1009: 54% AI-Generated and Climbing — State of AI

Scott and Wes react to the freshly released State of AI 2026 survey, covering everything from skyrocketing AI adoption and the rise of coding agents to the pain points, job security fears, and big philosophical questions developers are wrestling with right now. Show Notes 00:00 Welcome to Syntax! 00:06 Introduction to the State of AI in Web Development 02:47 Survey Insights: AI Coding Adoption and Sentiment 06:55 Models and Providers: Usage and Sentiment Analysis 09:14 Paid Agent Usage: Who’s Paying for What? 11:31 Agents and Assistants. 15:21 Personal Experiences with AI Tools 20:33 Paid Agent Usage 22:36 Programming Languages: Trends in AI Coding 23:54 Image Generation Tools: A Shift in Usage 26:48 The Evolution of Image Generation Tools 28:44 The Future of Video Generation 29:42 AI in App Development 30:54 Code Review Tools 31:27 Brought to you by Sentry.io 34:59 AI’s Role in Code Generation and Review 35:50 The Financial Impact of AI Tools 38:03 Local AI: Trends and Misconceptions 41:28 AI Risks and Pain Points 45:13 User Sentiment Towards AI 48:44 The AI Bubble: Current Perspectives 50:52 Survey Insights and Future Directions Hit us up on Socials! Syntax: X Instagram Tiktok LinkedIn Threads Wes: X Instagram Tiktok LinkedIn Threads Scott: X Instagram Tiktok LinkedIn Threads Randy: X Instagram YouTube Threads

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Wes Bos?

  Welcome to Syntax. Today, we're talking about the state of web dev AI, the state of AI results. There was a survey that was just completed about 10 days ago. This is from Sasha Greif. He runs the state of web dev API, the state of JavaScript, the state of CSS, you know, all of these things that we do. And he's been running one for two years called the state of AI. And it's actually kind of interesting because it gives you, I think it gives us a pretty good outlook as to like, what are developers feeling? What are they actually using right now? How much of their job are they using it in? And we're going to go through some of the best parts and sort of take a look at like, what are people using? What, what models are they using? How much of the code are they doing? All of that good stuff. My name is Wes. I'm here with Mr. Talinsky. How are you doing, Scott?

• 0:44Scott Tolinski?

  Hey, I'm doing good, doing good, doing good. Drought conditions in Denver and we're getting a ton of rain right now. So I'm loving having a great time, enjoying the lovely rain. Have you ever been to Hawaii, Wes?

• 0:58Wes Bos?

  No, no, I have not.

• 1:00Scott Tolinski?

  You must go. Oh, why, you simply must go. That was my favorite part about Hawaii was just that it rains all the dang time. I love that, or at least in Oahu, it does. So yeah, big, big fan of the rain. Got to love the rain and I moved to a place that does not rain very often. So gotta love that.

• 1:24Wes Bos?

  Real quick, Amsterdam. We are going to be at Amsterdam for a Syntax meetup on June 10. This is going to be a part of the opening party for JS Nation and React Summit. These are awesome conferences that are happening on June 11 and June 12. And you don't have to get a ticket to these conferences to come to the Syntax meetup. You can simply just go to syntax.fm/meetup and grab a free ticket for that. But we highly, highly recommend that you also come to the conferences. Some of the brightest minds in the industry are going to be at this thing. So check it out, jsnation.com/reactsummit. You can get a combo ticket, use coupon code SYNTEX for 15% off. Again, that's syntax.fm/meetup for tickets to the Syntax meetup.

• 2:07Wes Bos?

  And then also you go to jsnation.com and reactsummit.com to grab tickets to the conference. We'll see you there. Peace.

• 2:14Scott Tolinski?

  Well, let's talk about rain on some people's parades. We're going to be talking about the state of AI in web dev. I know there are some people who are still holding out on the fact that AI coding sucks and that it's never going to be any good. But from the results of this survey, AI coding has had a massive jump in terms of people who use it. And I think we were speculating a little bit beforehand, obviously, that it's a little bit of becoming more normalized. But also, it's gotten better over the past year, obviously. But I think people maybe are a little less eager to just say, F that, I'm not doing any AI coding at all. As you can see, the percentage change here has been pretty massive.

• 2:59Wes Bos?

  Yeah, I think there's, there's a mix of people realizing, oh, it actually is, is kind of handy. I could see that by if I post stuff on Blue Sky, I don't get absolutely annihilated.

• 3:12Scott Tolinski?

  Same, I've noticed that too, yeah.

• 3:14Wes Bos?

  I like, I don't want to be the like, like a hype master. I feel like I have a pretty good balanced take on this type of stuff and it's certainly not a silver bullet. But I don't think, I think this is a fairly monumental shift in our industry and I don't think it's going away. So it, I think in the last maybe six months or so, we've seen a lot of the people who have been very against it. To say, oh, this stuff is actually pretty helpful or, oh, it actually is, is much better at what I want to do right now. So there's certainly people on the other end of the spectrum, which are full psychosis and shipping all kinds of slop. But there's

• 3:55Scott Tolinski?

  37,000 lines a day, baby.

• 3:58Wes Bos?

  Yeah, yeah, for sure.

• 3:59Scott Tolinski?

  I, I, yeah, I think it's interesting here. We see a big jump where 18% of respondents to this, they have written, it seems like 75% of their code they write has been written with AI. And that's a big jump from the year prior than that. And, and even higher than that is 18% of folks write, let's say 85 to 90% of their code with AI. And granted, this survey, the results are going to be skewed towards people who are writing code with AI. Right? It's the state of AI. And it's also going to be skewed towards the people who responded to this survey.

• 4:43Scott Tolinski?

  So again, that's 1,200 people, which in the vast world of our industry is a small, small bit. But again, this is a sampling, right?

• 4:55Wes Bos?

  And what would you say you're at right now? Because if you were to ask me, I'd just be like, I don't know that I can put a number on that because it's just, it's just like part of my workflow of.

• 5:05Scott Tolinski?

  Yeah.

• 5:07Wes Bos?

  It's not like I was like, oh, now I'm going to go to this other place and only write with AI. And now I'm going back to my notepad, notepad plus plus and hand coding. It's, it's just a kind of a back and forth. And sometimes you're jumping in and you're factoring something, you're reviewing something.

• 5:24Wes Bos?

  It's, it's, I don't, I don't know where I would be at this, probably at like a 76. If I were to guess, but like it's similar to like, how much of your code is written with tab completion? A couple of years ago, it's, it's a kind of a hard thing to sort of guess there.

• 5:43Scott Tolinski?

  Well, mine is, mine is closer to 90% Wes, strictly because AI loves to write a shit ton of code when it does not need to. It's like.

• 5:51Wes Bos?

  Simply by volume of code?

• 5:54Scott Tolinski?

  By volume of code, it's, it's just like pouring it in there, even if it doesn't. I don't, I don't know if we're talking about like effectively, I'm not quite sure. Cause I did just like fully hand code a project the other day from scratch and it was great. I got a lot done, but again, it was probably more efficient than what the AI is writing. I would say mine has definitely obviously increased from the year prior to this. So just like many of these things, I think that number is just going to keep going up. And likewise, in terms of refactoring, 21% of respondents said that they're constantly refactoring their code with AI, which is a 10% jump from the last year. So. Yeah, that's a big old jump right there. I think we're gonna see big old jumps on all this

• 6:37Scott Tolinski?

  stuff from the sentiment in general from year to year. Now, this is only the second year they've done this survey, so there isn't a ton of historic data, and I do have some critiques about the survey. Love Sasha. I do think there's a number of things that would be very interesting that this survey is missing, which we'll get into those. Let's talk about models and providers. So this includes both experience and sentiment, whether you've heard of it, used it, never heard of it, blah, blah, blah. The first one being the most people have heard of and used it is ChatGPT. And I think this encompasses all OpenAI models, the way that this is listed.

• 7:20Wes Bos?

  Yeah, it's weird that they label it ChatGPT, and I guess like some of the most recent OpenAI models are called GPT. And we gotta do a better job at naming these models, you know? It's such a confusing thing. But I also do think there's quite a few people that are still just typing into ChatGPT, copy pasting the code in and out of their thing. I don't think it's a large amount, but I think there's quite a few people doing it still.

• 7:49Scott Tolinski?

  Totally. Yeah, I think there is. And that's honestly, it's a less scary introduction to coding with AI. Because you can say, is this, you know, I just have this quick little question. Okay, I got a little snippet. Okay, it works or it doesn't work, whatever. I can eyeball it from there. And that is like a big jump between that, and then you have the in-editor tab completion stuff, and then you get into the world of just like, hey, chat, make me an app, you know? I mean, there's different levels to this in terms of what you're using. But as far as models go, people have used it, and negative sentiment is 10% for ChatGPT, which I'm just going to chalk

• 8:29Scott Tolinski?

  this up to all GPT models, where Claude is used it and has negative is only 2%. So a 10% negative versus Claude has a 2%. And likewise, Claude also has that 46% positive. So people, and this is, this, I'm gonna say this mirrors my experience as well. I, in fact, when we did this, how much do you rage at AI? I raged at Claude considerably less, which I feel like has to tell you something.

• 9:01Wes Bos?

  If you scroll down underneath this, there is out of the models mentioned, which of these does your company actually pay or do you or your company actually pay for? I think this is probably a better outlook of like with these like I've heard of it and have positive sentiment. I think what happens is that people have like, yes, all of these people have clearly heard of chat GPT and Copilot, but like, like, what are people actually using right now? I think that this looks a little bit, it gives you a better look at it, right? Like 69 nice percent using Claude chat GPT. This is the worst. This should say OpenAI, right? 50% people at OpenAI, 32% Gemini, 6% Copilot.

• 9:44Wes Bos?

  So you're telling me that 32% of people are paying for Gemini, but only 6% are paying for Copilot? That doesn't add up to me, don't you think?

• 9:53Scott Tolinski?

  Well, it doesn't add up because Copilot is listed as an agent.

• 9:57Wes Bos?

  Oh, it was a write in. Okay.

• 9:59Scott Tolinski?

  These are a write in where people are like, I use Copilot as my model.

• 10:04Scott Tolinski?

  Like they don't know what model they're using. They're just using, I'm using.

• 10:08Wes Bos?

  I don't think people actually know what any of this stuff is. I had, I posted something the other day where I reverse engineered my lights with Claude. And the amount of comments I had from people saying, you could just use local models. And I said, which model could I use to do this? And they're all like Copilot or using, they were essentially just naming like free trials, which they have some, for some reason, think that that's just like a local model. Like, I think people are a lot more confused as to how this stuff actually works.

• 10:42Scott Tolinski?

  Yeah, I think that is evident if you look at this data further because you'll see that under models, Microsoft Copilot also a heavy write in. And then Kuro is a heavy write in, Kuro as a model.

• 10:55Wes Bos?

  So yeah. Copilot doesn't even have a model yet. I bet they'll probably be releasing one in the next month or so. Like they, they had this big thing a couple of months ago where they're like, we now can train on your, they had like a big banner at the top of GitHub being like opt out if you don't want us to train on your models. I bet they're going to roll out their own version pretty soon, but they don't, as of recording, they don't have their own model.

• 11:16Scott Tolinski?

  I think, I think this does illustrate something interesting beyond the data here is that people don't know what a model is.

• 11:21Wes Bos?

  No.

• 11:22Scott Tolinski?

  Yeah.

• 11:22Wes Bos?

  The app is the model, I guess.

• 11:25Scott Tolinski?

  The app is the model. The files are in the computer.

• 11:28Wes Bos?

  Yes.

• 11:29Scott Tolinski?

  The files are in the computer? Let's talk about agents then, because I think this leads in nicely to that, where again, Copilot seems to be the most used, but not the most loved, where it has a 13% sentiment, a negative sentiment. I think one of the most interesting things in this is a little bit of bubble breaking. You know, we hear about things like Devin on Twitter all the time, and according to respondents of the survey, most people have not heard of it. Forty-two percent of people have not heard of it. And the people who have heard of it don't like it. 14% of it have a negative sentiment about it, where only 2% have heard

• 12:09Scott Tolinski?

  of it and have a positive sentiment, and only 0.5% have used it and like it. So Devin is not either, either is outside of this bubble, or people who have actually tried it aren't, aren't a fan.

• 12:23Wes Bos?

  Oh yeah. I think there's going to be two things going forward with like, what do people use? Is first, is it included in your, whatever you're paying for, your Microsoft 365, your Google Workspace? Like, I think that those guys are going to have a huge leg up, because it simply is just part of the package, and then they pay extra for it, and it's integrated with absolutely everything. They don't go looking for another tool. And I think that's, that's partially why Claude and Anthropic right now are gunning so hard for like, the like, Office thing, because those enterprise contracts are, are massive.

• 13:00Scott Tolinski?

  Yeah. Yeah, I mean, it's evident here by Copilot being number one in terms of the most experience, and then Claude Code being number two in the most experience. Crazy enough, 34% positive sentiment from Claude Code over the 21 positive on Copilot, and then Codex, OpenAI Codex is only 16% of those who have used it like it.

• 13:22Wes Bos?

  Where?

• 13:23Scott Tolinski?

  Yeah.

• 13:24Wes Bos?

  What's wild to me is that Cursor is not even in this list anywhere.

• 13:29Scott Tolinski?

  Yes.

• 13:31Wes Bos?

  Tab 9 is in here. I didn't even know that was still a thing, you know?

• 13:37Scott Tolinski?

  And I have thoughts about this, Wes, and we'll talk about that, too. I have thoughts about that because if you get into the other coding agents, you'll see Open Code was a write-in for 389 people. Open Code is very highly used from people, but...

• 13:55Wes Bos?

  They have like 10 million, something like that, monthly or daily active users, monthly active users? It's insane how many people are using it.

• 14:02Scott Tolinski?

  Cursor, a write-in as well, and Py, a write-in. I think that Sasha needs to, I think, maybe grow with the industry here. And that Open Code, Cursor, Py, Kiro, all of these things need to be in the survey. Because, yeah, even Py has a decent usage. It's nowhere near these other big boys like Claude Code. But Open Code is a major player, and Cursor is a major player. It's a playa playa.

• 14:37Wes Bos?

  Yeah, and maybe, I think maybe that's part of maybe Cursor's problem is that people don't see it as a coding agent. They see it as like an editor. And that's why Cursor's been pushing so hard on like moving away from their IDE into like their agent interface, which absolutely everybody is doing. You know, you got your projects on the left, you got your chat in the middle. You know, that's the same UI everybody is building right now. But I think that's part of their problem is people see them as only an IDE. When in reality, it's like Cursor's actually a pretty good deal. And you still get access to, you get access to all the models. You get to try them all, which is pretty cool.

• 15:21Scott Tolinski?

  Yeah, I've been personally, Wes, this is a good time for us to talk about it while we're using.

• 15:29Wes Bos?

  Yeah.

• 15:30Scott Tolinski?

  What's your current usage of choice? You know, mine is changing all the time.

• 15:35Wes Bos?

  I'm probably 50/50 split between Claude and Cursor and

• 15:43Scott Tolinski?

  Like Claude Code or Claude the app?

• 15:46Wes Bos?

  No, Claude Claude Code. And then the other split is is probably Cursor. I do dip into Codex fairly often as well because I have a ChatGPT subscription. And just kind of get a get a feel for for how it works. And I do use the Codex models via Cursor probably more often. Because I still, I know what nobody likes having an IDE anymore, but I still find it nice to be able to like, click on something and see the code that that's being generated. I still like that UI the best. So that's kind of my my go to. But then I also like the like the terminal experience quite a bit still.

• 16:28Wes Bos?

  As much as I say I don't like Tuis for just quick stuff needing to to, I don't know, like FFmpeg resize a whole bunch of stuff or download a whole bunch of videos or upgrade and install, you know, I just Claude dangerously skip permissions. And I find that that's one of my favorites as well.

• 16:48Scott Tolinski?

  Yeah, I've been using I've been using Claude Code for like almost everything these days. And that's like in the past couple of weeks, I've like really landed on Claude Code to the point where I did cancel my ChatGPT subscription. I find Codex to be just an absolute slog to work in. And I know some people really like Codex because it's like more exact. But I just, I I just did not like any of the results of the front end code it was writing ever. And like to me, that was such a thing that I just was so annoyed with, like, it was hot potatoing data from function to function to function to function. It was writing so much unnecessary code

• 17:30Scott Tolinski?

  that I was just getting annoyed with it. And I'm I don't have that same experience with with Claude. And I feel like maybe my Claude Code setup is a little bit better. And then I'll also dip into, I have a a copilot subscription. So I'll use Copilot in open code now and in Py. I'll use Copilot models in both of those. And I'll use Py if I want to have something just run on my system at large. And I'm pretty much using Claude code only on my, like, in my projects. And then like, if I want to say, hey, go to my NAS and delete 100 files. I'm going to do that in Py.

• 18:09Scott Tolinski?

  So just because I have, I have access to those Copilot models in Py is pretty permissive on that kind of stuff. It's like where Claude code is going to be like, do you want me to read the file that you told me to read? Please, like when I read it? And I'm just like, shut up.

• 18:25Wes Bos?

  I threw Claude on my Synology. And like, the one thing that I hate doing is updating Docker images.

• 18:31Scott Tolinski?

  Same, oh same.

• 18:32Wes Bos?

  I've got Jellyfin on a Docker image. And I, for some reason, I, every time I need to update that thing, I'm just like, ugh. What do I do? And like, because there's, there's data that's been like mapped. Obviously with Docker, you can map things and you don't keep the files in the Docker image. You keep them on your file system and you sort of map them there. But so many times I've accidentally nuked all of my like library data. And I've had backups, but it's annoying. And now I just tell Claude code to update the thing and it rips on it. And I know that probably makes a lot of people mad. Yeah. Every time I post about that, they're like, just use X, Y, and Z. I'm like, I need to touch this thing like twice a year and I always forget what to do.

• 19:17Wes Bos?

  And now I just type in the box and it, it works great.

• 19:20Scott Tolinski?

  Yeah, I have that same thing with Image. Image I run on my NAS and like, I don't know, the updating process for it absolutely sucks. So because they don't have Synology, they don't have like an official Synology thing. So it's like, you gotta duplicate the container and do that.

• 19:38Wes Bos?

  Yeah.

• 19:39Scott Tolinski?

  I don't want to do all of that.

• 19:44Scott Tolinski?

  I, it's funny that you run Claude code on your NAS where what I do is I just have my local agent on my laptop have SSH.

• 19:56Wes Bos?

  Okay.

• 19:59Wes Bos?

  Is that better?

• 20:03Scott Tolinski?

  I have no idea.

• 20:04Wes Bos?

  Is there any downside to that? Like, what do you just tell it to go SSH in? And I

• 20:10Scott Tolinski?

  I just have all my stuff set up, my key set up and whatever. Everything's on my tail scale network. And it's just like, oh, just SSH into it. Yeah. And it doesn't, it's not like the password is getting passed into.

• 20:23Wes Bos?

  No.

• 20:24Scott Tolinski?

  Yeah.

• 20:25Wes Bos?

  Interesting. Maybe I need that as well.

• 20:29Scott Tolinski?

  Yeah.

• 20:29Wes Bos?

  But who knows?

• 20:29Scott Tolinski?

  Fuckin' the shot.

• 20:30Wes Bos?

  Cool.

• 20:30Scott Tolinski?

  Who knows? Okay, let's get back to the survey here. So let's look at paid agent usage, what people are actually paying for in terms of agents. We talked about models and paid usage. 58% of the people who responded to this are paying for Claude code, whether that is through a max plan or through tokens.

• 20:49Scott Tolinski?

  We have 42% are paying for Copilot and they're about to be paying

• 20:52Wes Bos?

  Yeah, see, that makes more sense.

• 20:53Scott Tolinski?

  Whole lot more.

• 20:56Scott Tolinski?

  Did you see that like, so they have like, Copilot has like a new dashboard that shows you like, how much you would be paying if you were actually paying per token.

• 21:04Wes Bos?

  Oh, really?

• 21:05Scott Tolinski?

  CJ was telling me that his, his cost was outrageous in terms of like, how he thinks, he thinks people are for a major root awakening in terms of

• 21:16Wes Bos?

  Oh, yeah.

• 21:16Scott Tolinski?

  Because it was like 60,000 bucks or something crazy.

• 21:19Wes Bos?

  Seriously?

• 21:22Wes Bos?

  That's insane. And I saw Peter Steinberger, the creator of OpenClaw. He had burned a million and a half dollars worth of tokens in a month or something nuts like that. Like

• 21:34Scott Tolinski?

  That guy's crazy.

• 21:35Wes Bos?

  And obviously that guy, that guy's trying to figure out what coding would look like if there were no, there was no like cost.

• 21:43Scott Tolinski?

  Yeah, yeah, yeah.

• 21:44Wes Bos?

  And like whether the cost for that will ever come down. Like it would, it would have to come like even $60,000, you know? Like even if that were to come down 60x.

• 21:56Scott Tolinski?

  Yeah.

• 21:57Wes Bos?

  Thousand bucks a month. That's a lot of money. That's a lot of dough.

• 22:02Scott Tolinski?

  I know. Um, further things, 24% are paying for OpenAI Codex. 5% are paying for Jetbrains. And then, yeah, the Cursor needs to be on here.

• 22:11Wes Bos?

  Four, it is on here. Oh, it was a write-in.

• 22:14Scott Tolinski?

  But it's a write-in.

• 22:15Wes Bos?

  I don't understand how it was not in here because a year ago Cursor was arguably more popular than it is right now.

• 22:22Scott Tolinski?

  Right. Yes, I know. That's agreed. And then also like, there's, there's no, like, even like Zed or something isn't on here. No, I don't, I don't know people are paying for Zed. I don't know, but that is like their whole paid product. Okay, let's get into other tools here. Um, it's no surprise that the primary amount of, uh, the primary language that people are writing with AI is TypeScript. I'm surprised to see JavaScript at 77%. Like, I'm writing TypeScript. I haven't touched a JavaScript file in a billion years at this point.

• 22:52Wes Bos?

  No, it's, I, I think, again, those are probably people being like, well, TypeScript is JavaScript, so I'll check that off. But like, I, I'm at a point where you don't even have to tell it to write JavaScript. You know, or you don't have to tell it to write TypeScript. It just knows, you know, like, I haven't had it accidentally write JavaScript without types in a long time.

• 23:10Scott Tolinski?

  Yeah, totally.

• 23:11Wes Bos?

  Python, 41% makes a lot of sense. Python is still big. I'm, I'm happy to see the, uh, the shift to more JavaScript. Because like, when we first started with a lot of this AI stuff, it was all Python all the way down.

• 23:26Wes Bos?

  And we were living in the reptile building for sure. Shift towards TypeScript.

• 23:30Scott Tolinski?

  Um, 11% in Rust, which Rust is actually, I will say, pretty good to write in AI because of, uh, the build system and everything. Or the, uh, just the general nature of the Rust language, I think lends itself well to this. But yeah, uh, HTML and CSS being in here as write-ins, I don't know, I don't know how to feel about that. I think they're, they're programming languages, whatever, who cares? I don't want to have that discussion. Um, image generation, most people are using Nano Banana and then just ChatGPT second on here. What do you pick up for image generation? I know you, you've tried a few things.

• 24:05Wes Bos?

  Yeah. I have been, I was Nano Banana for quite a long time. I do dip into ChatGPT's image generation every now and then. And then I recently started paying for Mid Journey again. Like, I was like, very early on Mid Journey customer. And then I can't, I canceled it. And then I, I kind of went back to it because it has like, has like the best tools. Like, it's almost like Photoshop, like those layers and things like that. And I don't, like, I, I, I'm trying to think like, like what I actually use it for. Obviously, for March Mad CSS, we made those hilarious images of everybody in like 90s basketball thing. I've used it to generate some icons here and there.

• 24:47Wes Bos?

  It's often some, like, I use it more for like assets, you know? Like, if I need like flames, if I'm doing like a thumbnail or something like that, and I need some like flames. Like, more of my like, like, I used to just Google image search for things. And then I would find the like stupid free pic with the fake transparent background. And then, you know, I, I just, I'm so, Google image search for stealing photos is awful now. So now I just reach for, for something like Mid Journey or whatever, just to get assets.

• 25:22Scott Tolinski?

  Yeah, Mid Journey looks pretty great. I, I reach, yeah, pretty much, I pretty much reach for Nano Banana for generating assets. Or, or just ChatGPT for assets. And I've been doing the same thing primarily, like, for the Faz's podcast that Courtney does. I was using like beach photos for her thumbnails, just because like that's her vibe. And she wanted something like nice and calming, so we were using waves. And, yeah, finding good royalty-free photos of waves was annoying. I'm just going, give me, give me 20 waves photos that are like calm waves and pristine beaches. And it's like, okay, here they are. So much endlessly easier.

• 26:05Wes Bos?

  It's so true. Those websites that bait you on images have absolutely ruined the process of finding images. So Midjourney is cool. I don't know. I keep paying for it. I really like using OpenRouter for a lot of these image models because you simply just pay for what you want. Like when we were doing the MAD CSS, I racked up like probably forty, fifty dollars worth of usage on these image things. And I don't mind that because then like there's months where I don't generate anything and you don't have to pay for it. Whereas like Midjourney, they don't have an API. It's simply you gotta pay the like eleven bucks a month. And I think that's where a lot of these models are going to be going.

• 26:48Wes Bos?

  Where they don't want people piecemealing. Eleven Labs is like this as well. They don't want people just like, oh, I'm gonna take a sip here. I'm gonna generate 12 seconds worth of audio and then I paid my 18 cents for that. And now I'm not going to use it for the next 20 years. They want you paying like 19 bucks a month and you have to go ahead and cancel if you don't want it and then they'll delete all of it. Unfortunately, I think that's where a lot of these are gonna go. And I bet a lot of the big models out there, like I bet we'll see it from Claude and I bet we'll see this from OpenAI. They're probably gonna release something where you, there's gonna be no API for it. You simply have to use it via their harness.

• 27:27Scott Tolinski?

  Yeah. Yeah, the OpenRouter stuff is really interesting. Yeah, because was it just even like yesterday, the Grok Imagine showed up on there. It's five cents per image. So five cents per image gen via the OpenRouter. And then there's also stuff on here I haven't used, like ProVector, ProVector SVG Craft in terms of like an image gen for vector-specific image gen. I haven't used that stuff and it looks pretty, pretty sick, actually. I did some of the like the playgrounds and it looked good.

• 28:03Wes Bos?

  I also have a replicate.com where I throw, every now and then I'll throw 20 bucks in it because they have a very good vectorization model. And you just use it as you need, as you need it, right? Every now and then I'll have something that needs to be vectorized. And they just, you can just upload it via the website. And then like, you don't have to use their API keys or anything. You just use the GUI on the website. Also, the Grok image one is actually very good and it's fast, super fast. And when you're trying to like, like spitball on ideas for things, I find that that's actually been very good.

• 28:41Scott Tolinski?

  Yeah, yeah. Open Rudder, shout out to Open Rudder. Video generation, we don't need to get too much into this.

• 28:47Wes Bos?

  No.

• 28:47Scott Tolinski?

  It's not super related.

• 28:48Wes Bos?

  I see, well, let's talk about this, like what? I see no purpose for having AI-generated videos aside from haha, a pig jumped, like the pig is doing high jump or something like that, like those funny videos every now, I'll get, I'll get a little laugh, but I don't.

• 29:06Wes Bos?

  I don't see the, the actual, maybe Hollywood is, wants this type of stuff, you know, like B-roll, like a, maybe that would be it. Like a, instead of having to like fly a drone over New York City, you could just generate it. But like, I just think a lot of this AI-generated stuff is just, that's a lot of money to spend on some, I've seen no good come of it. And in fact, I've seen a lot of awful stuff come from us. Every, all the boomers on Facebook are getting duped all day long.

• 29:34Scott Tolinski?

  Yeah, no kidding. Yeah, I don't, I don't have any, any use case for that as well. So I just don't get it. The next one is going to be app generation. So V zero, 12% of respondents have used V zero, 9% of yous lovable. I, again, these aren't tools that I, these are like, feels like these are like non-coder tools.

• 29:57Wes Bos?

  Yeah.

• 29:58Scott Tolinski?

  Yeah.

• 29:58Wes Bos?

  That's, it's interesting to hear it. I think in the early days, these things were absolutely blowing up. But now when I hear from like my non-coder friends, often I'll hear is, hey, did you know Gemini can make games? Someone sent me a, like a friend sent me a message the other day and they said, hey, is this code any good? They're asking me like, I know I made a game and. It's interesting to see like where people are coming at. Also, I think a lot of the like early hype on these things was very paid off. It's the like amount of people on TikTok or whatever being like, hey, I made this thing on lovable or whatever. I think a lot of those people were were being paid because you don't see them talking about it anymore.

• 30:42Scott Tolinski?

  Yeah, I know. I know lovable really burnt a ton of goodwill too with that whole. We meant to, we meant to expose all your data. No, wait, nevermind. No, we didn't, you know? Yeah, that, yeah. The code review tools are interesting here. 14% of people who are using CodeRabbit, 13% of people are using Copilot, which I'd assume they're just prompting Copilot, Claude Code, etc. So people aren't using like code review tools, even if they've responded to their saying that. And the ones that do actually show up are mostly write-ins like Bugbot and then Sentry Seer has 15 write-ins. So again, I don't know if people are actually, I think this whole section is mostly write-ins. So they need some actual code review tools.

• 31:27Scott Tolinski?

  And Seer is a great, it's a great time to talk about. Seer from Sentry is, it gives you the root cause analysis of any of the issues in your code and has access to, with your Sentry account, it has access to the whole context around why something broke and how to fix it. And Seer is really super cool. But there's actually some other interesting code review tools from Sentry. I don't know if you've seen Warden, Wes. Have you seen Warden?

• 31:56Wes Bos?

  Yes, I have.

• 31:57Scott Tolinski?

  Warden is a new tool from Sentry that I think is really cool. Also, they got this sick landing page. Have you seen this thing?

• 32:04Wes Bos?

  Yeah.

• 32:04Scott Tolinski?

  You click and a bug shows up and then a little laser comes and zaps the bug. This is so stinking cool. I love this landing page. Oh, my God. This is great. And what this thing does is it, it's a good tool to have in your stack that really is watching your codebase for all kinds of things to fix it, to Find issues automatically, whether that is code review or security review or dependency review, test coverage, those types of things. And it will alert you in your GitHub actions.

• 32:43Scott Tolinski?

  It is really pretty sick. The way I like to think about this is it's kind of being like hooks for finding issues in your code before you push them. If you were using something like Warden, it might find potential security holes before that you actually merged something. It's really pretty cool. Their little example up here is a Warden finds a SQL injection from unsanitized input possibility and even gives you a suggested fix directly from your GitHub issues. Man, I, this is a tool that I'm instantly going to add to all my stuff and not just because it's a Sentry tool and I work for Sentry, but because this is the type of thing that I think we all be needing

• 33:28Scott Tolinski?

  right now with the current security landscape.

• 33:31Wes Bos?

  You know what is one that everybody needs to add is error handling. So often AI write code where it doesn't handle the error properly, or it will simply just like put a like a label somewhere on there that something happened and just having like consistent, proper error handling techniques so that obviously the error gets sent off to Sentry, but you also show something to your user as to what happened or be able to handle it and the thing will fix itself in a lot of these cases. That needs to be added to absolutely every code base that people are doing because there's so many times you're just using an app and nothing happens or like something went wrong, you know?

• 34:14Scott Tolinski?

  Totally. And in perfect tool for that is Spotlight. With all this cool stuff Sentry's been shipping, Spotlight does exactly that. It's basically like local Sentry and you get access to all the errors in your app, especially in development, not just in production. So as you're going, your AI is having access to all of the errors in a way that it actually has access to solve them and then can use the actual CER AI. Tools to assist in debugging and fixing. These are the types of things, again, that as we shove more and more AI code out, we need automated and better tools to find our bugs, solve them and

• 34:52Wes Bos?

  Yes.

• 34:53Scott Tolinski?

  Potentially alert us to security issues. So shout out to the Sentry team for actually being so future thinking. So, okay, let's get into the survey, back into the survey now. So, how are people actually using AI in their coding? 90% use it for code generation specifically. 68% up from 8% before, so a 60% increase. Use it for code review and assistance.

• 35:21Wes Bos?

  That makes no sense. How the hell did that go from 1% to 68%?

• 35:26Scott Tolinski?

  Maybe it was a write-in before. Maybe it was a write-in before.

• 35:29Wes Bos?

  Yeah.

• 35:29Scott Tolinski?

  Yeah. 68% use it for learning and research. 67% debugging. And then you get into much smaller. 37% use it for image generation. And I don't even see video generation on here, so. Yeah, so again.

• 35:46Wes Bos?

  I don't know, it's the wrong industry for this, yeah.

• 35:48Scott Tolinski?

  Totally, yes, agreed.

• 35:51Wes Bos?

  Next one we have here is personal expenses. Like, how much do you spend on AI tools every single month? 40% are saying $0. So there's quite a few people that are just getting by on the free trials. And I think they're hopping from one trial to another trial. I think that probably the days of, those days are very limited. And I would probably say in the next six months or so, that we're not going to have those free plans or even those cheap plans anymore. Like, right now, what you're seeing is everybody is saying, "F you, Claude Code, I'm going to Codex where I can use it a lot longer." It's like, yeah, that's just because Codex, it's like a game of chicken right now. It's like, who can lose more money longer?

• 36:34Wes Bos?

  And if you think that, like, Codex is gonna be like that forever, they're happily taking you over right now. We're like, "Hey, come to us, we'll give you better limits." We won't, we'll let you use it with other harnesses. We'll let you do everything and people are coming to them. But you think that's not gonna happen with OpenAI? It's for sure gonna happen and the prices are gonna go up relatively soon.

• 36:57Scott Tolinski?

  Yeah, they are.

• 36:58Wes Bos?

  Really?

• 36:59Scott Tolinski?

  Just trying to get you to use it.

• 37:01Wes Bos?

  But let's take a listen. 10% of people are spending between $100 and $500. And then

• 37:08Scott Tolinski?

  That's up from 9%.

• 37:10Wes Bos?

  0.6% of people are using more than $500. And I guarantee you, if you're using AI every single day, like the results of these are saying, you are using thousands of dollars a month on this type of stuff. So that the disconnect between what people are actually paying or willing to pay and how much compute they're actually using is absolutely insane.

• 37:33Scott Tolinski?

  Yeah, it's insane. I would be surprised. I mean, well, I actually, here's why. The fact that 0 went 13% down and 100 to 500 went 10% or 9% higher than the year before. I think you'll see that, that, that, that, right? I think you're going to see those numbers kind of invert. People are paying a lot of money. It's getting more expensive and more people are finding it to be useful. So therefore they're going to continue to spend more money on it. Yes. Um, I think, let's take a look at local AI because local AI often a conversation topic. 20% are not interested, uh, 20 or 20% of people don't use it and are not interested. 31% of people are, don't use it, but are interested.

• 38:18Scott Tolinski?

  And then 49% of people are saying they use local AI. I think that is not accurate. There's no way that's accurate because look, what are people running their own rigs, their own GPUs? What are they running? The crappiest models possible there are? I don't think they are. I think people don't understand what local AI is.

• 38:38Wes Bos?

  I don't think so either.

• 38:40Scott Tolinski?

  Yes.

• 38:40Wes Bos?

  Like you can use, yeah, you can use Ollama and whatever, but if you're using this to code, uh, maybe for background images, if it's slower, but still good. I, I think if, if you're worried about Like, people are using local AI if they're worried about either cost or privacy, right? And I think a lot of the cost people have have moved to Kimi, DeepSeek, a lot of these like cheaper models. And I bet that's where we're going to see a lot of people go in in the next six months as the real prices for these things start to come out. We're gonna, people are gonna be figuring out how to get like decent results out of these

• 39:21Wes Bos?

  significantly cheaper models, or the models that are, I don't know if they're they're cheaper because they're they've been made like that, or they're just, again, being heavily subsidized to to get people to hop over. And then the like, the privacy people as well, like, I would love to have a rig on my desktop that I could just run all this stuff locally.

• 39:39Scott Tolinski?

  Brother, yeah.

• 39:40Wes Bos?

  But like, that's, I don't think people realize how beefy of a system you need. And CJ has a video on it, right? You gotta have a pretty sick setup. You got, he spent thousands and thousands of dollars on on his rig. And it's like, it's it's good, but it's like nowhere near what the things that we're used to using, all these Claude and OpenAI models.

• 40:03Scott Tolinski?

  Yeah, I think that's the whole thing is to get used to the Claude, OpenAI, AI models where you just type it in and it does its thing. You're going to have a rough transition. I think the benefit comes in from using these smaller models to do smaller tasks, and then really getting into the systems of it where, all right, this one is doing the to do list, and this one is doing the the prep, and this is writing the ADR, and this one is reviewing it. And then you have your big beefy frontier models doing the actual coding. But that is a lot of stinking work. Just, I mean, it depends on what you're using it for. My homie, Topher, who's been on the show before, he has two DGX Sparks from Nvidia,

• 40:45Scott Tolinski?

  and he does all of his OpenClaw stuff with those, all of it.

• 40:52Scott Tolinski?

  So.

• 40:53Wes Bos?

  And what does he use OpenClaw for, though? Like, cause that certainly you can do things like summarize a note or.

• 41:00Scott Tolinski?

  Right.

• 41:00Wes Bos?

  Even just like the programmatic stuff. Yeah, like moving files, categorizing things, taking a voice note and putting it into a to do app. Right? That kind of stuff. You don't need like these, these frontline models. That stuff worked great three years ago with what we had then.

• 41:18Scott Tolinski?

  Totally. Yes, yes, totally. So again, yeah, local AI, interesting thing. But again, yeah, I don't think these people are really using it like they say they are. One thing I wanted to hit on is this next section, which is the risks and pain points. Job displacement being the largest thing inside of AI risks, which, uh, yeah. I think that's a big, a big point of, I would say, stress amongst developers. But I don't think we've truly seen the end results of this just yet. We hear about layoffs all the time, but I think companies who have done good hiring practice from the start, they're still hiring. I mean, like Sentry is still hiring all the time and they haven't done any layoffs.

• 42:02Scott Tolinski?

  And they're, you know, it's like companies like that, that are, uh, that are using AI heavily, but still, you know, finding a way to make their employees effective with it. Military use of AI, 45%, yeah. Environmental impact, 40%.

• 42:17Wes Bos?

  Also concerning.

• 42:19Scott Tolinski?

  Also concerning. AI slop takeover, very concerning. Negative cognitive impacts, yes. Have you experienced any negative cognitive impacts? Impacts from using AI, you might be entitled.

• 42:29Wes Bos?

  Yeah. When's the class action of like, it ruined my brain.

• 42:34Scott Tolinski?

  Yes, I know. It, yes. Yes, I.

• 42:38Wes Bos?

  Only 16% think rising AI costs are a problem.

• 42:43Scott Tolinski?

  Yeah.

• 42:43Wes Bos?

  Interesting as well. But like you, you also hear like. Like people's companies are simply like make mandating them like when we were at GitHub Universe last year, they announced like the ability to track how much people are using. And right then and there, we're like, yeah, this is what like people want their employees using this stuff to to move faster and to be more productive. And you're seeing a lot of people who are just like, no, I don't want to use this, partially because I don't trust it. I don't want my job to be replaced with it. And partially because they just don't like a lot of people don't really want to learn how to how to use this effectively. Right. Like, there's a lot of people have simply just used the chat GPT typed into

• 43:28Wes Bos?

  the box, had it spit out 18 pages worth of garbage. And they're like, that's not very good. And they don't see the the actual benefit, you know, like, give somebody who just is like a chat GPT casual listener, give them like, show them how powerful like a Claude agent can be. I think that they might be may change their mind.

• 43:51Scott Tolinski?

  Yes, I do think there is a tremendous gap there. Now, the highest thing that people say is a pain point is hallucinate hallucinations and inaccuracies. Yeah, that's par for the course. I think the there's a meme that shows like, pre AI development, and it has like, you're building a scooter, and it's like, here's the wheel, here's the other wheel, here's the base, here's the thing, and you put them all together sequentially. And then it's like, post AI development, or using AI development, and it's like some monstrosity. And then you're refining it down into the actual scooter. So like, I think people who aren't used to this stuff, they don't realize that you almost start with like, you start

• 44:33Scott Tolinski?

  with this jumbled mess, and you refine it down, rather than you start small and refine up to, you know, code quality or higher things. But the hallucinations and inaccuracy, yeah, I mean, that's the name of the game is is trying to reduce these things, whether it's through deterministic means like Linters and things like fallow or whatever, you know, skills to help it along MCP, whatever, code quality being the second highest things, the second highest pain point. Yeah. Yeah, obviously. I think these are big things. Ethical concerns, a little blip on the pain points, but I don't think that belongs inside of ethical concerns. Okay, here's, uh, here's one thing I wanted to touch on. The happiness of people who have been,

• 45:18Scott Tolinski?

  people's happiness with AI has gone from 3.3 out of 5 to 3.4 out of 5. So people who have taken the survey are generally more happy with AI than they were ever so slightly than the year before. Which tracks, I think, I mean, how could it not? The models have gotten better, the tools have gotten more mature, everything's gotten a little bit better. I'm surprised the jump isn't a little bit higher, but again, 3.4 out of 5 is pretty high already in terms of what I would expect here from this survey. What I think would be more interesting than how much do you like AI, uh, 1 through 5, would be getting into the psychological stuff here, right?

• 46:02Scott Tolinski?

  Like, I, I think about this all the time, and maybe, I, I've been thinking, Wes, this would be a good episode to bring Courtney on, to talk about psychology here. How does coding with AI make you feel? How stressed out are you with working when, with AI? I'm more stressed when working AI because I don't have as much control over it. I'm like, do the thing I want you to do. How, how does it make you feel?

• 46:26Wes Bos?

  Yes, fragmented, ADHD. Developers are probably less happy. Like, you, you can make a lot more things, and like, I certainly have this problem, is that they, I have a million ideas, my brain is always buzzing, and like, having a box that can actually make these, put these things into reality, that's, that's a bit of a problem for me.

• 46:49Scott Tolinski?

  It's brutal for my mental health, because I'm like, yeah, I'm feeling like, oh, I gotta keep this going. What did I, I installed a caffeine app on my computer, so now I can shut it and have my agents run while I, my computer's shut. That's not healthy. That's not healthy at all.

• 47:04Wes Bos?

  Here's another question, though. It's like, is software getting better?

• 47:08Scott Tolinski?

  No, I mean, maybe it will, but like, who can use any single thing on their device and be like,

• 47:14Wes Bos?

  I'm just trying to think like, from like, a user point of view of like, stuff I use every single day. Like, like, there's certainly lots of like workflows in my life that are a lot better and easier. Of things that I use every single day, my day to day things, has that gotten better? You know, like my, I always love to talk about like, Freshbooks. Freshbooks was this beautiful thing that I love to use. It was sponsor of the podcast for many years and I loved it. And over the last like, five years, the quality has just gone downhill and there's been no features and nothing ever works. And then I'm just like, looking at it being like, shouldn't this be way better? Given that we have this

• 47:57Wes Bos?

  magical genie that you could just rub and it fixes all of our problems? Shouldn't it be better?

• 48:04Scott Tolinski?

  Shouldn't it be better? It's something I ask myself all the dang time. And it's not, yeah. Like, and too many of these companies are just spending time just shoving AI into their product to make it worse. I mean, Notion, oh my God, Notion is just a nightmare to use sometimes. So, yeah, I don't, I don't know. Yeah, I don't think things have gotten any better. At least, not from my perspective. Some things have, some things have gone just perfectly, but I think those are teams that are always caring, do care about craft. So, I think it's, yeah, it's interesting. But yeah, I would love to see this survey include those types of, of, of things. If we're getting into, you know, how do you, how do you feel about it? How does it make you feel? The opinion section of this is the last one.

• 48:47Scott Tolinski?

  Seventy-two percent of respondents now consider AI to be an integral part of their workflow. Seventy-two percent of respondents. That is...

• 49:00Scott Tolinski?

  Hi folks, for the people who are saying it's useless, 72%. Do you think 72% of people are being forced by their company? Cause I know that we'll get comments being like, well, their companies are making them use it.

• 49:12Wes Bos?

  100% of the people leaving comments on this video.

• 49:15Scott Tolinski?

  Yeah.

• 49:16Wes Bos?

  But like you're telling me that 28% of people don't find it as an integral part of the workflow. I guess that that's fair.

• 49:25Scott Tolinski?

  Yeah. Yeah.

• 49:27Wes Bos?

  I think that's fair. Like, is that like, maybe those people are like, they're certainly using it, but maybe they don't find it that it's a

• 49:35Scott Tolinski?

  Integral part.

• 49:36Wes Bos?

  Integral part of their workflow.

• 49:38Scott Tolinski?

  Integral, integral being the key word here, yes. Interesting, there are a lot, we hear a lot of talk about bubble. Many respondents think that we're currently in an AI bubble, but they also realize AI is here to stay, even after it eventually pops. 37% agree with that, 33% strongly agree with that. Very small percentage of people disagree with that and strongly disagree with that. So yeah, I think we're seeing some of that bubble in just terms of the cost that we've talked about. The cost exploding and it's going to get more, more interesting.

• 50:10Wes Bos?

  What is it, what does it look like when the, the bubble pops though? That's, that's my question here is that like, I certainly think everything is chaotic right now and everybody's trying to figure out.

• 50:21Scott Tolinski?

  Yeah.

• 50:21Wes Bos?

  Like what it looks like and there's a lot of money being thrown around. I certainly think we are in a bubble, but it, I don't think that like, when that pops, I don't think that we're going to go back to, to what the way it was before this.

• 50:35Scott Tolinski?

  Chiseling into rock, yeah.

• 50:36Wes Bos?

  Yeah, I think it's just like the, I think the way that we work is significantly been changed. And there's, I don't think there's any returning back to, to where we were.

• 50:46Scott Tolinski?

  Yeah, the genie's out of the bottle, you're

• 50:49Wes Bos?

  Genie's out of the bottle, yeah.

• 50:51Scott Tolinski?

  I, okay, so I, what are your thoughts here, Wes? Because I have some thoughts on the survey specifically. So here's what I would love to see. Hey, I want to know what you would want to see from the improvement of this survey overall. But I know Sasha, he's a great dude. Here's what I would love to see this survey improve on for the next one. I think it's missing a huge part of the actual AI landscape. It's missing things like skills, it's missing things like MCP. What are the most impactful AI features that people are actually using?

• 51:25Wes Bos?

  What are you actually using?

• 51:27Scott Tolinski?

  It's more than just model or agent. There's what skills do you think are effective? Do you think skills are more effective than MCP? Or do you think subagents? Are you using swarms? Are you using like?

• 51:41Wes Bos?

  Where do you run the code?

• 51:44Scott Tolinski?

  Yes.

• 51:45Wes Bos?

  How like, are you monitoring it? Because like, are you, are you just firing off several agents at once and not monitoring? Are you coming back to it? How are you kicking them off? Are you doing it from your phone? I think that this survey missed a lot of like, yeah, it's interesting that people are using OpenAI versus Claude and whatever. But I think like, how people's workflows have changed in the last six months have, has been the most interesting part of this landscape and how people are tackling different types of issues. You know, like PyAutoResearch, absolutely phenomenally interesting tool of like, these are new, these are new ways to solve problems.

• 52:28Wes Bos?

  Not like a lot, I think a lot of the survey was simply just like, we used to write code by hand. Now we are having the bot write the code for us and not just like, how, how are the new problems being solved here? I think that's the more interesting thing. Like, how are you rigging up your, your, your whole thing? What's MCP, the skills, all of that good stuff?

• 52:52Scott Tolinski?

  Yeah, I totally agree. Yeah, I even hear like strategies would have been useful. Like, what are people actually, are you using ARD documents? Are you using just markdown files for documents? You throw in a whole docs thing? Are you using like a task manager or something like beads? Like, these are the types of things that I think need to be in this survey if it's going to be useful. So Sasha, I, I, I hope you take this criticism as critique. I think you. This is only the second time they've done this survey. There's probably a lot more that can be learned from this. So I would love to see the survey be a little bit more representative of how people are actually

• 53:33Scott Tolinski?

  working in AI beyond just, you know, opening up their VS code and using Copilot or whatever.

• 53:39Wes Bos?

  Yeah, but like that, that's why I want to see our people, our developers actually ripping three agents at once, or is it just like us who are at it? I would, I would like to see, are people actually using MCP servers and skills? Or are most people just simply opening up, typing in the box, reviewing the output? You know, there's, there's a big difference between that. And I'd like to see some real world stats on it.

• 54:06Scott Tolinski?

  How much code do you review by hand? How much code are you actually reading that your agent writes? Those types of things? Yes, absolutely. Cool. Well, do you have anything for us before we get out of here, Wes?

• 54:16Wes Bos?

  That's it.

• 54:18Scott Tolinski?

  That's it.

• 54:20Wes Bos?

  Thanks for tuning in. Peace.

• 54:22Scott Tolinski?

  Peace.