A new kind of AI model from a ChatGPT inventor is thrilling developers

Tim Fernholz’s TechCrunch profile covers Diogo Almeida, an OpenAI researcher who helped build ChatGPT and invent RLHF, then left two years ago to found TypeSafe AI. This week TypeSafe released Jev, a transformer that is not an LLM: it outputs probabilities—“calibrated decisions”—not text, so users define outputs in advance and the model cannot hallucinate in the usual sense.

Almeida’s thesis: optimizing for human language is the wrong target for automation because computers speak a different language. Jev is positioned as cheap and fast—output tokens free, input metered by the billion not the million. Demand briefly overwhelmed the API. Vercel engineer Pranit Sharma said swapping ChatGPT Luna 5.6 for Jev on a safety classifier for commands yielded results five to 18 times faster with greater accuracy. Bryo AI’s Nikhil Mudholkar found Gemini slightly more accurate on business-email classification but 10–20× more expensive, and valued Jev’s real probability scores.

Earendil CTO Armin Ronacher notes the model shifts the hallucination problem to the user via thresholds, and sees routing and agent-trace monitoring as fits. Almeida says Jev is trained only on synthetic data via “reinforcement learning from calibrated decisions,” calls it a “System One” model, and stays tight-lipped on architecture that outsiders suspect sits atop an open-weight LLM.

We’re cheering models built for automation instead of chat theater—calibrated probabilities are how you wire intelligence into software without praying the next token behaves.

Almeida’s heartbreak line—“lightning in a bottle, and yet it is not useful”—names the gap between demos and production classifiers. If Jev’s early Vercel and Bryo numbers hold, a whole class of LLM babysitting jobs gets cheaper overnight. That is useful AI: smaller, faster, honest about uncertainty.

The risks are familiar. Architecture secrecy invites cargo-cult copies and trust gaps. “Cannot hallucinate” is true only inside the user-defined label set; bad schemas still automate bad decisions. Ronacher is right that 50% means coin toss—teams that ignore scores will recreate LLM failure modes at higher QPS.

I’m watching whether competitors ship similar System One models, whether TypeSafe opens enough method to audit calibration, and whether agent stacks standardize on cheap probability checks before calling expensive reasoners.

TechCrunch published on September 18, 2026. Almeida named the model after economist William Stanley Jevons; he rejected the “frontier lab” brand, saying he wants the product to be intelligence rather than fear or hype.