), its low cost makes real-time model sorting practical for the first time.
