The State of AI for Web Development

So hi, everybody, I'm Sasha, and just like CJ said, I'm here today to talk about the state of AI and more specifically, AI for web development. So you might be wondering what my qualifications are for speaking about such an important and complex topic today and opening this conference, pretty much. And yeah, I don't really have any. I'm not an AI researcher, I'm not an AI expert, I'm not even that big of an AI user myself, to be honest. But the one qualification that I do have is that I run online surveys, such as the state of JS, also on my t-shirt. Maybe some of you have taken that survey before just by a show of hands.

So yeah, a few people. Thank you, by the way. But there's also another survey called the state of web dev AI, which just ran for the second time earlier this year, collecting over 7,000 respondents. And in fact, the data is available right now at 2026.stateofai.dev. So I want to talk about this survey today, but I actually live in Japan, and so I took a plane here. On my way to the airport, in the taxi, there was an advertisement for an AI product. And then at the layover in China, another advertisement for a different AI product. Even in the plane on the entertainment system, another AI advertisement. So I don't think you need me to tell you that AI is all around us these days. Also look at the program for this conference, right?

But the survey... Is it working? No? Okay. The survey definitely confirms this trend of AI usage increasing. So one question was asking respondents what proportion of the code they produce is generated through AI tools. And so this grayed-out bar here is the Delta with the previous edition of the survey showing a decrease over the last year, whereas the white bar shows an increase. And as you can see, the chart goes from 0% AI generated at the top to 100% at the bottom. And the chart pretty much flipped year over year. Last year, most people were in that top half of the chart generating maybe 20%, 25% of their code using AI.

And this year, a lot more of them are generating 80, 90%. Another question that also shows the same trend, how frequently do you use AI tools to regenerate or refactor your code? And you can see that especially people who state that they use AI constantly, that has nearly doubled. And I think that goes to show that AI today is really present in our workflow and embedded in our workflows through IDEs, through coding agents in a way that it wasn't before. So along this AI usage increase, there's also a cost increase. And Sam Altman recently gave an interview in which he stated that whereas at the beginning of the year, AI costs never came up as an issue, today they have become a huge issue. And I think, again, a lot of it is due to agentic coding. Before we would just type in a prompt in ChatGPT or whatever and spend a couple thousand tokens.

Now it's an order of magnitude more. And that shows in the survey results as well. So this question isn't even about enterprise spend, but just personal spend. And you can already see that the amount of people who don't spend anything on AI has gone down. Whereas people who spend between $100 and $500, that has gone up dramatically.

Another question that shows the same trend was asking respondents what their top pain points were when it comes to AI tools. The number one here is hallucinations and inaccuracies, which shows a huge increase because it wasn't in the last survey. So all of it is new. But then if you look at the next pain points, all of them have decreased, meaning it's gotten better, except for one, which is rising financial costs.

So we've seen that costs are one of the issues facing us today when it comes to AI. And so I didn't want to just talk about all the wonderful things that you can do with AI, which we'll hear a lot about today. I also wanted to talk about the other side of the coin, the AI pain points, issues and risk factors. So along this AI usage increase, there's also a cost increase. And Sam Altman recently gave an interview in which he stated that whereas at the beginning of the year, AI costs never came up as an issue, today they have become a huge issue.

Yeah. So we've seen that costs are one of the issues facing us today when it comes to AI. And so I didn't want to just talk about all the wonderful things that you can do with AI, which we'll hear a lot about today. I also wanted to talk about the other side of the coin, the AI pain points, issues and risk factors. So one of them mentioned in the survey is the issue of cognitive decline, so whether relying on AI tools will make us less skilled as developers. And you can see most people agree with that. And I think it can be an issue on two levels. First of all, for us personally at an individual level, because the more we code with AI, the more we forget or become unable to code without those tools. But also on an industry-wide level, because there will be a new generation of coders who haven't really ever worked without AI. And you can imagine how that can have implications for the future.

But the biggest issue that people are concerned with by far is the issue of job security. So whereas last year most people disagreed or strongly disagreed with the statement that AI tools were a threat to their job, again this year the chart is flipped and most people agree or strongly agree. By show of hands, who here agrees or strongly agrees with that statement? Raise your hand. So I would say a minority, but maybe by the end of this talk there will be more of you. We'll see. One reason why I think it is a threat is as one commenter on the survey pointed out, AI companies can convince employers that AI can take your job even if it actually can't. So all that matters for you to lose your job is that your CEO thinks that AI can do it, or maybe not even today, but will be able to in the future. And again, when asking people about general AI risks and issues that they're concerned about, job displacement came in number one. Number two was the military use of AI, which makes sense because the survey happened during the Anthropic and Pentagon controversy. And number three is environmental impact, which I think is a very important issue but maybe not addressed enough.

An interesting thing that the data showed is that there is a gender divide when it comes to this issue of environmental impact, with men ranking it as number three, but non-men ranking it as number one. And so this kind of survey can't really tell you why respondents answer one way or another, but I did want to highlight this because I think it would be an interesting area to study more in an academic or social science context. Okay, so I kind of ran through the data because I didn't want the whole talk to just be me pointing at numbers on a screen, because that's a bit boring. And also, like I said, you can go online and check out the survey results for yourself right now. So for that second part of the talk, I want to talk about the future. Maybe not so much how I want things to go, but just one vision of how they might evolve. So who here knows about Moore's Law? Quite a few people. So Moore's Law states that every 18 months, basically, computing power increases, doubles, something like that, and it's held true for most of our lives, right? So we've all grown up in a world where next year's computer is going to be cheaper and faster and better than this year's. But there's signs that this might not quite hold true anymore. Like I said, AI costs are increasing, and that's having impact also on memory costs overall, on hardware costs, even in other areas like video game consoles.

So who knows if the AI of the future is going to be cheaper and faster. One thing I do know, though, is that AI companies are betting that it will. So OpenAI, Anthropic, Google, they are all acting under the assumption that AI is going to follow that same path of getting faster, cheaper. So for the rest of the talk, let's buy into that assumption and let's bet on an AI future where AI is basically everywhere and instant and almost free. What would that look like? Well, one thing we might have in that future is self-evolving code.

Now, Sentry is a sponsor of this event. Sentry is also a sponsor of the survey, but that's not why I'm talking about them today. It's just to talk about this product that they have called Sear. And Sear, what they say is that when a new issue hits, it will investigate, find what broke, and by the time you open your laptop, there will be a PR waiting. So this already exists. But if you keep following that train of thought and bet on AI getting faster and better, well, you can imagine a future where AI will detect your bugs, it will write the PR and it will approve the PR, all of that nearly instantly, basically.

Just like us humans, we have an immune system that can identify threats, viruses, bacteria, and address them and maybe even remember them for the future. In essence, apps could develop a kind of digital immune system that lets them evolve to deal with future threats or just bugs. Just like our human bodies can sometimes have allergic reactions to things that aren't really that dangerous, well, maybe apps will also decide to overreact and rewrite the whole thing in a different programming language just to fix the little bug. And so maybe we'll have safeguards in place to avoid that. Or maybe we'll just get used to it and that will be normal.

It'll be like, what language is the app written in this week? Has that changed over time? What I think is interesting is that you can imagine a spectrum that goes from deterministic systems to more probabilistic systems over time. And our human bodies, by default, are very probabilistic, right? It's hard to know what's going on. It's hard to reason about them. But our whole goal with medicine is to try to make them more like software, more like computers, right? So we want to have a pill that can cure our illnesses, maybe turn off our appetite, maybe regrow our hair. That would be nice.

So it's interesting that over the same time, software might follow the opposite path, where it's getting less deterministic and behaving more like an organism, which also means it gets harder to reason about and control. And yeah, this is, again, very speculative. But if you keep following that train of thought, well, you might also end up with personalized codebases. So, right now, Y Combinator, the Silicon Valley incubator, they have a request for startup called Dynamic Software Interfaces. And basically, a request for startup is a thing where they say what kind of projects they'd like to fund. And that request for startup states that coding agents have now gotten good enough to allow users to customize the software they consume.

What I mean by that is you can imagine having, instead of code, code becomes almost like an artifact of the production process for apps. Instead, you have a source prompt that describes what the app is supposed to do, and you can combine that with user profile, including the user's entire identity, because, of course, the AI will know that, but also if they have any disabilities, maybe they're colorblind, maybe their hobbies, preferences, maybe even their social graph. And if you combine those two things, well, then you get a codebase baby that has traits from both parents, and that's totally individualized to yourself.

A bit like we all have bodies. We're all the same human species, but we don't necessarily all look the same or behave the same way. And again, it's also very speculative, but if it does come true, you can imagine how much that would impact our jobs as software developers. And if we keep going forward into the future, maybe the very idea of code would cease to exist, being instead replaced by digital DNA.

What I mean by that is you can imagine having, instead of code, code becomes almost like an artifact of the production process for apps. Instead, you have a source prompt that describes what the app is supposed to do, and you can combine that with user profile, including the user's entire identity, because, of course, the AI will know that, but also if they have any disabilities, maybe they're colorblind, maybe their hobbies, preferences, maybe even their social graph. And if you combine those two things, well, then you get a codebase baby that has traits from both parents, and that's totally individualized to yourself.

Now, I'm not saying all of this will come true, but if it does, I think that kind of leads us to having AI as our operating system for our life, basically. Because today, we have operating systems which run apps, and one of these apps might be a browser which itself runs web apps, but why do we need all of this if we have AI, right? Because after all, if everything else comes true now, if you want to see the weather, you can just ask your AI operating system to show you the weather. Even if you want to play a game, you can ask it to implement the game to your own specifications.

And so who knows if this will happen, but what I do know is this would pretty much become the choke point that controls our entire digital lives. And this would give a huge amount of power to whoever implements it, so I'm willing to bet that people right now are working towards that vision, whether it comes true or not.

So I've spent a lot of time in my career working in Photoshop, retouching images, designing UIs, but these days, you don't really need Photoshop for a lot of things, right? If you want to retouch an image, you can just load it in ChatGPT and type in a prompt and get the results, right? So I think this is a clear example of how AI can make certain aspects of software pretty much obsolete. And so, where does that leave us?

Oh, sorry. Sorry, what did you think I meant by that? Yeah, it's hard to know. Some of you, after hearing this, might be inspired and ready to go out and bring on this new future, and some of you might be horrified and ready to go protest the construction of any new data centers. Speaking for myself, I have to say I'm pretty happy with the way things are today, and I don't know if I'm especially anxious to give AI companies even more control over our lives compared to what they already possess. But my goal today isn't really to tell you if AI is good or bad. My message, I guess, is more to encourage people to have an open mind, and whether it's through looking at data or just through your own speculation and thought experiments, keep thinking about these things critically, whether you like where it takes you or not.

So if after all of this you still want to hear about AI and learn more, you can go check out the survey results, 2026.stateofai.dev. And I hope despite all of the things that I've said, you can still enjoy today's talks, which hopefully will be more about the cool and interesting parts of AI. And with that being said, thank you very much. Great talk, and thank you for making this survey. I have read the results. I've done the State of JS survey and State of CSS.

But the first question is, how much time has passed between the last survey and the most recent one? So how much time for those increase and decrease in bars that we've seen? A little over one year, so maybe 14 months, I think. OK. And I did see a couple of questions asking if you had years of experience on there. Let me see. Yeah, so there's a lot of data you collect. So I think it's interesting to mix and match the data in a lot of different ways. This question is, why display the AI usage as a whole and not by years of experience? It would be interesting to see the differences between junior, intermediate, and senior developers. So do you have that data?

Yeah, so I do have that data, and I did look at it. And it's pretty much what you would expect, that junior developers tend to rely a little bit more on AI. But it wasn't that significant. I think there is a chart that shows that maybe people who rely a lot on AI have one or two years less of median experience compared to those who rely on it less, if that makes sense. Yeah, it's about what you would expect. And another one here is, what reported metrics surprised you the most compared to last year's data? I think usage, the one I showed. I did think usage would increase, but I didn't think it would be that much.

And another one here is, what reported metrics surprised you the most compared to last year's data? I think usage, the one I showed. I did think usage would increase, but I didn't think it would be that much. And generally, I think the impact of coding agents and cloud code was very clear throughout the survey. And that kind of took me by surprise, because like I said, I'm not a big AI user myself, so I think I wasn't aware of how big of an impact it's having in such a short time.

Right. And yeah, and that's exactly it. So just in 14 months, I think the end of last year is when a lot of agentic flows started happening. And even I, I've been an AI hater a little bit, but even this past few months, I've been trying to use it more. I mean, if even Sam Altman was surprised, I feel like you can't blame me for not anticipating that. Definitely. So a lot of these questions are getting a bit more philosophical and a little more like, let's really talk about this AI thing.

So this first one is, why do we embrace the term AI? It's not intelligent. Why don't we use the term LLM instead as a more precise one? Do you have any opinions on that? Well, I feel like if you talk about LLMs, an easy counter is like, oh yeah, but you know, LLMs, they're just like one thing. I don't know, I feel like LLM, you can risk getting bogged down in the technical discussion about that specific implementation as it stands today. Whereas I wanted to talk about AI more in a more philosophical sense, and I feel that's still the best term for it.

Yeah. I mean, it also, at this point it's kind of a marketing term, right? Because you just throw the word AI on anything. That's what people know about like... Yeah. And I read a book recently that mentioned instead of calling it artificial intelligence, we should call it alien intelligence, like, because it's totally different from like human intelligence. It's just a completely different thing. But yeah, it's curious, very curious.

All right. Let's find another one. Yeah. Another philosophical, maybe a hot topic. Since AI models are becoming so much better each year, what's the point in still learning how to code? Do you have any opinions on that? I mean, not much point. You can learn to code. I guess it depends if you want to learn to code because you like to learn to code, or to find a job. If you want to find a job, I think there will be a big contraction where the industry will get hollowed out. Hollowed out. And you'll have like consumers on one side and AI providers on the other. But I'm really afraid that anybody who's in the middle providing anything else will have to find a job at OpenAI.

Yeah. I mean, I do. So I've taught people how to code in my career and I do think that it's still necessary because I mean, if you think about when calculators were introduced, I'm sure there were doomsday people saying, oh, we don't have to do math anymore. But if you go to school to be a regular engineer, like building engines or cars, you still have to learn the things under the hood of how to do math. You don't just depend on the calculator. So I still see it as a tool. And I think knowing how to code is only going to make you that much more effective. If you know how to code, you can prompt better. So why not still learn under the hood? I want to be clear that I'm not saying that learning to code is not useful or that developers are going to disappear, but I do think that the number of places where developers can work will shrink. It's true. All right. Let's find another one.

Yeah. In a similar idea, what skills do software engineers need to adapt to or learn to thrive in the AI future you described? Well, maybe something like product management, because at least for now, you will probably still need to tell the AI what to build, tell the LLM what to build. But like I talked about with personalized apps, there will be a point where more and more users tell AIs what to build. And so I'm not even sure if that has a long shelf life. Yeah. I mean, I like that you brought up that point, because it got me thinking. AI makes things less deterministic, right? You can't expect what it outputs, but can we create frameworks that make it a little more deterministic so that if there is this future where end users can modify their code base, well, what if the code base is a little more modular or we've built in ways of allowing the user to modify the system itself? I guess the Herbie's agent or some of these agents people are using that modify their own code to add new features, yeah, it's a fascinating future to think about. Right. Yeah, this is a hot topic as well. There's been a lot of news about increase in price, and you kind of talked about it at the beginning of your talk. But Khan says, how do you envision the future of AI developers if the prices of AI models drastically increase? Oh, that's the best scenario. That's the dream. Right. Then we can all go back to doing things the good old way. All I said will be completely irrelevant. So I'm hoping that it happens. I think for me, the bad scenario is where AI costs decrease, and then I will have to figure out what to do. I imagine there might be a shift towards local models or smaller models. Yeah, that too. That's a very good point. We have open source models and Chinese models, local models. So there is quite a few alternatives, rather than just open AI and Anthropic and Google. But, yeah, who knows? Who knows? We'll see. Yeah, and we haven't really seen like a Opus level or a GPT-5 level open model yet. So if we're dependent on these more capable models, we might have to be dependent on larger corporations. I mean, I think it would be a good middle ground if these top-level models become too expensive for most use cases, but we are still able to use local models, open source models that maybe don't do everything and don't enable the future I was talking about, but still improve our lives as developers. That would be pretty cool. Yeah. I've been trying to think of AI more as a pair programmer than a full replacement for me. We trade off, like, I write the test, it implements, or it implements, I write the test.

There are ways to use it where it doesn't take over but makes our jobs easier, makes our code better. That's a very narrow path to tread.

Are you worried about AI-generated code that might lead to tech debt crisis if developers copy-paste suggestions without understanding the architecture? How can the community mitigate this? I'm going to start with AI. Yeah. The answer is more AI. Yep. Definitely, the answer is going to be more AI. Yeah, I will say from personal experience, I've created agentic workflows where AI validates AI. So you have one agent implement, another validate, another do a security review. Obviously, things are going to fall through the cracks, but I mean, I think even with real humans, things fall through the cracks as well. Yeah, I think the more checks and balances you can add, the better, just like with real humans.

After years of developers thinking their job security and desirability was untouchable, can AI be considered a wake-up call? And what's your point of view on the social impacts of AI? Yeah, I think it is a wake-up call. And what I find really ironic is that we are the ones who are funneling all our resources into the companies that are going to make us obsolete. Right. We did this. Yeah, we did this ourselves. We were too good at our jobs. So yeah, and social impacts, yeah, I think it will have social impacts. And, you know, one of Sam Altman's ideas is giving people AI tokens almost like as a universal basic income, like a universal basic form of tokens. So one thing I didn't talk about in my talk because it was too dystopian even for that is the idea that AI tokens will become like the global world currency, like kind of like we hope Bitcoin will become. Fascinating. Yeah, and so, yeah, the social impacts could be, you know, as dystopian as you want to imagine them, there's a way to justify that scenario. So yeah.

So yeah, and social impacts, yeah, I think it will have social impacts. And, you know, one of Sam Altman's ideas is giving people AI tokens almost like as a universal basic income, like a universal basic form of tokens. So one thing I didn't talk about in my talk because it was too dystopian even for that is the idea that AI tokens will become like the global world currency, like kind of like we hope Bitcoin will become. Fascinating. Yeah. And so, yeah, the social impacts could be, you know, as dystopian as you want to imagine them, there's a way to justify that scenario. So yeah. Let's hope not though. All the AI hype is just feeding into the companies that are giving us AI, so I don't know. This one is interesting, I guess more about how we feel when we use AI.

This says, with the use of AI, I feel like a fraud shipping code at my company. Have you ever felt that way or thought that way? Well, I think I already felt like a fraud before. Okay. And so, yeah, I think, I mean, I think it's good. Like you should feel like a fraud because I think those the people who are a bit resistant to AI are the ones we should encourage and kind of treasure and highlight because everybody else is just going to follow the path of least resistance. Right. So I think that that feeling of being ill at ease, you should treasure that. And kind of once it disappears, that's when you should be worried. I love that. Yeah. So stay vigilant, be persistent. Don't, don't just give into the AI because yeah, that's, that's totally if you're skeptical, you're doing your job like you're feeling like, oh, this code could be better. I could probably write a little bit more by hand. So yeah, I like that way of thinking.

This kind of gets at what you were talking about with like the AI OS. So how will AI replace the OS browsers and everything else? Do you have more thoughts on how that might work or what it might look like? Yeah. Again, I feel very under qualified to talk about any of that, but I do think, you know, it's such a big thing to go after, like literally controlling how you interface with the world that I think people will try to make it happen. And you can really imagine how, well, okay. Google, they announced search results moving to AI search, right? So the implication of that is basically no more websites because why would you go to a different website to look at information if you can have that right on the Google screen? And the next step is you don't even need to go to Google because your Android phone is the, you know, AI interface, whatever it is, is Gemini. So I see it kind of as a progressive cannibalizing, kind of expanding out, you know, like in The Last of Us, the fungus is kind of conquering everything.

Maybe that's how AI slowly grows. Let's hope not, because that's a bit of a doomsday way to end here. And again, I'm not, my goal wasn't to come here and say, this is what's going to happen. It's more like, let's think about the consequences. Let's think about the implications of what could happen. So that we are able to prevent it if we don't want it to happen. And we can prepare and start thinking about it now. Fantastic.

Well, that's all the time we have for questions. Thank you so much, Sascha. We appreciate it. Thank you. Thank you. Thank you. Thank you.