Jev Explained: Intelligence infused 'If-Statements' for Your Code | Allie form TypeSafe AI

We are enabling the builders. We are enabling the creatives of the world to build new things. Jev is just like the direct line. It's machine-to-machine intelligence. You just talked about more code, pushing more applications, pushing more products. Where do you see Jev in that equation? I see Jev as the intelligence per dollar. Intelligence per dollar is a cornerstone of what, at Type Safe, is our goal. We want it to be as cheap as possible because that's what enables everybody to take you to. I am Ali Lobs. I am the DevRel at Type Safe. I am our DevRel department. Okay. I am both our lead DevRel and our most junior DevRel. So, yeah. It has been, I mean, this is what the word unprecedented is made for, is defined for. It has been, it has been absolutely mind-blowing. It is just one of the most bizarre experiences of my entire life and I think probably everyone at the company could say the same thing. It has been, you know, we've been working on this a long time and we've been so close to this technology for a long time. And, you know, when you're really close to something, you get very familiar with it and it becomes very normalized in your mind and it's like, you remember, we remember, we know that this is like a really cool thing and something that doesn't exist yet and that it could be, that it's really, really useful in a lot of ways. But like, one, are we going to be able to convince people or the fear that can sometimes be in the back of the head is like, is it just actually not as good as we think it is? Like, you know what I mean? Is, are people going to be like, why would I want this? And so, for it to, for it to have caught on like this is so, it's so exciting and it is just mind-blowing. I'm like, I'm at, I'm at a, I'm at a loss for words to describe it because it is just something that is so outside of what we could have predicted. You know, something I like to say is, you know, how do we, had we prepared for a launch this big? That would have been an irresponsible use of resources. Like, that anyone would be like, no, like, don't prepare for one of the largest launches of all time. Like, it's just like, that would be bad planning. Yeah. And yet, here we are. I'm incredibly proud of like the team and especially our platform team and how things like things are online are so people are using it. It's in production and it's staying up for such a so much larger than we could have possibly anticipated. That's like such a huge achievement. And so like, shout out to our platform team big. What do you think besides obviously scaling the infrastructure, one of the biggest challenges that you're facing right now? Oh, biggest challenge we're facing right now. So there's like, there's a couple ways to answer that. There's from the personal perspective, there is the like, making sure that everyone's taken care of themselves as we are just like sprinting like crazy, you know what I mean? And I feel very, I feel very fortunate to be under the leadership that we're under, you know, Diego has multiple times in this past week and a half. Like, we do, like, we do stand-ups every morning. We have an all hands every week and we do like a lunch discussion every day. And so, and multiple times he's talked about like, the importance of like balance and making sure that people are taken care of each other, you know, and that we're taken care of ourselves. Because while we do need to like put in as much as we can right now to meet this wild time that we're in, like, he very much understands that like, if you put in too much, your productivity will decrease and it's worse than if you didn't put in as much. And so, like, we're, we're all like pushing as hard as we can, but also being reminded by ourselves and by, you know, everyone else to be like, hey, it's okay. Like, you can go home, like, this is like a, we'll get back to this part in the morning, you know what I mean? And it's like that, that's really cool. So that is, that is one of the challenges. The other challenge is, you know, it's for, I would say for me in DevRel, the challenge is I want, I want to please everybody. I want to answer every message. I want to answer every email. I want to answer every discord message request. I want to be there and discord. I want to be talking to the builders. I want to be helping every project and it's just impossible right now. And it is, it's impossible to hit them all. And so the challenge is figuring out like exactly where to focus. And it's, it's a, it's what we call, you know, it's what they sometimes call a champagne problem, right? It's like, it's a, it's a, it's a problem because of such great success. So it's not like the saddest problem. But, but we really, really, really want to do right by our users. We are a very like, like builders oriented company. We are, this is very all along. This, our product has been, we've wanted to be for the people. Right? This is for individual builders to build with. It is also hugely valuable to companies. And we've got that, you know, we've, we've got our sales pipeline going and like we're, you know, working with a lot of different companies right now. And you're, you're seeing us put into production from more and more places. And you know, that's, that's going great. But what we want to never lose side of is that we are, we want this to be for each individual software engineer. We want this to be a thing that everybody can use. And that requires like supporting individuals. We give $5 a month in free credits to all users. We, we, we have so far. And that is like a, companies don't need that, right? Companies, they can just like, whatever, throw $100,000, $1,000, whatever it is that they need into their account and just like get going and experimenting. But individual builders, sometimes even $5 for plenty of people in the world that we like too much. So we want you to be able to get in and build with it right away. And at our prices, $5 is a lot of tokens. Like you can do a lot. You can put something live in production on $5 a month for a lot of like indie projects. So that is such a, it's such a focus for us that we want to, we want to enable the builders of the world to get to like use this. Yeah, I think my next question, just because there's, there's so many people. Obviously, having heard about Jeff, we just, you just gave an amazing speech of all like all the tips of how to use it. Can you explain for somebody who's just heard about the hype on X? What exactly is the difference? Like, why does it such a big change? The succinct version is, Jeff is a way of accessing the intelligence that is within these, these AI's, these systems, you know, these, these models that we, in the present day call AI, like there's intelligence within them. And LLM's provided a way of tapping into that intelligence. But it provides a way of tapping into it in a very narrow way only through the context of like generative text output that has been reinforced, you know, the three reinforced learning has been designed to be like pleasing output to do to follow instructions. And this gives an entirely different way of accessing that same wealth of information, that wealth of intelligence inside these models, but through something that is far more expansive and is for machine-to-machine communication. You want your code to be able to use intelligence within its software. And that's not what LLM's were made for. They were made to chat. They were made to produce text. Yet we've bent them over backwards to get them to output text that represents a human user interface to a computer. My mind is blown when I think about the way we're using LLM's where we're like taking, we're taking the intelligence, we're turning it into text to interact with an interface that was already to access something in a computer. We turn it into a human interface. Then we get the LLM's to output text to be able to interact with the human interface. And then it goes back into machine language. And it's like, we started in machine space and we ended in machine space. Yet we did this extremely expensive bridge through like text and through like a human interface layer. Why? If it's from machine to machine, that's all that you actually need to get to. This is an incredibly expensive way of doing it. And Jev is just like the direct line. It is just its machine to machine intelligence. That's very interesting. It reminds me a lot of that discussion. I think a lot of, I speak to a lot of other dev worlds as well as engineers, founders in the valley. Some of them say that at some point we will just be writing mathematics at any point. So the translation layer will actually be a new programming language that is up solely based on linear algebra or yeah, just I could see that. I'm not as familiar with like the actual like the mathematics like within these things. But I do think I do think that is one of like the really interesting things about this like AI revolution that we're in is it's not like oh all the things that the AI can do. I like I like that it's a different approach to computing. You know what I mean? It's like it's not like you said it's like it's it's linear algebra instead of like binary arithmetic. It's such a it's such a shift in at the very lowest level of what computing means. And I think I think we've only like scratched the surface. of what that will actually mean in the long term, even in the medium term. - Yeah, beautifully said. Let's turn around a little bit more towards the product again. Can you give people a practical grasp of the possibilities that are now enabled through Jeff, like some use cases, some tips and tricks that you just mentioned for? - Yeah, a practical grasp. This is like, this is the hardest part about talking about Jeff and teaching people Jeff, is that it is such a, it's a category creation product. That's a term we used a lot leading up to it, 'cause that drives a lot of the strategic thinking about how to present this to the world. It is a category creator. A category creator can be very challenging to get people to wrap their heads around, 'cause it's not just coming in as a better version of something, we're not an LLM. We're not a better version of LLMs, like we sit in harmony with LLMs. We replace certain things that I think we're using LLMs to do in a really weird backwards and correct way, but we don't replace LLMs on the whole. So there are simple use cases to think about. The simplest way to think about where Jeff fits in practically is thinking about any place where you're using an LLM to do classification. Some people think of Jeff as a classification model. Fine, that is one of the ways it can be used. It is one of the things that it does. So if you have a classification shaped problem in your software and you're using LLMs to do it, Jeff is almost certainly a better drop-in for that, right? If it's categorizing emails, if it is categorizing financial transactions, if it's categorizing these, or if you're scoring things, you're having an LLM decide like how good, how bad, how much of X, Y or Z, Jeff fits into that really easily. And so those are the first use cases that come to mind, and those are the ones where people that are like putting us into production really quickly. It's because they already had problems. Like I call them Jeff-shaped problems. Like they already existed. They were maybe using a more expensive, slow, less effective tool for that, and now there's a better tool, great, drop it in. The other case though that is far more exciting to me is that true category creation side of things. And that is where because Jeff is so fast and because it is so inexpensive, and you can put Jeff inside the software loop. Like you can put, we call the questions that you send to Jeff, the decisions that it's making. We call them primitives on purpose because we want people to think of them as the lowest level components in your code. Some people, it's helpful to describe Jeff as a smart if statement or a smart switch statement is like an intelligence infused logic gate. So when you think of it like that and you really think about, oh, at the lowest level, anywhere where there is logic switching happening in my code, what if I had intelligence there instead? Maybe you're doing a complicated regular expression. Maybe you haven't even built it because you conceived of what the code would need to do and you're like, yeah, but that would be way too hard to like figure out if blank, therefore do X or Y or Z. But maybe Jeff can now unlock that. That it becomes, it's like a, we call it a primitive because I think in a lot of these use cases, it will be a truly new path of code that couldn't exist before and now because it's there, you can. I'm going to drill down a little bit deeper. I hope you okay with that. Just because it's been amazing to watch from the outside on X, Twitter, all the things that people put up. But then often enough, you see people building a snake game or building like some other thing that we've seen before. What are some of the use cases that have really surprised you and your team? You know, it's funny thing is people ask me this and I should have better answers off the top of my head. There have been so many in my brains like so fuzzy from the, the, the, the, the further. I'm also drawn to the very like, to the toys. Like my, what like stands out most to me, what I remember is the like silly experiments. People have been making Jeff talk using it to actually generate text by like looping and asking Jeff, like what words should come next? What words should come next? It's a, it's hilarious because it's like bending Jeff backwards to become like an LLM again. And it has a, like a personality to it that's very silly. Like we have it hooked up in our Slack. Like there was an open source code base of someone made, but now we can talk to it in our Slack. And we all think it's very like adorable and charming and like dumb and it's fun. - Do you think Jeff would have pink hair as well? - Oh, that's a good question. I, I feel like Jeff, like talks like a toddler kind of. It, it, it, it talks like Rocky from, from, you know, from a project Hail Mary, right? We, someone pointed out that out there, like it, it's like Rocky. It like, it speaks so like, yes, no, bad, bad, bad. You know, like, it, like literally we'll do that. And that is, like, so those are the things that like stick with me at the top of my head. And I think there's a reason, like it, that's also what keeps me going. Like with this wild fervor that we're like currently in, it's just like, because like I said earlier, you know, we, we are enabling the builders. We are enabling like the creatives of the world to build new things. And so those are what just like energized me the most. But it's really cool seeing things on all sides of the spectrum. And I think, I think one that has been really interesting to see is like real-time analysis of like speech, like people talking. I think there's a lot of potential in this space and have already seen some like really interesting projects come out, 11 labs put out, like 11 labs developers put out a, a video of like using it to, using their like real-time transcription, like feeding into Jeff, you can do like analysis of what you're saying. And it can sort of be like a speech coach in real time. I built a teleprompter project that does something very similar, shows you what like bullet points you have it. Like you could have it in a teleprompter right now of like, oh, all the questions that you wanted to ask. And it doesn't matter what order. And they'll just disappear once you've asked them because it's just, Jeff will be like, oh yeah, that question was asked and you can just call it over and over and over and over again in real time. It's those real-time use cases that I'm most excited about and that I think is really gonna be, that's gonna be really the cool stuff. - So you say like, for example, generate a few I, is that something? - Generative UI, oh, okay, there's a good one. Generative UI I think is a, that's a super cool one. And that's something that we, a lot of people's projects, I can name like a demo that I made on our lead up to launch like when we were experimenting with different demos for our launch. I made like a lot of tiny little experiments. And so a lot of what people have built been like better versions, more complete versions of stuff I experimented with. Generative UI never even thought of, like never even went into. And I have a personal belief that Generative UI is going to be a huge part of how we interact with computers in the future. I think that more and more what people interact with is just gonna be tailored to exactly what they need to their style, to their skill level, like just to the personal person. And why show a whole bunch of options that you don't need if the AI, if the system can be smart enough infused with intelligence to basically be like, well, you probably want one of these three and can quickly figure out the one that you do want. And that will make things so much better, frankly. Those of us that use computers all day long, we can take it for granted that we can use them effectively, that we're proficient in them. I try to really stay in touch with people that are like completely outside of the tech sphere, right? Even the things that we think of as like very, like public friendly, like, you know, modern smartphones, like things that are like made to be, you know, hardcore tech people will be annoyed, 'cause it's like, oh, you can't hack into it, and it's so simplified or whatever. Even then, these things are hard for people to use. They are people get so upset when there's like an OS update and now all the buttons have changed. Us techies, we can learn it really quick. There are people out there that like, that's the most dreadful part of their years when they're forced to upgrade their like phone OS and they have to like figure it out. This isn't just like your grandma. It's a lot of people. It's a lot more people than people in the tech bubble can oftentimes think, it can remember. So the world of generative UI and like making it so that it can truly be tailored to people, I think is going to increase the accessibility of software and everything that a computer can do to so many people. And I think that is a really, really like energizing thought. I'm a big proponent of like accessibility and like lowering the bar of entry into anything that I love doing. I want anyone in the world to be able to get started in that thing. I have no interest in gatekeeping. I have no interest in being like, oh no, you need to be like, you have to have like studied or done this to be able to like get into this. It's like, no, if you can just do it right away, that's better. That's better. It makes more creatives in the world. So yeah. Yeah. No, a wonderful rant. Let's talk maybe a segue on that. What do you think Jeff could, which role do you see Jeff playing in this entire builder economy that we are now experiencing right? You just talked about non-technical people being able to access UI better. Very likely they also, I mean, they're pushing right now, also more code, pushing more applications, pushing more products. Where do you see Jeff in that equation? I see Jeff as eventually. it will be a tool in every engineer's tool belt. It will be, I like to compare Jeff to Docker and AWS Lambda, serverless function execution. Both of these are products that I see as, when they first came out, they were hard to learn. It was a totally different way of thinking and a lot of people would be like, this isn't worth the trouble. Like, why do I want to learn containerization, how to spin up Docker? I already know how to set up a server. We've been doing it for decades. Like, what's the big deal? Or serverless function execution, even weirder. Why, what do you mean? Like, I'm going to run a JavaScript function without a server, like very bizarre. But now, you can't be in the infrastructure space without being super familiar with both of these, or at least like knowing that they're available tools and likely the right tool for a whole lot of jobs. And I think Jeff will be that for people in the future. Like, you will just always know that, yeah, of course you have a decision model. Like, you're going to use that. Like, like a regular expression. Like, people know that regular expressions exist. It's the wrong tool for a lot of jobs. It's very much the correct tool for some jobs. I think Jeff is the same way, you know. It is the wrong tool for plenty of jobs. When it is the right tool, it is the right tool. And it just needs to be there for everybody. And, you know, that's a big part of why we also just have such a focus on the intelligence per dollar. Like, intelligence per dollar is a cornerstone of what, at type safe, is like our goal. We want it to be as cheap as possible. Measuring intelligence. Whatever the ephemeral unit of measurement of intelligence is divided by dollars. We want that to be as close to zero as possible. Because that is what enables people. That's what enables everybody to use it. Beautiful. Yeah, you just talked about the learning curve on Jeff. You said it's not the tool for everything. Can you give us, like you just did before, some practical tips? Like, what's your piece of advice as somebody who's used the tool for much longer than most of us? How do we use it best? It's-- that's a really good question. I think Jeff is so much easier for people that have been engineers for a long time to grasp. But because people that have been programming, especially for, you know, decades, before the advent of AI, before the advent of the first time you could ever even ask ChatGPT to write a little bit of code for you. We had to design systems by hand. The old artisanal crafted by hand systems. And Jeff, for people that are already familiar with that way of thinking about software, Jeff, it's in very easily. It's like, oh, I get it. Because you've designed systems for so long. But we have this incredible world of people that are getting into creating software in the era of AI agents, in the era of coding agents. And I think that's wonderful. I have-- there's a whole 'nother-- Yeah. Passionate, romantic conversation to go on about why I think that's a wonderful, wonderful thing. So I would almost say, like, to the old timers like me, I don't really need to explain how to think about you. You're going to get it, right? You're going to understand. It's the people that are less familiar with that kind of systems engineering that I really want to-- that I really want to be able to speak to and have them understand. And there's a couple different approaches. One is tell your AI agent to look at our documentation. There are some great skills that the community has made. There's a couple of them that are both called Jevify. It's basically like a brainstorming skill. It's like, look at my project and understand this about Jev and let's brainstorm how it could fit into it. That can work really, really well. We were even going to make an official brainstorming skill as part of our skill package. We didn't get around to it before launch. And it's awesome to see that the community immediately did that. It's almost nothing that we need to make now because everyone's making it all that's so cool. But that's also a little bit hand-wavy. Oh, just tell your agent to do it. I do want people to be able to understand it. And it's thinking about it as the lowest level bits of a system is the most important part about wrapping your head around it. It is anything you can do to wrap your head around the idea of smart logic gates or smart if statements. And if that's not something that you're immersed in, it can be harder to visualize it. But you want to break it down into that-- into that, how does the code flow through your system? What is actually happening? Imagine the branching paths in like what can happen. Even if you've completely vibed it up and you haven't looked at the code, look at the behavior of your application. You click on a button, you do a dropdown, you do some user input. And this happens instead of that happening. Those are decision points. And if those decision points are deterministic, as most decision points are in software, it's just like comparing a true and false or comparing whatever. Which case? Yeah, that's happening like a bazillion D times of the microsecond. That's great. But then do you need to branch on something that requires some intelligence? Do you want different behavior based on something that is harder to do with just arithmetic? Something where it might be user input, semantic input, speech or typing. Is it something that is-- if you were to try to sketch out a flow chart of behavior, and for one of them, you're like, well, this is more like a gut feeling, right? I can't define it mathematically. I can be like, what if it's senile? It's like this. If it's this vibe, it should do that. And if it's this vibe, it should do that. That's probably a spot where Jeff could work, right? It's that when do you need probabilistic vibes based logic switching? And that's where to think about Jeff. Yeah, I really liked the other example as well that you taught me about earlier, which was the idea of speculative prompting. Yes. Yes. Speculative prompting is, it's one of the first things I ever wrote about in our documentation. Exactly the very, very, very first project I built with Jeff. I used that technique, like it was intuitive to me, but I also understood that it's probably wasn't going to be intuitive to people. It's not intuitive decoding agents. They don't do it by default. It seems inefficient, frankly, if you think about it from like-- In most traditional uses of APIs and stuff, it's not how you would use them. You would call the minimum amount you need. You don't want to do extra work before you need it. But-- and you're right, so this is a very unintuitive way of thinking about it, but speculative prompting is when you-- anything that you might need to know about the thing, about the state that you're evaluating, which is like the input to Jeff, ask it all at the same time, even if you don't know if you're going to need those questions. You only need the answer to that question if your code branches down this path, which is going to be dependent on the answers of some of the other questions. Ask them all up front, because you want to get them all in one round trip to our servers. You'll get that in like 150 milliseconds. Now you have all of those answers. Our costs are so cheap, Pratokin, that it's almost certainly going to be worth it. And now you have all of those in advance. And your code can do a whole bunch of interesting logic, because you already have those answers. But if at each point in that branching logic, you're reevaluating, oh, OK, it's a billing ticket. Or it's-- yeah, it's a billing ticket. Now I need to know, is it a refund request? Oh, and I've got to ask Jeff again. It's like, no, no, no, ask up front. They say, is it a billing ticket? And then also ask, is it a refund request? You're only going to care about the answer to that second one if the answer to the first one is true. But ask them at the same time. It's a very small amount of additional extremely cheap tokens to ask that second one. And now you don't have to do another round trip. And frankly, oftentimes, if you do need to do another round trip and send the entire state again, might actually cost you more. So that is like, it's such a powerful way of using it, because you lose the speed benefit if you do a bunch of sequential calls. A tip that I always-- when I'm vibe-coding up something with Jeff, after a first sort of pass on the project, I'll tell the coding agent, I'll say, hey, if we are ever sending the same state to the type safe API, in different places, we should almost certainly be bundling those into one call-up front. And pretty much always, the agent is like, oh, yeah, we are doing that, because we were doing it sequentially. But now I understand. And I should put them all on the same thing. So there's like, I think people are going to need to have to really push their agents to do that less and less over time. As system one models, as decision models become ubiquitous in the way that we think about it as it becomes in the coding agent's training data, they're going to be good at it. But right now, we really need to hold the agent's hands. Yeah. No, I think you're right. It is a paradigm shift. And we have to adopt a new model. I guess that's what comes when you open up a new category of the API model. Let me shift focus a little bit and ask you one more question before I'm going to let you go. Of course. We're here at the Jeff Aton with the AIR Collective at the code over at HQ. What are the projects you're looking at? You're excited about what have you seen so far? What are you looking forward to seeing that you've built? - I haven't seen any projects yet. I've been, I've been doing interviews since the intro. I haven't seen any yet. What I am, what I'm very excited about is that one of the judging criteria in today's hackathon is novelty, is like coming up with something like really new. And as I've talked about multiple times in this interview, like that's what gets me the most excited. Like I wanna see people, and I hope I gave enough. I think I gave enough in my introduction speech of how to think about Jev that it's gonna spark some of these ideas and people's heads. I hope we get some of the, what I call the drawing board ideas where people are gonna be like, oh my God. I hope, I really hope that someone pulls an idea that they have had for a long time, that always felt infeasible. And now that they learned about Jev, and they heard what I said, they're gonna go, I think I could actually do that. I want someone's old idea to come to life that was basically impossible before, infeasible before, and now they can make it today. That's what I wanna see. And I'm gonna be coming back almost certainly for the judging because I wanna see what people built. - Awesome. Well, what a wonderful way to round it up. Thank you so much for taking the time. And best of luck with everything you do. I think it's fair to say that we're all very excited to see. - Absolutely, thank you for having me on.