How I AI - Jev: 8 real use cases for the fastest, cheapest model I’ve ever used | John Lindquist

Episode Date: September 30, 2026

John Lindquist created egghead.io, a developer education platform used by hundreds of thousands of working engineers. These days he’s building mega.dev, a hands-on program specifically for developers who want to do real work with AI agents, not just prototype them.What you’ll learn:Why Jev is a decision engine, not a chatbot, and what that distinction actually changes about how you buildHow John built a real-time voice to-do app that classifies and executes commands with no visible pauseThe data deduplication pattern that merges messy records in milliseconds using confidence scoresWhy Jev works best as a router, and how a single text input can navigate users deep into an appWhat a chess match between Jev and a low-reasoning LLM reveals about speed, cost, and when to use whichThe multi-step classification pattern John reaches for when one Jev pass isn’t enoughWhere Jev falls short, and when you should still reach for a full generative model—Brought to you by:Vanta—Automate compliance and simplify security—In this episode, we cover:(00:00) John Lindquist returns for Jev week(04:32) What Jev actually outputs(06:15) Demo: real-time voice to-do app(08:17) How sequential Jev calls chain together(10:38) Demo: plain English to function name (grocery cart)(11:50) Demo: data deduplication and record merging(13:45) Confidence scores and multi-model validation(15:06) Demo: Jev as a multi-level app router(18:23) Architecting around Jev(19:35) Demo: Jev vs. traditional LLM at chess (speed and cost benchmarks)(24:29) DOM interactions as a decision set, not an infinite canvas(28:21) Demo: Wikipedia “path to philosophy” route mapper(30:28) Demo: multi-agent coordination and collision avoidance(33:36) Demo: real-time presentation coach(36:56) Quick recap(39:54) Lightning round and final thoughts—Tools referenced:• Jev (TypeSafe AI decision model): https://typesafe.ai/blog/introducing-system-one-models-and-jev• Vercel AI Gateway: https://vercel.com/docs/ai-gateway• OpenRouter: https://openrouter.ai• Opus 5.5 (mentioned in context of iterative demo building): https://www.anthropic.com/claude-opus-5-5—Where to find John Lindquist:LinkedIn: linkedin.com/in/john-lindquist-84230766X: https://x.com/johnlindquistMega.dev: https://mega.dev/Egghead.io: https://egghead.io/—Where to find Claire Vo:ChatPRD: https://www.chatprd.ai/Website: https://clairevo.com/LinkedIn: https://www.linkedin.com/in/clairevo/X: https://x.com/clairevo—Production and marketing by https://penname.co/. For inquiries about sponsoring the podcast, email [e-mail].

In some ways, managers are really at risk. This concept of like org design and role design, you can put those skills to use when crafting your agents, which is why I am currently running no joke 40 graph bots right now. Whenever you have constrained inputs that are interacting with apps, it's Jeb is a good thing to reach for. This one is a favorite of mine because it represents like a workplace where you have multiple agents that could all have their own tasks

and avoiding collisions so that they never end up on the same space. At each step, It's going to be able to control these different agents and make sure that they get to their specific task as quickly as possible and that they never collide or interfere with the other agents. This one's really fun to me because it's a presentation coach. I click the microphone. I'd start talking. I have a list of bullet points in either a podcast or something that I want to make sure I cover or an interview or anything. It listens through everything I say and then starts checking the boxes to make sure that I covered everything.

This model, which is basically free and incredibly fast, allows you to do discovery over data in a way that feels like it's opening up my opportunities and allowing me to look at things that I thought weren't high ROI before. Welcome back to How IAI. I'm Claire ofau, product leader and AI obsessive here on a mission to help you build better with these new tools. Today I have repeat guest John Lindquist who has done one of our most popular episodes to date on the podcast. And here he is joining us on Jev Week at How IAI to talk about our favorite new model and all the amazing product-facing use cases that he's discovered testing Jev from TypeSafe AI. Let's get to it. John, welcome back to How IAI. I have to say this is Jev Week on How I AI. And the reason this Jev Week is a lot of models have come out in the past seven days.

you and I do a lot of AI things and yet when we're DMing back and forth we're talking about one thing and one thing only which is which is Jeb so you know folks that maybe listen to my episode earlier this week about why I'm hyped on Jeb I would love to hear why you're so excited you know I'm looking up here you have 23 demos

we could probably look at so you are going deep Jeb as well what about this model this framework this this way of thinking about building has gotten you excited. Fast and free are both amazing. If you've ever been tired sending a basic request to an LLM and waiting around for a bit just to have it like call a function, this is now essentially instant. And I think of all the time I've spent on this and the hundreds of demos I made,

I think I've spent 73 cents on it. And I love the one quote. someone said, I think I'm going to pass this inheritance of Joe onto my children once I die. Yeah, I like the one where it was like, type safe. Type safe is really taking off. They have to have made double digits in revenue by now. Yes. So yes, fast and free.

I also have had that experience where it ran over like, I don't know, 20,000 records. And it was like, I have charged you 0.4 cents. Right. And it exposes those scenarios where I thought, I thought, like, I had five gigabytes of Jason. I'm like, I had never passed this into an LLM. I just, like, I don't have that budget on and on. I'm like, let's see what happens.

And it was, it was 40 cents, you know. And I wasted it on a really stupid thing, like try and organize this or whatever. But five gigabytes of Jason is a lot of text. And it went through and did the whole thing in probably a couple minutes. That's been my experience, too, which is I'm looking at corpuses of data. that I would never just dump into an LLM, both from a cost perspective and honestly, from a time perspective. It's like this is not worth my time to even see if something's here. But then this model, which is basically free and incredibly fast, allows you to do discovery over data in a way that feels like it's opening up my opportunities and allowing me to like look at things that I thought weren't high ROI before.

for. Yeah, absolutely same. It's just opening doors that weren't or closed before. So just quick reminder for folks, Jev is not like the LLMs that you're used to and love. It does not output text. It outputs basically like decisions and scores and yes, no probabilities. And so you kind of put information in out in and then you get out very structured, type safe,

very structured, limited set of options. But the ways you can use this. are almost unlimited. And so John's going to walk us through some of his favorite demos of how he's using it. Hopefully we'll get a combination of inspiration on where a real-time, fast, cheat model can build cool things

and a little inside of like, what else is around this? Because what I found is I see a lot of these like shiny demos on the timeline. And then you're like, you're hiding your, you know, your embeddings and you're hiding like how this is actually working. And so I think you can lift the veil for us on how some of this works.

and where it's Jev alone, where it's Jev Plus, that would be awesome. Yeah, I would just, just to add to your definition of Jev, of it being TypeSafe. I think of LLMs being unstructured to unstructured, where you put text in, you get text out. And this one is similar. You put in unstructured sentences and data, but then you get structured data out. And the power there is it's humans talking to machines, and they're the scenarios where you have functions and APIs and everything set up, but I want to communicate with them using unstructured data. And we're so used to chatbots, and it's just not a chatbot anymore.

You need services, you need functions, you need things you want to throw a data at first, but you can talk to it however you want, which is incredible. Yep. Well, let's talk about talking to it. All right, yeah. So let's go ahead and try this out right away. So traditional to-do app, We have a list of to-dos. And the fascinating thing here, which I'll try right away, is I'll just say something like, book dentist appointment, remove.

Buy oat milk, complete. Review pull request, low priority. Now, the speed there was all done through Jev. It does a few passes of looking through the text, where it ensures that it lines up, if there's any like dictation mistakes and you say it doesn't get oat milk quite right, it'll go through and see does this quite match up with something or do I need to do a different, do I need to categorize this into something that exists like a confidence score?

Like it's a lot of confidence scores there. It matches it with the thing in the list and then it matches it with the function that's going to call right based on based on the way that the operation that you want to perform on top of that data. So it's multiple layers of analyze the dictation itself to see if that's like a valid dictation, then analyze if it matches up with a task, and then analyze if it matches up with the operation you want to perform on the task. And it can do that all live,

and it even, while you're dictating, it classifies if you've spoken to a point where it can take an action. So as the words come in, it's like, is now a time where I can, we have enough data to make a function call? where I can reliably keep on talking. Because you noticed that entire time I was talking, I never paused, I never hit enter.

It was able to classify, is this a moment where I have enough information to take action? So it's just showing all these pieces of Jev fitting together to give you this real time, like give me kind of streaming data in, like unstructured streaming data in, and I'll start putting it in all places it needs to go. I have a question for folks that have just have no idea

how to actually put this into place and want more of the tactics, which is, are those, you know, are those like sequential Jev calls? Are you like, yes, no, it's time to analyze. Yes, no. It's, this is a good transcript. You score. Here's the top task that relates to it. And score, this is the top tool call that relates to it. You know, is this like, because people are, I'm sure, trying to put in their mental model, like a tool call where you get all that structured data back. Like, you know, How have you set this up so people can kind of understand how they chain these jev pools together? My mental model is, remember back when we used to write code by hand. There is no conditions.

If else switch statements, all these things that were the moment of classification, and if you just think of a line-by-line program where things execute, you think of where do I need those conditions? And so as you're breaking the problem apart, whatever the thing is that you're building, just think of all the conditions that if else, the moments where it either needs to be put into this bucket or that bucket. And so as I was going through this, I built a lot of it backwards because I started with what I was building, marking it to do as either done or removing it or heading a new one and such. And I'm like, oh, wait, no, now I want the dictation to be able to select it. And so I just kind of built it backwards. And it was so easy with Opus 5.5 recently just to kind of wing it and like take it different directions and go through scenarios and see the cases that were exposed.

Because I had that final idea of I just want to, if you've ever used a Siri or a Google assistant or any of those and you try to manage like add to my to do list. And they're all terrible. And like how can I build like does Jeff enable this to actually. be good and I think it does. Like it's one of the most common things like managing. You could integrate your calendar into this. You could integrate it because they're all behind function calls, right, or MCPs or however you want to build it out. You could integrate anything with this and just have this top level dictation live inference running and it's something anyone could build cheap and free. Like it's amazing. Amazing. Okay. So you and I both did live voice apps. Mine was much more

emo, which was just talking to the screen and getting a color and an inspirational quote. Yours is much more practical, which maybe reflects our different personalities. Let's see what other use cases you have. And what I like about what you're showing us is I talked a lot about like internal data analysis and clustering and all this stuff. You have such great inspiration for how you can use this inside products or to do things that were previously really tedious. So I'd love to see another one of your favorite demos. Yeah, let's dive in. So this is one of your, so this is kind of breaks it down into its smallest pieces. If you think of a grocery store and if you think of I have a function called like Get Cart if we remember programming and you type in something

like what's in my cart. And that's how if you think a classification of regular text to functions, like this is just demoing back if the previous demo didn't make enough sense, think of like plain English language mapped to a function name. And that's one of the core foundation foundational pieces of what Jeff does. That's that demo. Classifying documents. Sounds like you kind of covered that. This one, I think, is incredible because if you've ever run a scenario in, I use Google

contacts, and often you put in the same contact twice, and you've probably gone through the process of, like, merging them together. Very common scenario where you have messy data of someone typed in something wrong, and they hit submit, and now you have two. company entries in your database or whatever. This can go through and go through all of the records and kind of merge those records together. So let's click on that. And it can go through and say this one wants to merge. So Cedar Grove Office products obviously matches with Cedar Grove Office. Ridgeway data is Ridgeway Analytics. And it can go through and apply those merges for you by

matching together and classifying. Is this one close enough to that one? And again, It's huge data sets merged together in milliseconds. This is like my favorite use case of Jev, which is like pairwise comparison of a lot of data to create grouping and clusters. And so in the episode that I just did, I did that over PRs. I were like, were these PRs on the same product area or the same thing or the same theme? But you could do data reconciliation, I think is a really useful one. I went through recently and I had like 2,000 passwords in my one password. And you know how like whenever you do your one password, it's like update, save, make new.

And I'm just like all over the place, had a whole bunch of different ones. And like I was just thinking how much faster it would have been to use Jeb and some of the metadata to just pair those up and merge them. Absolutely. And again, it's not like you're giving an example where there's like six, you know, things. You're talking about when you have 60,000 or 600,000 things, those pairs feel very expensive until you use something like this. And something we haven't shown is that there is a confidence score that comes back. So if there is a North Star Clinic and there's North Star Clinic services and it's not, and they don't quite line up. Like if you want, you want them only to merge if you're like 99% or more confident, you can you can set those parameters.

in there. So that is a knob you can turn as you go through a few checks. And I think like you mentioned before is after you do this pass, you can send a smarter LLM against the result and say, we were, the data did look like this, please, you know, either random sampling or the entire data set, ensure that this was done correctly. And then you can gain more and more confidence over time as you run more and more, you know, large data runs. One other thing you can do on that in addition to validating the data pairs is you can describe them. And so what I've done is do the pairs and then run. It doesn't even have to be an expensive model, but say like, we've decided these two are the same.

Explain to me why. And it can give a short kind of like, it's the same because they both say services and, you know, whatever it is. And so you can get sort of a quantitative matching and then a like qualitative explanation with two relatively cheap models once you've narrowed. that down. Yeah. Yeah. Great stuff. Okay.

Great stuff. We are just psyched about it. I was like, I can watch all of these. This one's fun. It's kind of hard to show off. But essentially, if you think of having your entire application and so you have going back to the to do list and say that's one part of your much larger app, you can build additional distractions around that so that if on your landing page, someone,

one comes up and either has a command bar, like a command K, or other demos, or sorry, other parts, omnibar, any place to put in text. And they start typing in the dictate or voice or something. They can't remember what the to-do app was called. And you notice this is to-do-do-do-and-I-tied-to-d-do, and I typed in to-do. It's able to go in and match against the tool, and then you could type additional things where you'd say go to the to-do app and then take this action. Like I would know in my head, oh yeah, there was that pull request and then take all the pull requests and mark those as little priority. And so if you think of this multi-step multi-step classification where you can build your smallest tools and then build another

abstraction of Jev around them where it can pick which of the smallest tools to use, it kind of goes back if people have set up MCPs where there's abstractions around them I think Executor and others where it picks which MCP to use there's the tool itself

and then there's a layer where you're picking which tool to use and you can build that as many layers of that as you want that's what this is kind of talking about it's difficult to show off but just imagine saying go to the two app do this and yeah

what I would say is like Jev is a router is maybe like what's helpful. Jen is a router. I love it. Very, very fast router. And so, you know, this use case, you're saying like, as I have text, route to the right tool.

And then if I can infer from the text, the job to do in the tool, I might as well infer that and like route as far down the user journey as I can infer from these like structured options. Other practical use case and things, a lot of people start to build these like jev powered coding harnesses where it's like, you. given the first input, let's route to the right model, let's pick the right tool, all this kind of stuff. And so you can imagine very fast configuration, very fast routing, very fast, like I don't just have to take the first step and then think and take the next step. I can actually just build the chain and then execute it.

And so I think these decision models like Jeff, if they remain fast and cheap, can be a new way to think about navigation of your user experience, whether that's a kind of like front-end experience or a dev tool or whatever it might be. Yeah. Even search to a certain extent of search for a long time has had a lot of really expensive options out there. And I don't know how far you could push this into replacing certain search engines. But yeah. You know, I gave an example when I did this, like, I had 4,000 YouTube comments and I was trying to do live search through this. And I want to demystify some stuff for folks because it actually isn't super fast to like search over all of these results and say like yes, no, this is a good result and rank it.

But you can like cluster like groups of like 30 or 50 or 60 and score them and rank them. And that is a much faster way to build quote unquote real time search with Jev. So, you know, for folks, again, I just want to like give people a little. little bit more specifics about what you're seeing on the timeline, which is you do have to think about how you're going to architect around Jev. It is fast. It is efficient, but it's not like completely latency-free. And so you may need to go look around, see how people have done search. And then if you find one you like, go into GitHub, figure out how they actually made that search super performant, because I've seen a couple different strategies out there to how to do this.

Yeah. And caching and all that stuff. All the, all the hacking. Yeah. It's easy when everything's locally cached and just to go over the full corporate. Let's talk about games because this is something I haven't covered yet, but I think is a really cool demonstration of some of the ideas inside Jev. Yeah, I built this one because I wanted to show off. Jev versus a traditional model. I picked a free model on OpenRouter. This is like a low reasoning model. I think it's Akimi or something, I'm not sure. And for people that aren't watching, it's playing chess. Yes.

This is playing a game of chess, and Jev are white pieces, the LLM are the black pieces. And here, this is set so that Jev is able to think through all the possible moves, and then it ranks the highest three moves, and then it thinks through all the next possible moves there. and based on that two-step reasoning, picks the best next move. And so kind of like routering, it's kind of like sputtering into the potential best moves and then coming back to what the next best move would be.

And it's still doing this in sub one second times and using AI inference. Whereas if you pass the same data set to an LLM, the amount of reasoning that takes, This is playing Blitz Chess, so there's a one-minute timer. Jev would be able to complete this entire game and beat it. And the LLMs would eventually get there. They'd be much more expensive and they might not even perform better.

Because in scenarios like this where it's focused on routing and the next best thing to happen, the amount of data that you can crawl into and find the next best move can kind of balance out between these two. Unless you're getting to like really super high level like creative chess, where you're trying to stump someone by fainting certain moves or whatever, but I'm not it. I'm decent at chess. I can beat my kids. Perfect.

For now. Sometimes. For now. And if you look at the, it has some benchmarks over here that Jev was 10 times faster in the average move. And it was four times cheaper. And this was a very space money alphazon, low reasoning. So with very, very tiny context windows. Yeah.

What I think this demonstrates is, one, there are scenarios in which speed is a, is an advantage. This is putting in the context of a game, but it's very obvious how much. I mean, 10x faster is not incremental. And so it really does unlock this concept of real time. And I do think about going back to when I was a. a young product manager and did a lot of like conversion testing and in growth work and like speed did matter in a lot of user experiences, especially consumer on what one. And so I do think, you know, we hear a lot on the coding side about model optimization from a cost perspective, but I think we're

going to start to see like latency wars really heat up as well, which is like the faster the response, the more like more magic you can unlock that feels like an if else statement the the better. And so I think this is showing one where speed really matters. And then the other thing it's demonstrating is this like multi-step routing that we talked about where you can think through turn one, turn two, turn three. And then because it's 10 times faster, you could apply another model on top at that point, a more intelligent model at that point with kind of like a narrowed set of strategies

and probably make a stronger decision faster. Right. And I think this also demonstrates the, again, to the concept of router, there are a limited set of options in chess and other games. There's a lot of options, but it's a limited set. And you can pick from the most likely set and narrow it down and optimize cost and everything.

Data structures and your APIs are all the same. Like they all have a limited set of things you can do, even though it might seem infinite to you because you never know what somebody's going to type or do in your application. This handles those, when you have those scenarios, it's time to look for a Jev first, I would say, whereas the LLMs are much more, I don't, I have no idea what I want the action to be,

or I don't even know if I want it to take an action or things where it's much more brainstorming and creative and finding what you want to do. But once the actions are set and you have that limited scope, that's where a decision model can really excel, at least from what I've experienced so far. And I think this can demystify, too, some of the other kind of fun, eye-popping use cases

that I have seen like browser use, right? If you look at a website, navigating it with a mouse feels like an image, set of possibilities, right? You can, like, move it anywhere. Any pixel on the screen is where your mouse can go and click. But if you actually look at the DOM, there's probably like 10 clickable buttons on a page. And so if you can very quickly say there's 10 clickable buttons on this page, which one do I click? Like, which one do I want to click? That, that demystifies that concept of browser use through Jev because you're really just taking a limited set of interactions available to us and our human brains. It looks like an infinite camera. of pixels, but to a model like this, it can just be a set of decisions. You know, I've also seen on gaming, you know, Jev plays Tetris. It's because Tetris is four options of your shape. And then like 10 options left or right. That's all you, that's all you have. And so, you know, these like horizontal scroller games are just you can like jump, you can walk,

you can do, you know, any of those things. And so because these decisions can happen so fast, if the interaction of your app, even if it's a game that feels complex or a web app that feels complex, it probably can be distilled down to a dozen choices. And then you can score those choices. And because Jeff is so fast,

it almost feels instantaneous to chain those together. Yeah, I'd almost say in, like you mentioned in a game in a platformer or in a 3D game, you might be thinking about all the things the character could do, but the input is actually the controller.

And so the limited amount of options you have are all the buttons you could press, how you could press in various ways, which is still a lot. But there's a very constrained input there. And whenever you have constrained inputs that are interacting with apps, it's, Jev is a good thing to reach for. And to contrast the DOM example, if you think of taking a screenshot of a web page and saying what on the screenshot, or what on our dot, what on our, what on our, product or website might be confusing to a user.

That's not a Jev thing. That's an LLM that can look through an image. It can compare it to all the images and products in the past. It can say it can look through the layouts and the colors and the contrasts and all the sorts of things and come up with a discussion for you of here's here's some points that might be because the fonts are gnarly or whatever. And so that's where trying to delineate between the two when to go which way. As AI adoption grows, so do your company's security risks and requirements. New frameworks, audits, and vendors keep piling on, but your team isn't getting any bigger.

Most compliance tools promise automation, but you're really stuck doing all the work by hand. Chasing screenshots, validated controls, and playing catch-up when audit season comes around again. Vanta works differently. Their agentic trust platform is built to scale with you, not slow you down. With over 1,400 automated tests across 400 plus integrations, Vanta collects evidence and monitors your controls year-round, cutting your audit prep by 82%. Plus, you can access the Vanta agent everywhere you work, whether your team is in Vanta every day or living in tools like Claude or Cursor. Want proof? Over 16,000 fast-growing companies like Ramp, Writer, and Harvey. Save time and resources while scaling fast with Vanta. My listeners can get $1,000 off by going to vanta.com slash how I-A-I-I.

That's V-A-N-T-A-com slash how I-A-I for $1,000 off. Okay, let's do one. Do you have a last favorite one? I love to. I know there's so many. This is super fast. Okay. On Wikipedia, there's every.

page leads to philosophy, if you've ever done that. So you can put in any person in the world or any topic, so like the Braun James or whatever, run to philosophy. So you give it an end goal and it looks at the entire page and it's going through the Wikipedia API and mapping that route and finding from the person to philosophy. So mapping routes as well, like giving it a final destination and it can crawl its way there. This one is a favorite of mine because it represents represents, if you think of a workplace where you have multiple agents that could be doing multiple, they could all have their own tasks and avoiding collisions so that they never end up on the same space. So this is going to, the job brief is deliver the fragile blue bin to staging, bring medicine to the board, and inspect the spill.

Then you have three people that you can assign to. So this can move one tick and you can see at each step it's going to be able to control these. different agents and make sure that they get to their specific task as quickly as possible and that they never collide or interfere with the other agents. So if you think of this in like agentic programming or any sort of parallel work you'd be doing to ensure that you have things running and you want to like analyze to help steer them in different directions so they never collide or disrupt each other or touch the same files or whatever. You can ask Jeff

to take in like every single, it's fine if it goes in every single step and it checks. It doesn't have to plan out the entire thing ahead of time. It can do just in time. Obviously, planning out ahead of time with a decent route probably helps it out. But I think as we get more and more parallel agents and swarms and whatever words you want to use for that, this concept will become even more important. Well, and what, you know, this has made me think is Jev unlocks efficient inefficiency, which is like, it is kind of like inefficient to map out all possible routes, rank them, double check that they're not going to collide.

Like that is actually pretty inefficient. It's an effective way to solve the problem. Like it's an accurate way to solve the problem, but it feels inefficient. And in the past, we were just like kind of like tossing these problems in the past last week. We were tossing these like big brainy LLMs and being like, think really hard. Just think really hard and come back with the plan. And instead here, you can really, very quickly evaluate all those options, stack rank them, and go through the sort of like inefficient universe of options and come to the right,

kind of like the right, right conclusions. This is why I'm like very excited about this model because it's all these things that in the past, I think with maybe with regular regular LLMs, you know, we're trying to do this delineation. it feels like you can do things you never imagined before, right? Like I can create a 3D video game. I can like generate images out of words. Like these LLMs are generative and they allow you to create things you never be able to create before. Jeff, I feel like allows me to do all the things I wanted to do, but they felt too

cost ineffective, too slow and not really worth it. And there are a lot more, a lot more things for me on that pile that there are. that there are on my like 3D gaming idea pile. And so that's why maybe as a builder I'm so excited about this particular model. Yeah, I love that. I 100% agree. It's even like I have so many APIs and so many apps and so many things available. How could I?

And the exploring with this is quick and the turnaround is quick. And it's exciting to see like you feel progress as you're working with it. Whereas with generative. And I feel you almost because if you're a developer, you understand the APIs, you understand the reasoning, you understand classification, whereas with the LLMs, a lot of it is guessing. And if something goes wrong, you need to think of a different sentence to say with decision models. You can think the decisions have made. You're like, oh, well, let's take this different route. It feels much more like programming, to be honest, where you have, I feel much more in control when I see something goes wrong because I know all.

of its options and I know what I want. And it's fun. It's just so fun. It's so fun. And I love that you say it feels like program. It feels like the smartest function. Like it just feels like a function that has a lot of intelligence built into it. But it's still a function. Like it's still input specific things and I get out very specific things. I can just reason with it a lot more. Okay. So if you if you've stayed with this podcast this long, it's just John Claire, two Jeff boys, excited about, excited about decision, decision models. John, okay, I'm going to give you, because we're at example, 9 of 23. I'm going to give you one last chance to pick a favorite from your remaining demos before we do lightning around and kind of get you, get you out of here.

This one's really fun to me because it's a presentation coach. And so I clicked the microphone, I'd start talking. And as I was talking, I'd have a list of bullet points in either a podcast or something that I want to make sure I cover or an interview or anything. And as I start talking, it listens through everything I say and then starts checking the boxes to make sure that I covered everything. And so let's say this real time, as someone who likes to go off on tangents, you know, I'll get passionate about something. I'll go a different way. you're under the clock, and Jeff can be your stay on track. And I think it's kind of almost a life coach of like,

you need to make sure you hit all these things. I'm going to watch every single thing you do, every single word you say, and make sure that, and you could start doing like red flashing lights of, you know, you have limited time left. You still need to say all this sort of stuff. So this one speaks to me as a teacher, presenter, workshop giver. I love this.

I just did Lenny's, letting his summit recently and gave the kind of like opening talk. And I hate those monitor because they're like, hey, you can only have three bullet points on here. I'm like, that's fine, whatever. But then I like to ramble. I like to work the crowd. I don't know if I've said that that bullet point or not. And so imagine you speaker notes could check off as you go and then even progress your slide for you. Like you're done here. Let's move on. And does this just take? I do have a question about your real-time, again, because I like to make this very tactical for people, and I'm curious how you approach this. On these real-time voice ones, are you constantly putting in the long stream of text to the moment?

Or are you doing like a moving window of snippets? I'm just curious how you're doing it. It's doing a, I can't remember exactly the logic, but it is analyzing what you've said up until then. It is like concatenating each word until it reaches a point where it can take an action. And it essentially transforms that into the payload that it's going to send over the function. And so that one is, it's stored in the, it's in the history. Like you could have undo and whatever because they turn unstructured into structured data. But yeah, it's a window which then gets chopped off and then you start your new window of the text. This is a good moment to tell people exactly how much this thing costs.

I feel like last time I checked it was like four cents per million input tokens and nothing for output tokens unless you like me are using VersaI-Gaithway right now. And then it's completely free. Completely free. So again, like I can talk. I can yap, but I don't know if I can yap a million tokens quite yet. Challenge accepted, right? Yeah, exactly. It's pretty amazing.

Well, these have been such good use cases. We've had a lot of real-time voice use cases, sort of like pairing deduplication use cases, routing and navigation and like multi-turn, multi-path kind of scoring, which I think is awesome. And then this one, which is like just keeping you on track, did you do the things you said you would do while you're talking? Did you present the point? The monitor, right? The thing that's always watching. and making sure you're

Love it. Let's double check. Did we do it? We covered what Jev is. Why speed? What it costs? Does the app stay in charge? Where does it fall short?

Did we talk about where it falls short? A little bit about where you'd want to pull in an LLM. But have you ever tried to throw Jeb at something and didn't do a good job? Not yet. I don't think I've spent enough time since it's been out for a matter of days. really analyzing how smart it is versus what I would say there's times it falls short where I add another layer of classification in. So I do multi-step. And I've seen people critique it where they only do a single pass.

And they think that's not good enough. I'm not going to use this anymore. When I run into those scenarios, I'll do multi-pass where I'll do one classification layer and then another and then another. So I think it might be a disconnect there between some people who try it out first and they're like, this isn't good enough at classification or they don't like define the classifications or they don't have strong enough APIs or. So I think just when it falls short, because it's fast and free, don't worry about adding just another layer. And then over time, you could compress those down into a single layer as you get more and more data that reinforces what the exact form. flows are that you need to go from from step to step. Again, it's all so brand new. Yeah. And it's getting smarter. Like they said, these models are, you know, there's demos will

ever be. Yep. And even these models will get much, much smarter. Well, and I'll give folks one other tactical piece of advice, which is you may use a specific type of Jav call, like a, like a, like a noel. And you may say, I think this is a yes, no question. And you may run it. You may be like, actually, it's a choice question or actually it's a score question. And so you can also test different ways to ask for different types of decisions to get it more accurate. So I've had that experience a couple times where like my first concept of how I would classify or make a decision was just ultimately like not the right one or had to layer one and then the other so that something is for people to look at. Okay. As you said, we could just, I could go on truly forever about this, but I won't. I will spare our

audience. Let's get to lightning round and then you and I will go back to just classifying all the things for zero dollars. Sure. It has been, I don't know, I was reflecting. It's been almost a year since we last chatted. It was so cute. We were talking about Claude Code and like aliases to dangerously skip permissions and just like babes in the wood and now life is completely different. You know, what, what are you other than Jeb? What are you really into these days? What has changed for you in terms of? AI engineering. What are the big, big moves in the last year?

Some of the funniest stuff recently with the latest models, when it gets into multimedia and you're working with video and 3D and everything, is buying extra hard drives and distributing work between Mac minis and such, which is not a problem, like, it's a problem. It's a problem. And, like, having AI is, like, watch my disk space on my laptop. being like, bro, it's time to hand some over to the SSD that, anyway. Hold on. I have to pause. You and I are exactly the same person because truly yesterday,

I put the SSD in. I have a pinned codex chat and the pinned codex chat is clear up disk space and it has two skills in it I call it a regular basis. One is move media files onto the hard drive. So it's like it takes it finds all the places I store all my MP4s and it goes finds them all different places and then it organizes them on my SSD. And the other one is it goes like in the crevices of my computer and finds work trees that are like still running local servers and like clear like bus some of the cash clears. I like I was at a client's office and they saw my computer like freeze up and you're out of memory and it's like they were like can I buy you like a new computer? I was like do, this is a new computer.

This is just the problem. And then I set up, I kill the open clause, RIP, but then I set up the stack of Mac minis as remote codex machines to run all my, all my code on because I, it is a real problem. It's a problem. I'm sure we'll all just, we'll go to the cloud, right? I like all the stuff, but. I like local.

I like working on. desktop apps. Jev is going to unlock so many things, though, with, like, removing UI that, I don't know, pushing UI later and stuff, but I don't know. I would say the other thing, the most fascinating thing to me right now are the paradigms around you've seen Grockbot at all and similar tools where the traditionally as developers we think in terms of projects and file structure and code. and files and everything.

And assigning agents roles and tasks where you don't worry about any of that stuff and you just have a list of agents and then you can bring them into a group chat where you have, oh, I have this task

and I wanted this research guy and this little code review person and this other agent and you bring them all into a group. It's kind of inverted the concept of agents being the top layer and now they care about the structures and the code and get and versioning and all that. I think it's a really big turning point in like our relationships with our

tools as we stop thinking about everything we thought about before other than buying more hard disks for agents to use. Yeah. I completely agree. And I say like, you know, in some ways. In some ways, managers are really at risk, like the job that you did, people don't want you to. In some ways, this concept of like org design and role design and like you can put those skills to use when crafting your agents, which is why I am currently running no joke 40

grok bots right now. Wow. Across the business, I will do an updated episode on all of them. I covered seven in my last episode. So we'll get to the remainder in another one. John, this has been so, so fun. Jeff has, I agree, you've been so, so fun. So I really appreciate you coming on, showing us some of your examples.

Where can we find you and how can we be helpful to you? Oh, my next big stop, I don't know for sharing the screen still, about mega. orgia. Is me and Theo and Kent and Angie. We're all doing this big thing in a couple of months. If people want to check that out, all of my effort for the next couple of months is going into this. So I'm putting everything else on pause just to make this the best experience possible. And I'm going to hype you up a little bit. It is, I'm going to say it's a kind of like course

slash program slash workshop to teach you how to turn agents. I'm going to read the landing page, turn agents into leverage. You can reason about price debug and ship. So be an incredible AI engineer. Also, hottest website on the internet right now. I saw it the other day and it's just, it's real lovely. Love it. So check it out. Check out John. Check out mega.com. Really cool program. And John, we'll have you back maybe in another year. And who knows? We won't have hardware or hardware at all. Maybe we'll be carrying around our little meta-buse key chains and talking to a little fuzzy Yeti who's doing all our coding for us. Be embedded into your brain at that point, I think. Sounds good. The year is too far away. All right. Thanks for joining. How I'm

AI again. Thanks so much for watching. If you enjoyed this 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 how IAIipod.com. See you next time.