Probability Model Jev Replaces Language Models in Automation
A Researcher’s Disappointment
Diogo
Almeida helped build ChatGPT. [1] He then helped invent the training technique — reinforcement learning from human feedback — that made it work. And then he walked away, because he had concluded that the thing he had built was, in a specific and practical sense, useless.
His complaint was not that the technology failed. His complaint was that it succeeded at the wrong task. “We have lightning in a bottle, and yet it is not useful,” he said. Language models had become extraordinarily good at producing human language. But automation does not run on human language. It runs on decisions — yes or no, safe or unsafe, route to this model or that one. Four years of progress in generating prose had produced something that could not reliably tell a machine what to do next.
Almeida left OpenAI two years ago and founded TypeSafe AI to close that gap. In 2025, the company released Jev, a transformer-based model that is not a large language model. [2] It does not write. It does not chat. It returns probabilities — what the company calls calibrated decisions — and nothing else.
What Happens When You Remove the Words
The design choice sounds technical. Its consequences are not. By refusing to generate language, Jev sidesteps the failure mode that has defined the LLM era: hallucination. A model that outputs text can invent facts, sources, and confidence it does not possess. A model that outputs a number between zero and one cannot invent anything. It can only be right or wrong, and it tells you which it thinks it is.
That single property changes what a machine is allowed to decide. When a language model reviews a command for safety, a human still has to read the reasoning and judge whether the reasoning holds. When a probability model reviews the same command, the threshold does the judging. Set it at 95 % and the model acts. Set it at 50 % and the model abstains. The human who once weighed the evidence has been replaced by a number the human chose in advance.
Armin Ronacher, CTO of Earendil, described the trade plainly: the model “delegates the hallucination problem a little bit to the user.” [3] If the output comes back at 50 %, you treat it as a coin toss and discard it. If it comes back at 95 %, you act. The judgment has not disappeared. It has been moved — from the moment of decision to the moment of configuration, where it is made once and then applied a billion times without further thought.
The Economics of Not Thinking
Cost is not a side effect here. It is the mechanism. Jev’s output tokens are free; input tokens are metered by the billion rather than the million. Pranit Sharma, a software engineer at Vercel, had been using a large language model to classify commands for safety. [4] Swapping in Jev made the classification five to 18 times faster and with greater accuracy. [4]
Nikhil Mudholkar, CTO of Bryo AI, tested Jev against a leading language model for sorting business email. [5] The language model was slightly more accurate. It was also 10 to 20 times more expensive. [5] What decided the question for Mudholkar was not the accuracy gap but the confidence score — the fact that Jev “hands back a real probability,” which he called ideal for automating workflows.
Read that sequence again. A human task — reading an email and deciding where it belongs — is handed to a model. The model that is slightly worse at the task wins, because it is cheaper and because it admits uncertainty in a form a machine can act on. The human who used to sort the email is not replaced by the best available intelligence. They are replaced by the cheapest sufficient one.
The Check That Watches the Check
The most consequential use case is not replacement but supervision. AI agents are increasingly deployed to monitor other AI agents — watching for jailbreaks, flagging misbehavior, tracing what a model did and why. Running a language model to watch a language model is expensive, and the cost scales with every action the monitored agent takes.
Almeida argues that Jev makes this kind of oversight practical. A probability model can sit in the trace of an LLM agent and flag anomalies at a fraction of the cost. The supervisor no longer reasons about the supervised. It scores it.
This is where the human role thins most quietly. Oversight sounds like a human function. In practice, oversight of automated systems at scale has already been delegated to automated systems, because no organization can afford to read every trace. What Jev changes is not whether the delegation happens but how cheap it becomes — and cheap oversight is oversight that gets applied everywhere, including to decisions no one thought needed watching.
The Name Is the Argument
The model is named after William Stanley Jevons, the 19th-century economist who noticed that making coal engines more efficient increased total coal consumption rather than reducing it. [6] Efficiency lowers cost; lower cost expands use; expanded use overwhelms the savings. The Jevons paradox has been applied to energy, to highways, to computing. Almeida applies it to intelligence.
His bet is that cheaper decisions mean more decisions get automated — not fewer. “Smart software all over the place,” he calls it, “emergent and distributed, much more like the early internet” than the mega-apps of the current moment. The vision is not a single brilliant assistant. It is a million small judgments happening in code, each one too minor to have warranted a human even when a human was available.
That is the shape of the displacement. It does not arrive as a robot that does your job. It arrives as a threshold you set once, in a configuration file, for a task you never thought of as a task — and then it runs without you, at a cost low enough that no one bothers to ask whether you should have been consulted.
The Frontier Lab That Refuses the Label
Asked whether TypeSafe is a frontier lab, Almeida declined the frame. “The main product of frontier labs is fear or hype,” he said. “I would like our main product to be intelligence.” He added that his company is not “a lab in the sense of betting on infinite wealth, or a religion, or building God in a data center.”
The refusal is not modesty. It is a statement about where the value is. Frontier labs compete on capability — the largest model, the hardest benchmark, the most impressive demonstration. TypeSafe competes on a narrower claim: that the useful unit of AI is not a conversation but a decision, and that decisions can be made by something far smaller and far cheaper than the systems currently dominating the field.
Almeida is tight-lipped about Jev’s architecture; outside observers suspect it is built on top of an open-weight language model. The company calls it a “System One model” — fast, intuitive, focused on one task rather than general reasoning. It is trained, Almeida says, exclusively on synthetic data using a technique he calls reinforcement learning from calibrated decisions. He calls that bet — making all of the company’s own training data — one of the best of his life, better than RLHF. [2]
The Gap Between the Promise and the Practice
Ronacher expects competitors to appear now that Jev’s utility is visible. “We should have seen this earlier in many ways,” he said, “but presumably because the LLMs are so cheap and subsidized, you often don’t have to be creative yet.” (Source name)
That sentence contains the whole story. The reason no one built a decision model sooner is not that it was hard. It is that language models were cheap enough, and subsidized enough, that no one needed to. The industry spent four years optimizing the wrong output because the wrong output was affordable.
Now the affordable option has changed, and with it the boundary between what a person decides and what a threshold decides. The promise is calibrated confidence — a machine that knows what it does not know and says so. The practice is a number in a config file, set by someone who has already moved on, running a billion times a day, deciding things no one is watching.
Sources
1. OpenAI
2. TypeSafe AI
3. Earendil
4. Vercel
5. Bryo AI