Diogo Almeida helped invent reinforcement learning from human feedback at OpenAI and was part of the team that built ChatGPT. But watching the chatbot take off left him disillusioned. Despite its conversational fluency, the technology wasn't built for the kind of software automation developers actually needed. Two years after leaving OpenAI, he's released a model that sidesteps language entirely.
Why Almeida Left OpenAI and What He Built
Almeida spent years at OpenAI perfecting human language processing. RLHF, the technique he helped develop, became the backbone of modern chatbots by training models to produce outputs that humans rate highly. But Almeida came to see this as a dead end for automation. Computers, he realized, don't speak human language. Optimizing for fluent text output was solving the wrong problem.
He founded TypeSafe AI to address this gap. This week, the company released Jev, a transformer-based model that produces calibrated probabilities instead of text. The model doesn't generate language, so it cannot hallucinate. Users define the output categories in advance, and Jev returns confidence scores for each. Output tokens are free. Input tokens are metered by the billion, not the million.
Jev is named after William Stanley Jevons, the 19th-century economist whose paradox describes how reducing the cost of a resource tends to increase its consumption. Almeida is betting that making intelligence cheap and deterministic will lead to its deployment everywhere, in ways that don't resemble the monolithic AI applications dominating headlines.
How Jev Works and Why It's Different
The model is classified as a "System One model," a reference to Daniel Kahneman's framework of fast, intuitive cognition. Where reasoning models like OpenAI's o1 or StepFun's Step 5 Preview work through problems step by step, Jev makes quick, probabilistic judgments about predefined categories. It does not generate text, explain its reasoning, or produce open-ended answers.
TypeSafe trains Jev exclusively on synthetic data using a technique Almeida calls "reinforcement learning from calibrated decisions." The approach produces a model that returns mathematically meaningful probabilities rather than the vague confidence-like outputs typical of language models. This distinction matters for automation. A 95% confidence score from Jev carries a different statistical weight than a language model saying it's "pretty sure."
The architecture details remain closely guarded. Outside observers suspect Jev is built on top of an open-weight large language model, though Almeida declined to confirm this. What he did emphasize is that the company's early bet on synthetic data generation paid off. Half of TypeSafe operates as a lab focused on statistically well-understood synthetic training data, which Almeida considers the company's most important innovation.
Developers Are Already Swapping In Jev
Demand spiked immediately after launch. TypeSafe briefly lost the ability to serve API requests because too many developers tried to use the model at once.
Pranit Sharma, a software engineer at Vercel, reported replacing OpenAI's Luna 5.6 classifier with Jev for reviewing commands for safety. The switch produced results five to 18 times faster with higher accuracy. Vercel builds agentic infrastructure, so the speed and reliability gains matter for their core product.
Nikhil Mudholkar, CTO of Bryo AI, tested Jev against Gemini for classifying business emails. Gemini held a slight edge in accuracy, but cost 10 to 20 times more. Mudholkar highlighted Jev's confidence scores as the key differentiator. Unlike language models that produce hedged text, Jev returns actual probability distributions, which makes it suitable for automated workflow decisions where human review of every output isn't feasible.
Armin Ronacher, CTO of Earendil and creator of the Pi model harness, noted that Jev shifts responsibility for interpreting confidence to the user. A 50% score signals a coin toss. A 95% score indicates actionable certainty. This transparency makes it practical for production systems that need deterministic behavior.
Using Jev to Keep Language Models in Line
One of the more compelling applications is using Jev as a monitor for language model agents. Running one LLM to watch another for jailbreaks or misbehavior gets expensive quickly. Jev's low cost and fast inference make it viable as a lightweight sentinel that flags suspicious outputs without burning through tokens at LLM rates.
Ronacher also pointed to model routing as a natural fit. Determining whether a given prompt requires a powerful reasoning model or a cheaper one is itself a classification problem. Jev could make that decision in real time, sending complex queries to expensive models and routing routine requests elsewhere.
What This Means for the AI Industry
Language models have dominated AI development for four years. Jev represents a parallel track, one where stripping away language generation produces something faster, cheaper, and more predictable. For developers building automation systems, this isn't a replacement for LLMs but a complement. The model fits into pipelines where classification, routing, and monitoring need to happen at scale without the cost and unpredictability of text generation.
Ronacher expects competitors to follow now that the utility is proven. The economics are straightforward. When intelligence becomes cheap enough to embed in every decision point, software starts behaving differently. Almeida's vision is not another mega-app but distributed intelligence woven throughout systems, more like the early internet than the centralized platforms that followed.
TypeSafe plans to expand Jev into new modalities. Almeida dismissed the framing of frontier labs chasing artificial general intelligence, saying the main product of such labs is fear or hype. His company's product, he said, is intelligence that works.