I’m using Jev more than Opus 5.5 or GPT-6. Here’s why.

1. 0:00 – 2:49

   Jev launch and what makes it different from every other model

   1. CVClaire Vo0:00

      Jev, Jev, Jev. Welcome to Jev week on How I AI. We have seen a lot of new models be released in the last five days. We saw Opus 5.5, we saw GPT-6 Sol, GPT-6 Luna. Muse is blowing up the timeline. Everybody still loves their GroqBots, and yet there is one thing that I wanna talk about in AI right now, and that is this fast, cheap, doesn't speak System 1 decision model from TypeSafe AI. As soon as I saw this trending on X, as soon as I saw it launched, I immediately started testing it. Now, what I will say is, more than any other model I've experienced lately, Jev has been the one that has exploded use cases, personal productivity use cases, code use cases, product use cases. Today, I'm gonna give you a very quick whirlwind tour of what Jev is, what it will give you back, and what it won't, and then how I have used it over the last week to do work that I think is worth hundreds of thousands, if not millions of dollars. And I have probably spent sub $10 on Jev tokens. A lot of it has been subsidized because Jev is currently free on the AI Gateway by Vercel. But even at list price [chuckles] it is a very inexpensive model, it is a very effective model, and it is going to be in the middle of almost everything I build from here on out. So let's get to it. This episode is brought to you by OpenArt Arena, the global leaderboard for creative intelligence. Every week, new AI models launch, and every one claims to be the best, but best at what? OpenArt Arena is built to answer the question that actually matters: Which model is best for your specific job? Instead of one overall winner, OpenArt Arena ranks models across real creative tasks, from advertising and film to animation, product, graphic design, editing, and lip sync, covering both image and video. And these rankings aren't based on hype. They're judged by professionals, industry leaders, and working creators through blind evaluations, so judges never know which model produced which output. That means you can see how models actually perform when it comes to the creative work you're doing. So stop guessing which model to use. Explore rankings based on real creative work and find the right model for your project and save time and cost. See the rankings at OpenArt Arena.

2. 2:49 – 5:28

   Type-safe values explained

   1. CVClaire Vo2:49

      So if I were to explain Jev to you, I would go to this table on the TypeSafe blog post announcing Jev, and it basically compares normal LLMs on the left, Jev style LLMs on the right. Both take in unstructured data as inputs. Both take in text as inputs. Jev does not take in images, but it takes in text, and it takes in text descriptions of images if you really need to get there. The outputs, though, are very different. With the standard LLMs that you're used to working with, you are getting strings and generated text out. So you are getting text in, text out. With Jev, you are getting text in, type-safe values out, and what I mean by type-safe values is these are values that are predefined that then Jev picks from and selects and returns to you. We will show you what those values are, but essentially they're like it's this or that, it's yes or no, it's one through 10. Like, it's pretty simple. Now it sounds simple, but it is incredibly powerful when you put it against the right problem. The other thing that you will notice about Jev is it is cheap AF, and it is fast AF. And so if you look at the current cost of input and output tokens, it can be pennies to tens or hundreds of dollars per million output tokens. Jev only charges you on input tokens 'cause it barely outputs anything, and it is 4 cents per million input tokens. It is, like, dirt freaking cheap. And because they output basically nothing, they don't even charge you for output tokens, um, whereas the standard LLMs are gonna charge you tons for the output tokens. And then I'll go into use cases, which is how you would use, like, a standard LLM versus why you would use Jev. You know, standard LLMs, chatbots where you want text in, you want text out, coding where you want, you know, code to be produced. Um, and so it's, it's great for things where you need stuff generated, but Jev is really great for making decisions. That's ultimately why I've been calling this... It's like a decision model. So you give it a decision, and it makes a decision, and an, and it's high agency, I would say. So if you need smart if statements, should I go left or right? If it's good, put it to this person. If it's bad, send it to support. All that kind of stuff Jev is really good at. It's really good at classifying data. That's a lot of what I'm gonna talk to you about today, and it's really good at real time. So it is fast, fast, fast. So you can put an LLM in a real-time loop in a way that was not performant enough with these other models.

3. 5:28 – 7:39

   Understanding Jev outputs

   1. CVClaire Vo5:28

      I do wanna explain exactly what Jev outputs, and so I would highly recommend if you're gonna use Jev, go to their docs. The two places that I'm using the most are the primitives, so what it returns, and then I use some of the patterns in cookbooks. So what will Jev return to you? It'll essentially return one of three things. It will return a choice, which means you can give it text, you can give it a list of choices, and it will pick a choice. So if my question is, "What should I wear on this date?" And my choice is a dress, jeans, workout clothes, ski gear, or nothing It'll like maybe pick the dress, right? And so that is choice. If I select score, it's going to say like how severe is this? Or on a score from one to five what it is. This is really good for severity rankings. So say the example here is a great one, which is like if you get a bug, you wanna rank it like cosmetic, broken, or blocking. Jev will triage that bug and give it a score. And then finally, there's a null. It's a version of a Boolean. It basically tells you what the likelihood that the answer to a question is yes. So if the answer to a question is yes or no, is Claire a podcaster? It's gonna give you like a 99% null because there's a 99% chance that I am a podcaster. Okay. So this is how, you know, you would think about the three things it can return. And again, very simple, but if you've been a software engineer, this is like 90% of software engineering is like doing these things, returning a choice, scoring something, routing, saying yes or no. And so it is just so, so, so powerful and the number one thing I have been using it for is classification of vast sets of unstructured data that would've been annoying to classify but is very high value. So I'm gonna show you those use cases and hopefully this will inspire you about how you can use

4. 7:39 – 11:12

   Use case 1: PR categorization and pairwise clustering

   1. CVClaire Vo7:39

      Jev. Okay, so I'm gonna pull up Codex and show you like two or three examples that are super simple to run, incredibly cheap, and incredibly powerful. This is one CTOs, VPs of product, V, you know, chief product officers listen up. This is the one that three years ago I would've paid truly $100,000 for. So what I had Jev do is look at thousands of PRs. I got PRs. I have connection to do-- to GitHub. Pull every single PR. And then what I had it do, which would've been so impractical in the past, is I said, "Categorize all the PRs." And so use Jev to do pairwise relations, and so this is something that I found Jev is really good at, is compare like PR A to PR B and say, "Are you the same? Are you working on the same thematic area or not?" Again, Jev's not gonna tell you what the thematic area is unless you give it choices. Instead it's gonna say like, "Yes, these two are related," or, "No, they're not." And I've been doing this a lot where I'm taking vast amounts of data and I'm saying like, "Yes, related, not related. Yes related, not, not related." And it's doing clustering, which I find very, very useful. So like create these clusters of data sets, and that's what Jev did. And then I used a really cheap model, I used Gemini Flash-Lite to then take those clusters and categorize them. And s- I first ran it on our marketing site where we only had like 112 PRs. It cost me 1.1 cent. I guess they wouldn't round down that .1 penny. And you can see here it lets me see how much of my work is related to maintenance, content, tools, site and conversion services, or docs. Super useful. Again, it ran very, very, very fast and very cheap, but then I ran it on my ChatPRD app, which has about 2,000 PRs on it to date in this calendar year. Because it has a lot more PRs it cost me a lot more money, and by a lot more money I mean nine whole cents. It cost me 9 cents to do this. It took probably about two minutes. What it did is it analyzed 1,700 PRs. It did 17, found 17,000 pairs in those PRs of things that could be matched, and then the themes were labeled by Gemini Flash-Lite, and it pulled all of the data out. I mean, again, CTOs, CEOs, like I know you're asked this by the board and by your team and by your boss all the time, like what percentage of work is going to what initiatives? Like tell me the percentage to tech debt. Tell me the t- percentage to this product or that product. And you can see here it got down, it excluded our docs PRs because those are completely separate. Almost 30% of the effort, the PRs that we do at ChatPRD are around platform, security, and infrastructure, so I'm being a good citizen and investing in security, performance, and infrastructure. And then of course, because we're a chatbot, a conversational AI, reliability, and then some product areas that we invest in are like data and integrations, document editing, and prototyping, which is new. And so this was so cheap, so fast. It is accurate. Like I can just tell you, yes, this is where we're investing our time, and you can even see as we invest more in different things over the last couple months it goes up. And so this is just an example of how powerful, cheap, and good Jev can be on large data

5. 11:12 – 13:00

   Use case 2: analyzing your own local Claude Code and Codex sessions

   1. CVClaire Vo11:12

      sets. And then I wanna give you all one that even if you don't have like a big repo with PRs everywhere that you can run, which is you can run this on your local Claude Code and Codex sessions. So all of your Claude Code and Codex sessions are stored locally, and so you can actually run this analysis on everything stored on your local machine. So I did that while Codex kicked off a thread to do that, and you can see, depending on what you count as like effort, how much I'm spending on different things. So I do a lot of product engineering, and so if we're grouping by user turns, in January I was almost exclusively doing engineering tasks in Claude Code and Codex. And then as you kind of like come into this new world, now in September less than 40% it looks like of my tasks are actually engineering tasks. I'm doing a lot more work with agents, which makes a lot of sense. And then I'm doing a lot more, like, publishing media. I'm doing a lot of our videos through Codex. And then you can see I have a new business, so client delivery, family and personal stuff, and then where it couldn't categorize or it was unclear. If I take it to session days, you can see again most of my sessions were coding and now it's, like, more equally split across different use cases. And so if you think about this, it is just super useful to do meta-analysis on all this data that's sitting on your desktop that LLMs can totally parse. It will cost you basically no money and give you a lot of insight that I think previously would have been hard to get. So these are two use cases I think everybody who is coding should do.

6. 13:00 – 14:30

   Use case 3: Gmail triage with Jev scoring and LLM follow-up

   1. CVClaire Vo13:00

      The third one, and I won't show it because it's a lot of my private information, is I did run this on my personal Gmail. So I basically said, like, given a subject line and a snippet, categorize whether or not I can, like, delete this email. Like score whether or not I can delete this email. It did it very fast, and then it gave me very clean tags that I could go through and then have another model work through categorized emails. And so this is where I would say, like, Jev alone is okay. Jev with an LLM buddy is super powerful. So what I like to do with Jev is I like to take a big corpus of information, tag it, categorize it, cluster it, filter it, and then apply really precise AI actions to the right clusters. And so that could be take your bugs, take your high severity ones, and really triage them deeply. It could be take your emails, group them into, like, you can definitely delete them and ignore them, delete them, and then work your other emails with an agent. It's just all those things that you can do where a very fast but accurate filter can be helpful. So I think this is so powerful. I hope it unlocks your minds on what Jev can do. I wanna show you, like, the biggest version of this that I've done just to, like, kind of give you numbers and ideas of what I'm working on. And then I'll show a couple, like, fun little apps that you can build with Jev that I think would've been hard to build before. So

7. 14:30 – 18:17

   Use case 4: ChatPRD’s product insights graph

   1. CVClaire Vo14:30

      the example I wanna give is, you all have heard me talk about this product. So in ChatPRD, I'm trying to build, like, a product insights graph, basically. I'm, like, trying to suck in everybody's data and tell you really interesting insights about it, and I have used every frontier model there is to try to figure out. And it's just really hard. There's just too much nuance in all the data to kind of, like, grok all this, get it in structured format, do the right, like, structured, unstructured. It's super expensive. I've spent thousands of dollars, if not tens of thousands of dollars trying to prototype this. It's just hard. And I think I cracked this baby with Jev. I kind of wanna show you what that means, [lips smack] which is it's not that I-- like Jev has done it all, at all. It is actually I have figured out where classification, clustering, and mapping are important, and I put Jev there. And then I figure out where, like, the big brains, like analysis, strategy, insights extraction, and I put, like, Astra there. And then, like, where generation is important, I use like a Sol or a Luna. And so if you look at this map, and it's, like, probably not that interesting to you all because it's very internal to ChatPRD. But you can see, like, I extract things, I extract labels. I then quickly assign everything to those labels with Jev. Then I pair those clusters and those groups with Astra and I say, like, "What the heck's going on across all of this?" And then it extracts more insights, and all of a sudden I have this really cool product. And just to give you a sense of, like, the scope of data here, [lips smack] I'm probably pulling about 1,000, um, 1,100 individual signals into this. So these are PRs, these are support tickets, these are Granola conversations, these are, um, Linear tickets. They're, like, kind of, like, all the signals in our business about what people are worried about and what is going on. And then I'm getting now, like, this gap between what are customers telling us they want and what are we actually working on. And I can get those trends over time. I can get them categorized. I can get details about them. So it's just really super fascinating, very helpful, has made this feature, I would say, probably more margin accretive to my business, if you all know what that means. Because before I was really, like, trying to brute force with these, like, brainy models what I needed to work on and how I could do this data, and now I can, like, combine these, like, very cheap decision models like Jev with a really brainy model and get this, like, interesting, completely hard to generate set of data, context, like auto wiki. It's, like, really unlocked this product, and it has taken about 1,100 raw sources of data and done over, like, 200,000 classification and pairwise groupings, and it probably cost me four bucks on the Jev side. It's cost me a lot more on the Astra side. [lips smack] But I just think, like, thinking through where classification, like super smart classification clustering decisions could unlock really complex products and how you might use, as I'm showing here, Jev alongside some smarter models. It's, like, really blowing my mind right now and I think is, it sh- should be interesting for those out there building interesting data

8. 18:17 – 22:14

   Demo: How I AI audience signal dashboard

   1. CVClaire Vo18:17

      products. So, so far I've told you, like, what Jev is, how you can use it on PR review or data analysis, how you can use it on your local sessions, how you could even use it on your email. But I'm gonna show you two things that I built with it in an afternoon that I think are really cool. And this is my attempt to, like, go against the, like, eye-popping demos that you're seeing on X. I think people have shown interesting things, but they actually haven't shown how you have to build it and how it works to get that real-time effect. And so I'm gonna show you these apps live because I think it's important to kinda, like, understand where the latency comes from and what the real experience is as opposed to, like, kind of a 30-second demo that goes, goes viral. But I do think it shows some, like, pretty cool stuff. So the first one I want to do is How I AI audience signals. So you wonderful people give us comments on the How I AI podcast and I go through them. I read them every day and I reply to them when I can, but I've never done analysis on them. And there's about 4,500 comments and I just wanted to know, like, are they good? Are they bad? Are they happy? Are they sad? And I also wanted to know if you all had any episode ideas for me where I could extract them and come up with episodes that I could do like this Jev one. And so I hooked up the YouTube v3 API. You just have to enable it in Google Console. And then I said, "Pull all the comments and categorize them into positive, negative or neutral." So that would be a choice, right? Or a score probably. And then I said, "Also use Jev to identify whether or not includes a idea for a future episode and then analyze all the data and give me a dashboard to look at the data." And so you can see here about half are positive. There are 58 comments in here with episode ideas and then there's, like, quality praise, which is q- praise for our production team, which is like how nice the podcast looks. And so then it made me this dashboard which is telling me is it positive, neutral, negative or mixed. And then it gave me an audience request board so you all want like open weight models, you want instinct, you really want a comparison of Groq and Muse for personal use. That's coming soon. And then do I use Groq or GroqBot or like what am I using? And then you can also see by episode the sentiment, which is Ryan's very popular three-step AI coding workflow, 61% positive. We had an awesome Claude Code for product managers episode, 80% positive comments. And then we can actually go through the comments themselves and I, using Jev, built live search against these comments. So if I wanted to find comments about screen share, it would search and find very quickly something that says screen share. If I wanted to say, "What are the comments about SLOP?" I could do SLOP and then very quickly it scanned all 4,400 and found all the comments that are related to SLOP. So again, this is very performant, very fast. Just for like behind the scenes you have to like batch the results and then score them and then push the high scores up to get that kind of performance. So there is some architecture here. I don't wanna pretend like Jev, you just like slap Jev in the middle and evaluate all 4,400 that quickly. But you can imagine this is very useful. And so one, thanks for all the comments and two, this was really, really cool to build and then finally I have to give Astra props. Really nailed the How I AI podcast styling so good job on the front end.

9. 22:14 – 25:16

   Demo: voice-to-color emotion-mapping app

   1. CVClaire Vo22:14

      But I wanna show one last one before I get us out of here and then I said this was Jev week so like and subscribe for a second follow-up Jev episode with one of our How I AI guests that's gonna come later this week. But I wanted to show like kind of a fun real-time use case for Jev. So if you have been seeing any of these like real-time app applications of Jev like playing a video game or, you know, playing, playing Tetris, doing live search, all these things. You can do a lot of real-time stuff where making a very quick decision or returning a set of choices can be very powerful. Okay, so I built a real-time app that takes in voice and then uses Jev to determine what color is like my emotion and then returns a quote in reflection of my emotion. It uses OpenAI's Realtime Voice API, it uses Jev and then it uses like API Ninjas Quotes API. There's apparently an API for quotes. And so we're gonna see if this works and how fast it is. I'm feeling so tired today. I'm feeling very happy today. Okay, the quote's... Oh, there we [laughs] . See? The quote worked. Okay, hold on. Ooh, it's excited. I'm in love today. I can't wait for the weekend. Doing this podcast is the best job ever. Look at that. So we got all these-- Okay, I'm gonna turn it off 'cause it's gonna like keep listening to me. Look at all these quotes where like immediately it picked the right color, it picked the right quote. You can tell the quote API has a little latency in it but- You can imagine how fun it would be to be able to build these real-time experiences. And again, when it's like pick the right color to match the sentence, it can do that really well. Just for you all to know kind of how it works, I gave it a list of hex values, so a list of colors. Every sentence or, like, phrase it ingests from the Realtime API, it asks what color or it scores the colors. It gives the top ranking score color, and then it also picks from a set list of filters for the Quotes API. So it filters the quotes, and then it scores them against my sentiment, and then it shows it on the screen. So again, like, running locally, not the fastest. I'm sure I could optimize it by caching a bunch of this stuff and, you know, doing all sorts of things. But again, like, a little, like, magic use case that I think is really indicative of the kinds of things you can build with

10. 25:16 – 26:22

    Jev week recap and what’s coming in episode 2

    1. CVClaire Vo25:16

       Jev. So that is my Jev 101, how to use it on your own data, how to use it with other models, and how to build some fun real-time experiences in your app intro. Again, as I said, this is Jev episode number one this week, so we're gonna have another one midweek with one of our most popular guests. So if you are excited about that episode, subscribe, comment, let me know what you want us to talk about, what questions I can answer about Jev. And until then, it has been so nice showing you my favorite new model. Thanks for joining How I AI. [upbeat music] Thanks so much for watching. If you enjoyed the show, please like and subscribe here on YouTube, or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiai pod.com. See you next time.

Episode duration: 26:24