Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invented reinforcement learning from human feedback (RLHF), said he became disillusioned with language-centered models. "We have lightning in a bottle, and yet it is not useful," Almeida told TechCrunch, arguing that focusing optimization on human language has limited usefulness for automation because "computers speak a different language."
Two years after leaving OpenAI, Almeida founded TypeSafe AI to address that gap. This week the company released Jev, a transformer-based model that the company says is not a large language model (LLM). Instead of producing text, Jev outputs probabilities — what TypeSafe calls "calibrated decisions."
Why Jev avoids natural language
By abandoning natural language outputs, Jev gains several practical advantages: the model is substantially cheaper and faster to run; outputs cannot hallucinate because users define the permissible outputs in advance; output tokens are effectively free; and input token accounting happens at billion-scale rather than million-scale.
Demand for Jev was strong enough that TypeSafe briefly lost the ability to serve API users. The model is being positioned primarily for software automation, where developers see it as a lower-cost and more robust way to add intelligence into applications.
Early user results
Developers testing Jev report significant performance and cost benefits. Pranit Sharma, a software engineer at Vercel, said his company had used OpenAI's ChatGPT Luna 5.6 to run a safety classifier for commands. Replacing Luna with Jev produced results 5 to 18 times faster and with greater accuracy. Nikhil Mudholkar, CTO of Bryo AI, compared Jev and Google Gemini on classifying business emails: Gemini was slightly more accurate but 10 to 20 times more expensive. Mudholkar highlighted Jev's confidence scores: "it is the only one that hands back a real probability which makes it ideal for automating workflows!!"
Complementing and supervising LLMs
Beyond replacing LLMs for some tasks, Jev can augment them by acting as a low-cost monitor or gatekeeper. Almeida suggests using Jev to track LLM agent traces and prevent jailbreaks — an approach cheaper than deploying agents to monitor agents. Armin Ronacher, CTO of Earendil (developer of the open-source model harness Pi), said Jev "delegates the hallucination problem a little bit to the user": users decide how to act based on returned probabilities (for example, treating a 50% result as ambiguous but trusting a 95% score).
Ronacher also pointed to model routing as a promising use case: predicting whether a workload requires a more expensive model. Using an LLM to route workloads would be costly; Jev's low latency and price make real-time routing feasible.
Architecture, training and name
TypeSafe is restrained about Jev's exact architecture; outside observers suspect it may be built atop an open-weight LLM. The company describes Jev as a "System One model" that prioritizes intuition over multi-step reasoning and is tailored to the right task. According to Almeida, Jev is trained exclusively on synthetic data via a method TypeSafe calls "reinforcement learning from calibrated decisions."
Almeida said an early bet to generate all training data in-house was one of the best decisions the company made, even better than their launch or RLHF in his view. He described roughly half the company as a lab focused on statistically well-understood synthetic data generation.
The model's name honors William Stanley Jevons, the 19th-century economist known for the Jevons paradox: as the cost of a commodity falls, its use can increase. Almeida hopes that the falling cost of computing intelligence will similarly lead to widespread, distributed deployment: "We think that there's just going to be smart software all over the place in a way that's emergent and distributed… much more like the early internet than… the mega apps that people are trying to build right now."
Outlook and competition
Jev is, for now, unusual in its approach, but Ronacher expects competitors to appear as Jev's practical utility becomes clearer. He suggested that widespread reliance on subsidized LLMs may have delayed creative alternatives.
TypeSafe plans to develop additional versions of Jev in new modalities. Asked whether TypeSafe is a "frontier lab," Almeida argued that frontier labs often sell fear or hype, whereas he wants his company's chief product to be usable intelligence rather than speculative promises or ideological projects.
