A former OpenAI researcher who helped invent the training technique behind ChatGPT has released a new kind of AI model that doesn't produce text at all. Diogo Almeida's startup TypeSafe AI this week launched Jev, a transformer-based system that outputs probabilities instead of language, according to a report published Thursday by TechCrunch. Almeida, who co-invented reinforcement learning from human feedback (RLHF) at OpenAI, left the company two years ago because he believed optimizing for human language made AI unsuitable for automation, since computers operate in a different language.
The new model's design enables dramatically lower costs and faster performance than traditional large language models. Output tokens are provided at no charge, while input tokens are billed by the billion rather than the million, the report states. Because users specify outputs in advance, the system can't hallucinate. Developer response overwhelmed the company's infrastructure — demand briefly knocked out API service, according to the report. Software engineer Pranit Sharma at Vercel reported that swapping OpenAI's ChatGPT Luna 5.6 for Jev in a safety classifier delivered results five to 18 times faster with improved accuracy. In a separate test, Bryo AI's CTO Nikhil Mudholkar found Gemini slightly more accurate than Jev for email classification but 10 to 20 times costlier.
Almeida describes the model as a "System One model" focused on intuition rather than reasoning, trained exclusively on synthetic data using a technique he calls "reinforcement learning from calibrated decisions," the report notes. The company produces all its own training data internally. Almeida told TechCrunch that decision represented "one of the best bets I've ever made in my life — better than our launch, in my opinion, better than RLHF." What the model delivers that sets it apart are confidence scores, which Mudholkar called "a real probability" that makes workflow automation practical. Armin Ronacher, CTO of Earendil, explained that users can set thresholds — disregarding 50% probability responses as coin tosses while acting on 95% confidence outputs.
The report identifies several use cases where Jev's economics shift what's feasible. Monitoring AI agent behavior to prevent jailbreaks becomes affordable when Jev replaces more expensive large language models in oversight roles, according to Almeida. Model routing — predicting which specific model a workload requires — becomes viable at Jev's price point and speed, Ronacher told TechCrunch, though using an LLM for that task would be prohibitively expensive. The model takes its name from William Stanley Jevons, the 19th-century economist whose paradox describes how falling commodity costs drive increased usage. Almeida envisions intelligence becoming cheap enough to embed everywhere, creating "smart software all over the place in a way that's emergent and distributed," rather than centralized mega-applications. Ronacher expects competitors will emerge now that this model category has proven its utility, noting that subsidized LLM pricing has reduced pressure to innovate. For businesses weighing whether to automate decisions with AI, the trade-off has shifted: probability-based models sacrifice the flexibility of language but gain speed, cost efficiency, and reliability that general-purpose chatbots can't match.