[0:00] Ilan: Think you need to be a developer to use Jev, but we're going to show you how easy it is to set up and two ways that you can use it that have nothing to do with engineering.

[0:09] David: Right. Let's go.

[0:25] Ilan: Before we get into that, what is Jev? It's making its way across X these days, and people are hyped about it. It's a new model coming from TypeSafe AI. It's a company that was started by one of the co-founders of OpenAI, and this model makes quick decisions. They call it a System One model. If you've read Daniel Kahneman's Thinking Fast and Slow,

[0:51] David: That's cool. When you were explaining it earlier, I was thinking exactly that, like Thinking Fast and Slow.

[0:56] Ilan: Right. That's where it comes from. System One models, they make quick decisions, and they say that LLMs are System Two models. They make slow reasoned decisions, and it actually comes into the type of model itself. It is not an LLM. It is not a large language model. How is it different? A typical LLM, like Claude or ChatGPT, uses RLHF or Reinforcement Learning through human feedback. That's where you tell it, hey, you did a good job, or no, you didn't do a good job. And it uses that to train its answers over time to be more like what people need to hear.

[1:38] David: And this is where the sycophantic behavior originated.

[1:42] Ilan: Really wants to please you. It really wants to get that thumbs up at the end that it did a good job. Jev uses a different method. It's called Reinforcement Learning for Calibrated Decisions, or RLCD. What that means is that it's trained to make predictions on some piece of data. And it gets reinforced when it makes a correct prediction on that data. So if it says that something should happen 20 % of the time, the data should show that it happens 20 % of the time.

[2:12] David: Right. And so it is optimized for being correct as opposed to being preferable, And because of this difference, it is up to 200 times faster and 400 times cheaper.

[2:22] Ilan: And it's not hard to use. So let's show you how we set it up in Claude and how you can use it for some real world examples. Setting up Jev in Claude chat GPT or whatever agent tool you use is really not that hard. It uses an API just like most models. And you can get access to that API through a couple of different sources. They can come from type safe AI itself. It can also come from open router, which hosts basically every model that you can possibly imagine. It can also come directly from Vercel or from Cloudflare's AI tools. However, use type safe directly. And you know why? They give you five bucks a month of free credit. And Jev is so cheap that that'll probably cover all of your costs if you're testing it out. So the best way to use Jev inside your environment is to set up a skill for it. And that skill will allow you to say, hey, use Jev to do X and it will just use it correctly. And if you're not sure how to do that, all you need to do is copy this prompt. You get that prompt in, your agent will set up the Jev environment and they'll even do a quick test for you just to make sure that it's working correctly. Do you use Jev once you have it set up? Are two examples of how we used it right away. The first one is for classification. We're both product people. We get requests from customers all the time, asking us to build new things for them. And we have to decide, is it going to make an impact and is it worth the effort? Out the difference using Jev to do the classification versus Sonnet, We tested it out on 50 feature requests from customers Jev got the task done in half a second for less than a tenth of a cent. Sonnet, on the other hand, took two and a half seconds and it took two cents to get it done. You can imagine how quickly that scales if

[4:31] David+Ilan: So Jev was over four times faster and over 20 times cheaper than Sonnet. And that comes because Jev's input token price is only four cents per million tokens, and they don't charge for output tokens.

[4:46] David: But it's not just us, others have replaced GPT 5.6 Luna in a classification scenario, where they... They found that Jev was five to 18 times faster and more accurate.

[4:59] Ilan: That's insane, man. It is so cool what this model can do and how fast it can do it. All right, here's a second example of how you can use Jev. Jev is great at semantic search. What does that mean? Taking some language that you're putting in and very quickly answering you where you should look in a database, in a document, in a Wiki. So we ran a test with this. In a 100-member Wiki, we searched two questions. In the first question, which had a clear answer, Jev answered in 370 milliseconds And cost way less than a hundredth of a cent.

[5:39] Ilan+David: On the other hand, took 1.3 seconds and it cost 48 times more. So you got the same right answer for way less cost and way less time. On a second test, we gave it a trick question where it couldn't answer the question correctly. And Jev answered in less than half a second still cost less than a hundredth of a cent. Sonnet gave the right answer as well, but it took one and a half seconds and it cost 61 times more to get that answer right. The reason that happens is Jev doesn't have to reread over the entire database each time. That's right. Jev parallelizes its processing.

[6:25] David: And so if you want to think about other scenarios, just remember that there are three things that Jev can output. A choice. So if you give it, hey, A, B, or C, which one is it? It will tell you just that one choice. It can be a score. So a range from like one to 10 on a scale and it'll give you that number. It can be a binary question. So just yes or no. True or false? And in all three of these cases, it'll give you a confidence in its answer. And with that, you have what you need to go out and try Jev. So let us know what you did in the comments.

[6:57] Ilan: And subscribe. Talk to you next time.

[7:00] David: You next time.