Jev AI Just Changed AI Agents Forever

Julian Goldie (0:00)

Jev AI just changed AI agents forever. And the person behind it is one of the people who actually co-invented Chat Cheaper TA. His name is Diogo Almeida. He spent the last two years quiet, building a new way to train models. And this week, he released Jev, J-E-V.

And the numbers he put on it are 20 to 200 times 500 That's a lot faster than a normal AI model, and 40 to 400 times cheaper. And then somebody actually plugged it into Claude Code as a plugin. One second later, that Claude session went from nearly like 1 million tokens down to 86,000. One second, same work, same session, but a fraction of the context, which is pretty wild.

So what actually is this thing? And why is everyone building with it inside a week? Well, Jeff is not a chat bot. It's not a coding model. It doesn't write at all. You can't ask it for an email. You can't ask it for code even. You can't even ask it to explain why it picks something. What it does do is make decisions, fast ones, fastest of them. So you can hand it the current decision and the current situation. You hand it a question with set answers. It picks one and tells you how sure it is. Now, TypeSafe AI, the company behind it, calls this a System 1 model.

So you want to think of it like the fast part of your brain. When you see a red light, you don't write a paragraph about the red light. You just stop. That's System 1 Writing the paragraph is System 2 And right now, every AI agent you run uses the paragraph brain for everything, including the red lights. Now, here's what that actually looks like in real life. Let's say, for example, you have an agent doing research for you. Along the way, it has to answer a pile of tiny questions. Do I have enough sources yet? Should I right now keep looking? Is this good enough to send? Or should a person check it? Which tool do I use next? And every one of those tiny questions currently goes to a big model, a big model that reads the whole history, thinks about it, writes out its reasoning, formats it, hands it back. Seconds go by, tokens get spent, and the actual answer was one word. That's a gap. Jev feels. And once you understand that gap, you start seeing everywhere in your own setup. And there are three kinds of questions you can ask it, and they cover almost everyone. So the first one is choice. You pick from a list. Which step goes next? Which folder does this go in? Which of my writers handles this? It gives you the pick plus a probability for every option, plus how confident it is overall. The second is score. So you would rate this against levels you define yourself. How relevant is this source on a scale from unrelated to directly on topic? How strong is this lead? It gives you a number back, and it can land between levels, not just on them. And the third one is called null. It's a yes or no.

Does this message sound urgent? Does this need a human to approve it? It gives you the probability that the answer is yes. So 0.999 means almost certainly yes. Someone sitting near 0.5 means it genuinely doesn't know, which is useful information on its own. And that last part is the bit that most people skip over. Jev doesn't just answer. It tells you how sure it is. So you can set a line above this level, carry on automatically, blow it, stop and ask me. And that single idea changes how much you can safely handle over. Because the thing that stops most people automate the work isn't the AI being wrong. It's not knowing when it's wrong. And you can ask it lots of questions all at once. All the questions in one request get answered in parallel. So asking five questions in one request means instead of one barely changes, it doesn't really change the speed and barely changes the cost. You can ask, for example, which step next? How urgent is this? And does this need approval in a single go and get all three back together? Now, let me run you through what people actually built this in the first few days, because the results are specific and they're all public. Let's start with research.

So a developer called Hassan ran 1,018 AI research papers for it. He summarized each paper first, then sent the title, the summary, in 24 possible topics to Jeff and let Jeff pick the category. All 1,018 papers cost 8 cents total. The middle time per paper start to finish was 256 milliseconds, a quarter of a second per paper. Next, we've got email sorting.