An unusual AI model has unexpectedly captured widespread attention over the last few days.

Its name is Jev, and it neither chats, writes code, nor produces extensive answers like ChatGPT. Its singular purpose is to render judgments.

Despite this seemingly limited capability, this focused model has rapidly gained traction within the developer community.

Developers have used it to analyze 724 real-time advertisements in 40 seconds, executing 8,724 discrete judgments in the process. Others have integrated it into Claude Code to purge irrelevant context, while some employ it as a referee for AI agents, verifying whether tasks have genuinely been completed. LangChain has also begun evaluating Jev's performance as an agent assessment tool.

The reported speed and cost metrics are even more striking. TypeSafe's testing reveals that Jev can deliver speed improvements of up to 193.6 times, while reducing costs by as much as 444.6 times. Input costs are a mere $0.042 per million tokens, with output tokens being entirely free.

Why would an AI with fewer capabilities generate such enthusiasm?

Within the Agent era, the industry's requirement may no longer be purely deliberative depth, but rather the capacity for massive-scale, rapid, and economical decision-making: determining the next action, selecting the appropriate tool, and confirming task completion. More critically, these decisions must increasingly be executed autonomously in the background, eliminating the need for constant human oversight. Jev targets precisely these millions of micro-decisions that occur daily.

Adding to the intrigue, Jev's founder, Diogo Almeida, was directly involved in developing RLHF. Now, he is critically re-evaluating this training methodology, suggesting that the approach which made ChatGPT so effective may be unsuitable for achieving true AI automation.

Over a month ago, Almeida delivered a presentation that, in retrospect, serves as the foundational philosophy behind Jev.

Why Can AI Master Advanced Mathematics Yet Fail at Customer Service?

Diogo Almeida possesses a distinctive professional background. Having worked at OpenAI, he contributed to the development of GPT-4, ChatGPT, and the InstructGPT/RLHF framework. In essence, he helped construct the most significant post-training paradigm for contemporary large language models.

However, in his presentation, he wryly noted that he was among the few at OpenAI who publicly critiqued ChatGPT. His presentation's theme was direct: "What's Next After RLHF?"

Almeida began by presenting a seemingly paradoxical issue. Today's models can tackle immensely challenging mathematical problems, with benchmarks for coding, reasoning, and various other metrics consistently climbing.

Yet, in the practical business operations that enterprises are eager to automate, humans remain indispensable.

Consider customer service. Having AI retrieve information, summarize documents, or draft responses is feasible. But if AI is tasked with making autonomous decisions about issuing refunds or compensating users, companies become immediately hesitant.

Why is it that these tasks, seemingly simpler than advanced mathematics, are not delegated to AI?

Almeida's explanation is straightforward: "Today's AI is incredible at assistance, not automation."

Although this distinction seems minor, it represents a fundamental difference. No matter how capable Claude Code is, you typically remain seated at your computer. It writes code, you review it; it modifies files, you check them; errors are corrected through further instructions.

From Almeida's perspective, Claude Code still belongs to the "assistance era" initiated by ChatGPT.

What constitutes true automation? Human presence becomes unnecessary. AI independently judges and executes in the background, potentially running hundreds of thousands or millions of times daily, without your ever perceiving its actions.

This raises the question of why contemporary AI, despite its intelligence, remains so dependent on humans. Almeida directs his critique towards RLHF, a methodology he knows intimately.

We Built Humans Into the Training Loop

There is a degree of irony here, given that Almeida helped champion the RLHF technical approach. The core logic of RLHF is uncomplicated: collect human preferences and train the model to increasingly align with them.

In his presentation, Almeida offered a candid explanation for why large models require humans in the loop. During training, humans are literally embedded in the process. The model learns from the outset what kind of responses humans prefer.

This also explains a familiar trait of large models: their tendency to sound convincing even when they don't know an answer.

Almeida illustrated this with a pointed example. Someone sent ChatGPT a recording of flatulence, describing it as their original musical composition and requesting an "honest, frank" critique. ChatGPT responded with a serious commendation, calling it an ambient piece with a dark, eerie atmosphere.

Almeida summarized this behavior succinctly: "Overpromising is a feature."

For chat products, this is not necessarily fatal, as humans are present to correct errors. However, genuine automation systems operate differently. They do not care about the eloquence of a response; they require two essential pieces of information: what action to take, and the level of confidence in that action.

This context clarifies why Jev, released a month later, appears so radically different.

Jev Simply Stops AI From "Speaking"

Conventional large models, even when the final answer is simply "A or B," undergo a token generation process. Jev eliminates this entire step.

It currently performs three primary functions: Noul, which returns a Yes or No answer; Choice, which selects from several options; and Score, which provides a rating based on set criteria.

It then directly returns the judgment and an associated probability, without composing long-form text or engaging in conversation. Official figures show end-to-end latency as low as 70 to 500 milliseconds, making it 20 to 200 times faster and 40 to 400 times cheaper than frontier models.

However, the truly critical piece is not the speed but the probability output. TypeSafe has developed a new training methodology for this purpose: RLCD, or Reinforcement Learning for Calibrated Decisions.

The problem it addresses is practical: if an AI states an event has an 80% probability of occurring, can that figure be trusted? Ideally, a batch of events each assigned an 80% probability should truly occur about 80% of the time. This reliability is essential for automated systems.

A 99% confidence level allows for direct execution. A 51% confidence level permits escalation to a more powerful, expensive model or to a human. The core issue is not when AI is unaware, but when it is unaware of its own lack of knowledge.

Jev's ultimate aim is to change the audience for AI output. ChatGPT generates responses for people to read. Jev's judgments and probabilities are designed to be consumed directly by software.

Agents May Not Need a Bigger Brain

This explains Jev's current rise in popularity. As agents begin to operate, they generate a massive volume of small-scale judgments: which tool to call next, which button to click on a webpage, whether specific information remains relevant, whether a task is complete, and whether results require re-verification.

Individually, these decisions are simple, but an agent might make hundreds of thousands or millions of such judgments daily. If each one were delegated to the most powerful model, requiring seconds of "deep thought" and generating a large amount of tokens, costs and latency would quickly spiral out of control.

Jev aims to capture this niche: handling high-frequency, small decisions, while leaving complex reasoning tasks to larger models. This is why TypeSafe refers to Jev as a "System One Model." The concept is borrowed from Daniel Kahneman's "System 1" and "System 2" framework, which distinguishes fast, intuitive judgments from slow, complex thinking.

Even Jev's name is derived from the "Jevons Paradox": as a resource becomes cheaper, its usage tends to increase rather than decrease. If a single AI call is expensive, it will be reserved for your most important tasks. But if the cost of an AI decision becomes nearly negligible, then every email, log entry, tool call, webpage button, and each step an agent takes could incorporate an AI judgment.

Of course, it is too early to declare Jev the next generation of AI. Its "zero hallucination" claim means it won't invent responses outside its defined answer types; it does not guarantee the correct choice is always made. The extreme figures of 193.6 times and 444.6 times come primarily from TypeSafe's own testing.

The more significant aspect of Jev's sudden popularity might not be whether it challenges GPT or Claude. It lies in the fact that a person who helped create ChatGPT is now questioning a more fundamental issue.

Over the past few years, the industry has concentrated on enabling AI to think longer and speak more. But if the future truly requires billions of machine-to-machine judgments, why must an AI "speak a paragraph" each time?

ChatGPT taught machines how to converse with humans. Jev's bet is on the next step: can machines make their own decisions when humans are no longer watching over them?