Jev made headlines over the weekend: one of the people involved in creating GPT is starting to reflect on RLHF—AI shouldn't always need human supervision.
In the past two days, a somewhat unusual AI model has suddenly gone viral.
Its name is Jev .
It doesn't chat, it doesn't write code, and it won't even generate a long answer like ChatGPT. It only does one thing: make judgments.
But this model, which seems to have had its capabilities "cut in half", suddenly became popular among developers.
Some have used it to analyze 724 real-time ads in 40 seconds, making a total of 8,724 judgments; others have integrated it into Claude Code to clean up useless context; still others have it act as a "referee" for AI agents, checking whether tasks have actually been completed. LangChain has also begun testing Jev's performance as an agent evaluator.
Even more outrageous are the speed and price.
In tests published by TypeSafe, Jev achieved speed improvements of up to approximately 193.6 times and cost reductions of up to 444.6 times. Input costs were only $0.042 per million tokens, and output tokens were even free.
Why has an AI with seemingly fewer capabilities become so popular?
Because in the agent era, AI may need more than just "deliberation"; it also needs massive, rapid, and inexpensive judgments : what to do next, which tool to use, and whether the task is complete. More importantly, these judgments will need to be made by AI in the background, eliminating the need for humans to constantly sit in front of a screen. What Jev values are these small decisions that may occur millions of times every day.
What's even more interesting is that Diogo Almeida, the founder of the Jev model , was involved in RLHF, and now he is beginning to reflect on RLHF: the training method that made ChatGPT useful may not be suitable for AI to truly become automated.
A little over a month ago, Almeida gave a speech. Looking back now, that speech was practically a "thought manual" for Jev.
01 AI can do advanced mathematics, so why can't it do customer service well?
Diogo Almeida's resume is quite unique.
He previously worked at OpenAI and was involved in projects related to GPT-4, ChatGPT, and InstructGPT/RLHF. In other words, he was directly involved in building one of the most important post-training paradigms for today's large-scale models.
But in that speech, he started by joking that he was one of the few people within OpenAI who openly criticized ChatGPT.
His speech focused more directly on the topic: What's Next After RLHF?
Diogo first raised a seemingly contradictory question.
Today's large models can tackle very difficult mathematical problems, and they are constantly improving in terms of code, reasoning, and various benchmarks.
However, when it comes to the business operations that many companies really want to automate, they can never get rid of the people.
For example, customer service.
Let AI search for information, summarize documents, and draft responses—no problem.
But what if we let AI decide for itself: Should the money be refunded? Should the user seek compensation?
Businesses immediately became cautious.
These things seem much simpler than advanced mathematics, so why are we afraid to entrust them to AI?
Diogo's answer is simple: Today's AI is incredible at assistance, not automation.
Today's AI is very good at helping you with tasks , but it's not quite able to complete them on its own yet.
These two things may seem to differ only slightly, but they are actually completely different.
No matter how powerful Claude Code is, you're still usually sitting in front of the computer. It writes code, you read it; it modifies files, you check them; if there's a mistake, you have it fix it again.
Therefore, in Diogo's view, Claude Code still belongs to the "assisted era" initiated by ChatGPT.
What is true automation?
The person wasn't even there.
The AI makes its own judgments and executes its own actions in the background, possibly running hundreds of thousands or even millions of times a day, and you may never even see what it does.
This raises the question: Why is it that today's AI, despite being so intelligent, still cannot function without humans?
Diogo pointed the finger at something he was very familiar with— RLHF.
02 When we train AI, we put humans into the circuit.
This is somewhat ironic.
Because RLHF was one of the technical routes that Diogo helped promote back then.
The basic logic of RLHF is actually not complicated: collect people's preferences and then make the model more and more in line with people's preferences.
So Diogo gave a very straightforward explanation in his speech: Why do today's large models always need someone in the loop?
Because when we train it, we literally put a person into that circuit.
The model learns from the very beginning: what kind of answers do people prefer?
This also explains a characteristic of large models that we are already very familiar with— even if we don't know it, it can often sound very plausible.
Diogo gave a very harsh example at the event.
Someone sent ChatGPT a recording of farting, telling it it was a piece of music they had composed, and asking it to give a "sincere and frank" evaluation.
ChatGPT then seriously praised it, saying it was a piece of ambient music with a very eerie and mysterious atmosphere.
Diogo even summarized it in one sentence: "Overpromising is a feature."
Over-promising is not a bug, it's a feature.
For chat products, this may not be fatal. The user is still in front of the screen, and mistakes can be corrected.
But true automated systems are completely different.
The machine doesn't care whether your answer sounds good or not; it only needs to know two things: what to do and how confident you are.
This explains why Jev, released more than a month later, looks so unusual.
03 So, Jev simply stopped letting AI "speak".
Even if a typical large model only needs to answer "A or B" in the end, it often still needs to go through the process of generating tokens.
Jev simply cut that part out.
It currently does three main things:
- No, answer Yes or No;
- Choice: Choose one from several options;
- Score, a score based on standards.
Then it directly returns the judgment and probability .
I won't write you long essays, nor will I chat with you.
The official end-to-end latency is as low as 70-500 milliseconds, which is 20-200 times faster than cutting-edge models and 40-400 times cheaper.
But the real key is not "speed," but the probability that follows.
To address this, TypeSafe proposes a novel training method: RLCD, Reinforcement Learning for Calibrated Decisions.
The problem it aims to solve is very real: if AI tells you that there is an 80% probability that something will happen, can you really trust that 80%?
Ideally, a group of events that the model judges to have an 80% probability should eventually have about an 80% probability of actually happening.
This is very important in automated systems.
With a 99% certainty, it can be executed directly.
With a 51% certainty, you can throw it at a stronger, more expensive, larger model, or even transfer it to someone else.
The real problem isn't that AI doesn't know, but that AI doesn't know what it doesn't know.
So what Jev really wants to change is the object of AI output.
In the past, the answers generated by ChatGPT were primarily intended for human use . Jev's judgments and probabilities, however, are designed for direct use by software .
04 In the Agent Era, what may be needed is not a larger brain.
This also explains why Jev has become popular now.
Because once the agent is actually running, it will generate a massive number of small judgments:
Which tool should I use next? Which button should I click on this webpage? Is this message still useful? Has the task been completed? Should I check the results again?
These problems are not difficult individually, but an agent may need to make hundreds of thousands or millions of judgments a day.
If every time we call for the strongest model, spend a few seconds "deliberating" before spitting out a large number of tokens, the cost and latency will quickly increase.
What Jev wanted to take was this floor.
A large number of frequent, small decisions are handled by Jev, while tasks that truly require complex reasoning are handled by the larger model.
This is why TypeSafe calls Jev the System One Model .
This concept comes from Daniel Kahneman’s “System 1” and “System 2”: one is responsible for quick, intuitive judgments, and the other is responsible for slow, complex thinking.
Jev's name even comes from the "Jevens Paradox":
As a resource becomes cheaper, people may not necessarily use it less; on the contrary, they may use it more.
If using AI is expensive, you'll only use it where it's most important.
But what if an AI-generated decision is so cheap that it's almost negligible?
An email, a log entry, a tool call, a webpage button, and every step executed by the agent can potentially incorporate an AI judgment.
Of course, it's too early to say that Jev represents the next generation of AI.
Its so-called "zero illusion" means more that it will not deviate from the prescribed answer types and make up random answers, but it does not mean that it will not make the wrong choice; extreme data such as 193.6 times and 444.6 times also mainly come from TypeSafe's own tests.
But what's truly noteworthy about Jev's recent surge in popularity might not be its ability to challenge GPT or Claude.
Instead, it is someone involved in creating ChatGPT who is now re-examining a more fundamental question:
Over the past few years, the entire industry has been thinking about how to make AI think longer and talk more.
But if what's truly needed in the future are billions of machine-to-machine judgments, why does AI need to "say something" every time?
ChatGPT taught machines how to talk to people.
Jev's next bet is: when people are no longer sitting in front of the screen, can machines make decisions on their own?
Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.