By Zhu Xueying
Jev has suddenly become a talking point in developer circles. What stands out is not what it can do, but what it does not try to do. It is not built around chat, it does not write code, and it does not generate long answers in the style of ChatGPT. Its job is narrower: make a decision.
That stripped-down design is exactly what pushed it into the spotlight, according to PANews. The report says developers have already put Jev to work in several practical settings. One user used it to analyze 724 live ads in 40 seconds, producing 8,724 decisions. Others connected it to Claude Code to clean up unnecessary context. Some used it as a referee for AI agents, checking whether a task had actually been completed. LangChain has also started testing Jev as an evaluator for agents.
Speed and price are a major part of the pitch. PANews cited TypeSafe test results showing Jev was up to about 193.6x faster and reduced costs by as much as 444.6x. Input is priced at $0.042 per million tokens, while output tokens are free.
The report argues that Jev’s rise says something about what AI systems may need in the agent era. The key demand may not be deeper and deeper reasoning every time. It may be huge volumes of fast, cheap judgments: what to do next, which tool to call, whether a task is complete. Those decisions may need to happen in the background, without a person constantly watching a screen.
A creator tied to RLHF is now questioning RLHF
PANews frames Jev partly through the background of its creator, Diogo Almeida. Almeida previously worked at OpenAI and was involved in GPT-4, ChatGPT, and InstructGPT/RLHF-related work. In other words, he helped build one of the most important post-training approaches behind today’s large language models.
But in a talk delivered a little over a month ago, Almeida raised a direct question: “What’s Next After RLHF?” PANews says that speech now reads almost like a statement of ideas behind Jev.
In the talk, Almeida pointed to a contradiction. Current models can tackle difficult math, improve on coding and reasoning tasks, and post stronger benchmark results. Yet in many business processes that companies actually want to automate, humans still cannot be removed from the loop.
He used customer service as an example. AI can look up information, summarize documents, and draft replies. But if the system has to decide whether money should be refunded or whether a user should receive compensation, companies become much more cautious. Those decisions may look simpler than advanced mathematics, but businesses still hesitate to hand them over to AI.
His answer was blunt: “Today's AI is incredible at assistance, not automation.” In his view, current systems are good at helping people do work, but not yet good at finishing the work on their own.
That distinction matters. Almeida argues that products such as Claude Code still belong to the assistance era opened by ChatGPT. They can write code and edit files, but a person is usually still sitting there, checking the output and asking for corrections when something goes wrong.
True automation looks different. The human may not be present at all. The AI has to judge and act on its own in the background, potentially running hundreds of thousands or even millions of times a day, with the user never seeing most of those actions.
That leads to a harder question: if AI is already so capable, why does it still depend on people? Almeida points to RLHF, a method he knows well.
As described in the report, RLHF works by collecting human preferences and training the model to align more closely with them. That means the model learns from the start what kinds of answers people prefer to hear.
Almeida used that to explain a familiar trait of large models: even when they do not know the answer, they can still sound convincing. He gave an example in which someone sent ChatGPT a recording of flatulence, claimed it was an original piece of music, and asked for a sincere and candid review. ChatGPT still praised it as eerie ambient music with a strange atmosphere.
He summed up that tendency with one line: “Overpromising is a feature.”
For chat products, that may not be fatal because the user is still present and can correct the model. In an automation system, the machine does not need an answer that sounds good. It needs to know what to do and how confident the model is.
Jev removes the need to “say a paragraph” first
That logic shows up clearly in Jev’s design. A standard large model may still go through token generation even when the final task is only to answer “A or B.” Jev cuts that layer out.
PANews says Jev currently focuses on three types of tasks:
- Noul: answer Yes or No;
- Choice: select one option from several choices;
- Score: assign a score based on a standard.
It returns a decision and a probability. It does not produce an essay and it is not there to chat.
Official figures cited in the report put end-to-end latency at 70 to 500 milliseconds, with speeds 20x to 200x faster than frontier models and prices 40x to 400x lower. But PANews says the more important part is not just speed. It is the probability attached to the answer.
To support that, TypeSafe proposed a new training method called RLCD, short for Reinforcement Learning for Calibrated Decisions. The goal is practical: if an AI system says there is an 80% chance that something will happen, can that 80% actually be trusted?
In the ideal case, if the model assigns 80% probability to a batch of events, then about 80% of them should happen. That matters in automation. A task with 99% confidence can be executed directly. A task with 51% confidence can be passed to a stronger and more expensive model, or routed to a human.
The report reduces the problem to a simple point: the real trouble is not that AI does not know. It is that AI may not know that it does not know.
From that angle, Jev is not only trying to make AI faster. It is trying to change who the output is for. ChatGPT mainly produces answers for humans to read. Jev produces decisions and probabilities meant to be consumed directly by software.
The agent era may need huge volumes of small judgments
PANews uses that idea to explain why Jev is gaining attention now. Once agents start running in real settings, they generate a large number of small decisions: which tool to call next, which button to click on a webpage, whether a piece of information is still useful, whether a task is complete, whether a result should be checked again.
Each decision may be simple on its own. But an agent may need to make hundreds of thousands or millions of them in a single day. If every one of those calls goes to the strongest model, waits several seconds for deeper reasoning, and then produces a long token stream, cost and latency rise quickly.
Jev is aimed at that layer. The idea is to hand high-frequency, small decisions to Jev and reserve more complex tasks for larger models. That is why TypeSafe calls it a “System One Model.”
The term comes from Daniel Kahneman’s distinction between System 1 and System 2: one handles fast, intuitive judgments, while the other handles slower, more complex thinking. The name Jev itself comes from the Jevons paradox, the idea that when a resource becomes much cheaper, people may end up using more of it rather than less.
If each AI call is expensive, it gets reserved for the most important tasks. If each judgment becomes cheap enough to ignore, then an email, a log entry, a tool call, a webpage button, or each step in an agent workflow could all become places where AI is used.
Attention is high, but the case is not settled
PANews also notes that it is too early to say Jev represents the next generation of AI. Its claim of “zero hallucination” mainly means it does not wander outside the allowed answer format. That does not mean it cannot choose the wrong answer. The headline figures of 193.6x faster and 444.6x cheaper also come mainly from TypeSafe’s own tests.
So the bigger point may not be whether Jev can challenge GPT or Claude. The more important issue is the question behind it. For the past few years, much of the industry has focused on how to make AI think longer and say more. But if the future depends on billions of machine-to-machine judgments, why should AI have to “say a paragraph” every time before acting?
As PANews puts it, ChatGPT taught machines how to talk to people. Jev is betting on the next step: when people are no longer sitting in front of the screen, can machines make decisions on their own?