AI and Vue.js (with Daniel Kelly and Patrick van Everdingen)

Welcome to DejaVue.

It's your favorite Vue podcast. You just don't know it yet. I guess I mean, maybe some of you are here since the very first episode, and, we are live with the panel at Vue.js nation 2025, everybody. Yeah. Michael, how do you feel about that?

I'm feeling great. We're going to talk about AI and Vue, and it's going to be awesome. So if you have any questions for these fine folks for this panel, put them in the chat, and we might be able to get to them. And we'll, yeah, we'll see where this whole panel goes.

Absolutely. And if you wanna wait. Wait. Wait. What's what's, like, Deja Vue? Never never heard about that? Well, we are a weekly podcast all around the Vue ecosystem. We have amazing guests. We had like Evan Uon, of course, a couple times, Daniel Roe, Lead of Nuxt team, but also Matt Pocock, right, the TypeScript wizard, CJ from Syntax who talked about his, well, first true love, Vue.js.

Yeah. So our first guest is Daniel Kelly, and you may have seen him yesterday giving an awesome talk on Vue and AI. He is the lead instructor at Vue School. And so he's he's got, like, ten years of full stack experience. He does lots of teaching, you know, all of that amazing stuff. So can't wait to hear from him about, what you're what you're doing with AI. How are you doing today? Are you feeling a little bit better?

Good, Michael. Yeah. Yeah. Doing pretty good, man. Doing pretty good.

Perfect. And, of course, with, like, we can't only do it with one guest. We have to have two people on here at least. And our other lovely guest is a full stack developer. He has a knack for TypeScript, really good expertise, also in GenAI.

I'm doing great, Alex. Thank you. How are you?

Good. Yeah. Psyched about the panel. And, yeah, as as Michael said, really curious what the the chat is saying. A whole big claps right now for all our amazing guests, and we should start straight away with the first question, which is a very simple one.

Patrick, you wanna take a stab at that one first?

I think we'll not lose our jobs to AI. I think we'll have, new jobs coming in thanks to AI, and we'll always have some sort of job where AI will play a big role in our lives. Yeah. I don't think we'll lose our jobs. No. What do you think, Daniel?

Yeah. I I definitely agree with that. We're gonna lose our jobs as they currently stand. Right? They're they're not gonna be the same jobs. They're gonna be same in some respect, in in a lot of the same respects. I mean, we we're we're still needed in that that development life cycle, but, yeah, they're gonna be different. It's not what we originally, you know, trained and and things for a %. Right? It's it's gonna look different.

So with these AI models getting better every couple months, we see that they're capable of more and more and and, you know, surpassing all these benchmarks and whatever hype we see on social media. Why why do you think that they aren't gonna take our jobs? Like, there's a there's obviously the other side that it's that's like, oh, they they can code better than the best humans. So, of course, like, it's only a matter of time. But, like, what would you say to that point?

Well, I think an important aspect of using AI responsibly is having the human in the loop, someone who keeps track of what an AI system is doing, whether that's be it's coding, marketing, writing content, you always need to have some sort of human approval in this entire process. I think we're doing that right now with coding. Cursor Cloud generates code for us. We check if it's valid. Hopefully yeah.

Yeah. Most certainly. And I, you know, I I think it's in in the short term, it does take over the responsibilities of, you know, the things that some people will be getting hired for now. Right? Because the people that are already on the team can do those things faster.

Yep. To add on that, a that, a while ago, there was this article, The Rise of the AI Engineer. I still remember when I read that, I was a bit skeptical. Like, what is this what's the definition of an AI engineer? Are you responsible for training the models? Are you responsible for implementing those models in in software?

I've heard it said, in the past before that software developers are not programmers. They're like, we we solve problems with technology, and writing code just happens to be the output of that. It's like an implementation detail. And this was, like, well before, this the AI in the last couple years. And I think that that's, that's that's still true. And it's just like coding is just one of the the things that we do, and maybe that changes. We we're wrangling AI a bit more and more over time.

Yeah. And I think it's also a key part, Michael, what you just mentioned. Like, not the programmer to see as a code monkey. And hopefully, maybe that's, let's say, annoying part of, like, I don't know, that the boring tasks can be automated away and be stripped away like it is right now of, like, okay. Maybe scaffold some test cases for us.

Yep. Yeah. Garbage in garbage out, they say. The quality the output of the quality, the quality of the output of your generated code is as good as you how good you prompt, the AI, how well you communicate, what the context is of your project, how, what what it needs to do. And my experience with these kind of things is the longer your the the task is that you define that the AI needs to solve, the higher the probability that the output of the code becomes, less usable.

Yeah. For sure. And it's it's kind of like having, a little bit of the I mean, extended memory as less so than, like, another strategic part of your brain. Right? It it helps you remember where some things are and what patterns are kind of in the code, but it's not necessarily always thinking strategically as a as a human could.

Yep.

So we'll go to our next question, which is how have you integrated AI into your daily workflow when you're writing code? Why don't I go to you, Daniel? You gave a talk yesterday. I think you covered some things about this. Maybe you could give just, like, a couple things from that talk. Don't don't redo your whole talk here.

Yeah. Watch watch the talk if you haven't seen it. Yeah. Exactly.

Honestly, for for me, so in content creation, I don't code every single day. But when I do, a lot of the ways I'm using AI is to brainstorm, directions for for my app. Like, you know, if I'm I'm building this, you know, example example application for some, you know, video series, what, you know, what features can I build in this app to showcase the the the concepts that I'm teaching? Right? And it can help me say, okay.

Do you trust composer mode with the YOLO mode enabled?

Have you heard of it?

No. I do not.

Maybe

Have you tried it.

Give maybe a little context to everybody out there who's not knowing what the composers or YOLO modes of people not using Cursor. Maybe that's a a good part there, and then I'm happy to hear more about experiences with that.

Yeah. So, composer mode is a feature within Cursor IDE, which in its term is a fork of Visual Studio Code. Looks exactly the same as Visual Studio Code, has the same extensions, plug ins, and it is a feature that allows, a large language model to, I would say, semi autonomously, construct a new feature for you, and that is not tied or limited to just code generation in a file that you're working in. It also allows a large language model to, access the file system, create new files, run stuff in the terminal. So, I think it's a really innovative feature that kind of helps you get in the right direction given that you have enough, supervision.

Forty two forty two new files. That's that's wild. Yeah. No. I I I haven't tried it. I the fact that it had access to the to run commands scared me, and I just didn't wanna trust it to do that. And I didn't wanna take the time to, like, learn how to sandbox it or something. So yeah.

It's also, like, spamming is like, oh, r m r f slash oh, oops.

Yes. Right. Yeah. Yeah.

Or, like, oh, just the wrong prefix. I want to remove the component.

There's a checkbox that says disallow, you know, removing files. So they may

Ah, okay.

Built that in to to it so it's a little safer.

Than it can only execute some malicious things. That's good. Yeah.

Exactly.

Yeah. Yeah. But at least that's a that's a good way to, like, make sure that not everything is lost accidentally. Like, oh, yeah. The git folder is not gone too bad.

Yeah.

But, I mean, it also fits in a way what what you two of you described of, like, okay. AI is useful if you have the supervision if you check your code. I think maybe that also comes pretty nicely in, with another point, which is what is your best tip or advice for, like, other developers doing AI assist development in in Vue.js specifically? Besides, of course, check what the output is. Mhmm.

I would say just, you know, take it in small chunks. And we kinda already said that. Right? Don't try to do everything at at one time, and this is going to be the most helpful thing. Remember, you're still the brain in the operation. It's just it's just helping.

Yep. Yep. I, agree. I agree. My advice would be, use Cursor's documentation feature.

Also expensive, I heard.

Sorry?

Also expensive, I heard, or at least compared to other models.

Quite alright nowadays. I, have a Cursor subscription, $20 a month. It basically gives you 500 fast requests with Cloud Sonnet, which I think, yeah, suits me pretty well. Yeah.

They they have, so called, like, premium models or a few that you can choose from, and Claude's no more than the, than the latest OpenAI model on there.

Yeah. That's nice. Yeah.

And what's also good to know is that, Claude is also available within Visual Studio Copilot as of a couple weeks ago. So if your company has a, GitHub Copilot subscription, you can always, switch from GPT to to Claude.

Yeah. That's what I did for most of the things because I'm mainly using GitHub Copilot and VS Code, and then it's like, oh, yeah. Claude, let's try. Gives better results, actually. So I was really happy about that that it happened.

Perfect results.

What I kind of dislike about the versioning of these large language models is that, they kept referring to Cloud 3.5 Sonnet as Cloud 3.5, and then they released a new Cloud 3.5, which they didn't call 3.6. They just called it 3.5 new. So, yeah, ensure that you refer to the right version of Claude, which I think has a date, time stamp in the model itself. So, always check because your your experiences may vary depending on the model that you're using. It's sometimes confusing to, yeah, in large language model land.

It also sounds a bit like versioning strategies should have been, well, improved a little bit. I mean, we, at least in software development, had have figured it out kind of with semantic versioning. But there are also suggestion other suggestions. But, yeah, that's good.

Jsut ask AI to what to name it.

Just ask. That's a good idea. Yeah. Alright. There is one question.

The way that we create docs visually for people to ingest, yeah, I think the need for that does diminish some, like having a website you can go to with docs that look nice. But does your app still needs doc does your app still need documentation? Yes. Absolutely. I mean, that's how you're teaching the the language model about your your app in a lot of ways.

And do you think that, like, from a maintainer's perspective, couldn't you just say, okay. I have a lot of tests, and the code is expressive enough. Then, like, as you just said with, like, okay. If AI generates the code or, like, a AI reads through code base and generates docs from that, do you think this will have, like, a similar, let's say, impact in writing the docs by hand manually right now?

I mean, I think part of it could, but, imagine writing, a a super expressive expressive variable name versus writing a comment. Sometimes you've got variable names that are are good enough, and then you've got variable names that are precise. And a lot of those variable names that are really precise are actually a pain to write every time. Right? And so I I kinda see the difference there being writing things out in the comment doc style versus and and capturing more of the essence of what's happening versus, you know, something that's that's happening in in the actual code and variable names and things.

I'll say from, a content creation, like, technical writing standpoint, I've experimented a bit with trying to get AI to help with writing articles even or anything like that. And it's not very good, especially compared to, like like, really well written documentation. It just doesn't have, all of the nuance that you would expect or references to, like, oh, like like, design decisions, like view in the view docs, it might say, well, we did this because of this, or it'll reference another part of the documentation. And in theory, you could build a whole system that, like, that that work does this all, but it's would be very complex. I think the best use is in that, like, question answering, like, dynamic.

Yeah. I also made similar experience, actually. Like, it's especially because it also doesn't have all the context. If you would have let's say, you start a project and you feed everything into, like, the context window of the l m. Like, here are all our transcripts from our meetings and why we did certain decisions, which no one ever have, like, from years ago.

Well, I think the limitations of the context window might be a problem that gets solved in a nearby future. I mean, when compared to GPT 3.5, which I believe had a context window of 16,000 tokens or 32,000. Whereas right now, we have models that are capable of handling, I think, 256,000 tokens to a million tokens, I think. So I think, well, the technology is, gets better. We can stuff more context in the context window, so to say.

But you wouldn't do that.

No. I wouldn't do that. No. It you there are people who might have this, preference. You can also stuff kinda all kinds of decisions in there. You're not limited to just code as a language, but you can also, yeah, put normal English in there. It will help you, steer the the output of your code. So, I think we're getting there, that we're able to stuff all kinds of things in the context, and it doesn't matter how big the context is. But we'll see.

I guess the problem is also, like, you have to have the information somehow. Right? If it's only existing in the heads of the people, then you can, well, stuff in. But that's that's a good point to mention that, like, the the size the window size problem might be solved at some point.

Yep. Yep.

And in terms of, my apologies, Michael. Go ahead.

No. You you have something to say. Go ahead.

Well, and it's just as far as docs are concerned, and maybe not even necessarily docs being used for a context for, the AI, but just docs being useful for the developer. You know, one of the reasons we well, I say we we all like TypeScript, but one of the reasons that people who like TypeScript like TypeScript is because it keeps you from jumping out of your IDE and to your browser to documentation as much. Right? And so I think, AI will allow us to remain inside of our IDEs even more. It already is.

Maybe that's also a good point to mention that some frameworks, for example, I know Svelte is doing that, is providing LLM friendly documentation, just like an LLM dot TXT file with, like, hey, this is Svelte. Here's how you do things. A bit like how you could build your custom cursor rules saying, hey, this is how you use certain methods of the framework you use, but by the framework by default. So maybe that's a good segue.

That's a very good question.

It's difficult. Right?

Yeah. Yeah. How can yeah. It's not easy.

Maybe while you're thinking about it, just to add on that, especially nowadays, AI is really good at generating React code, obviously, because it has a lot of training data on that, given the, like, big React ecosystem and a lot of AI tools targeting React by default. v0, for example, has Vue support by now, but, had mainly React support before. But on the other hand, we can't just say, hey. We only use the most popular tool forever now just because AI is good at it, let's say. And I mean, AI can also generate valid Nuxt and Vue, and so on quote.

It was a while ago, I read an article by, written by some folks at Anthropic, the the company behind Claude. And they, had some suggestions, to get better structured output from, from Claude given a certain prompt as input, and they recommended placing, like, structuring your prompts in XML tags. And it could be XML tags that were nonexistent, for example, instructions or context. And I can't remember the name of the research paper anymore, but it mentioned that it could greatly improve the quality of the output of an LLM. So my first thought would be, when writing documentation for frameworks and libraries, place code examples in a XML tag named coach example or, a discussion tag if you have a controversial take or a hot item of something.

Yeah. I think that's very interesting. I think maybe more people will even rely on the more traditional markdown style, documentation that you see on on GitHub. Right? And you won't have a link to a web page from the GitHub docs or Readme. You just, expose that that markdown file easily so you can add it to the context of your your ID's composer, and, you've got a lot going for you right there, I think.

Yeah. So, basically, the way having it all structured, so to say, either for, like, XML text, like, okay. I have my own, let's say, structured system or well, Markdown is also structured in a way, and not for, like, rely on fancy websites with, like, big animations. Like, here's here's the info as a dumb for maybe the LLM to learn. Yeah.

Yeah. Yeah. But but yeah. I mean, to your point, Patrick, could have have even more structure over the current, you know, just markdown format being headings and things like that using using XML to really tell it more semantic things. And maybe if there becomes some kind of standard around that, I don't I don't know. That could be interesting.

Yeah. Yeah. I've seen recently, so there's this l m dot t x t file that's, that people are trying to popularize for websites. So it's similar to, like, your robots dot t x t file where, like, crawlers will will read that file for some information. But, basically, this file is, like, take all of the content of your website and just stuff it into a single single file.

Mhmm.

Then you can just put it in there. So I wonder if that combined with well, you could make this l m's dot t x t file structured with the, the XML tags and stuff like that.

Yep. Yep.

That's also what Svelte is is doing. For example, I have, like, a system tag making it saying, this is the developer documentation for, I don't know, for example, SvelteKit. And then they have markdowns, like, nodes and, and so on so on. So it's kind of a mix. It also goes in that direction.

Yeah, we have a couple of questions about environmental responsibility. So, I'm gonna pop up this one. Yes. Where does the environmental responsibility intersect with our adoption of LLMs? Because we know that these LLMs, they're using tons and tons of energy and, even, like, Meta investing and building their own nuclear reactors and, like, you know, crazy things like that because they're just so power hungry.

And before you answer, AGI will solve that is not a valid answer. I'm not making it easy here.

I would say before you consider using AI or implementing AI in existing software or applications, try to think, do I need AI? Do I need AI? There are these examples I see online, and it's primarily, within a realm of AILM engineering called agentic workflows. They often give, like, an example of, okay. Let's build an agent.

Well, going off of that, Patrick, the the, thing you mentioned about training being one of the most energy, you know, heavy parts, the the deep seek model, like, I saw someone just, mentioned in the comments as as well. You know, it took them far, less time to train this deep seek model. I mean, this exploded on the Internet, like, like, Monday, and they did it with, with with graphics cards or, you know, that were extremely they were the old stuff. Right? And I imagine that uses less power as well.

Yep. Pretty good points. And I think you're you're touching on two topics that we also had in mind. I mean, we can stick with DeepSeek for a little bit. Do you have a take on on DeepSeek r one, on the the big thing in AI recently? Well, a lot of big things, but one of the most recent ones.

I'd say, DeepSeek, it's pretty impressive that you can run a reasoning model because that's what it is, a reasoning model that, is able to compete with OpenAI's flagship reasoning model, which costs, I don't know, $200 a month, I think. I think it's They

paid no profit on that. They're not profitable in a turn dollars. That's the craziest. Yeah. So

I think it's very good for our field of work that these new innovations, which, drive, innovation and, lower costs are coming out. I think it's awesome that you can just run it yourself. It helps you with certain reasoning tasks. Given the Vue.js ecosystem and coding, I've tried to use it to code. I think I'll stick with Claude.

Oh really? I think it's still a bit faster.

And, for example, I try to, give an example, create a server route, use use a async data, and it just made up something with Express JS. And I'm like, yeah. It doesn't it's probably not trained on Nuxt documentation, maybe briefly or not as much as GPT four or Claude is doing. So my experience, it's awesome that it's open source.

Patrick, do you have any, were you interacting with it, like, through the the chat interface, or do you have any, IDE integrations with with DeepSeek that work yet?

Yeah. So I implemented it in my cursor workflow using Ollama.

Okay.

If you use it to run local models, you can pull the latest, large language models and then, use that. So I run it in the terminal, actually. And what's interesting also is that, of course, it's a little political take, but, I believe it's also a model that's censored. You you talk about certain models certain topics, it will refuse to answer, especially in the web UI. But this is also the case with some topics that you, use in OpenAI, for example. There's always some sort of censorship.

Some sort of biased. Yeah.

Yeah. Bias, censorship, but you can circumvent this with certain jailbreaking prompts, which I think is interesting as well. So all in all, yeah, I think it's awesome that it's has come out. You everybody can use it, but I'm not using it for anything right now. No. What about you, Daniel?

No. I I honestly, I haven't tried it yet. I've I've heard good things, obviously, but, yeah, I'm I'd be excited to give it a whirl.

So our next question, and I think maybe our last question here as we, start to wrap up, is,

We still have a few, but let's let's find it.

So we'll do a question on data privacy.

Yes.

So the question is, here we go. Perfect. We're sending all sorts of data over to these LLMs. And how how do you think about privacy? Are you concerned about this at all?

So in terms of using Cursor, like, there are ways to ignore certain files, you know, prevent cursor from sending data and certain files to the LLM altogether. And I think that's a good start. But, obviously, you can't hide all the files. There else is useless. Right?

I guess, also depends on what you work on, so to say. I guess, when you work, I don't know, on a super secret government project and they're like, yeah. All the files are sent over to, I don't know, some server. Then people might be like, oh, yeah. Let's see. Let's maybe not do that. I don't know. Yeah. But, Patrick, what do you think?

Well, I know for a fact I have a couple of colleagues who work at a, financial institution here in Holland, and they have a policy of, not letting their engineers work with AI coding. As a matter of fact, everything is blocked from OpenAI. Not even ChatGPT can be used. So they are a resort, yeah, they have to use local coding models, which you can also pull with Ollama. They were quite great.

Yeah. I guess it also applies if we think about things beyond coding, like applications, including, I dealing with sensitive data, be it, like, I don't know, medical data, whatnot, then that would also be a a valid way to do that.

Yep. Yep. Yeah. Certainly. One thing you could also check out if you're not too technical, if you haven't worked with Alarm or if you can't work with it yet, there's this thing called the WebLLM project.

Yeah. The WebLLM thing is very interesting. I'll have to check that out a bit more.

Yeah. Do it.

I mean, that's just a start as well. Like you said, it's very much in development, and I guess there there are a lot of things coming no matter if, like, okay. Decrease the size while keeping the same, well, re reasoning capabilities. Yeah, versus running them. And I think especially given that I mean, I'm living in the EU, Patrick, you as well.

I think we're looking at two processes that run parallel. I think on one hand, the tech market is quite bad in general. I mean, I'm speaking for for Europe. They aren't the golden days anymore. I remember, like, two, three years ago during COVID when a lot of companies had much more money.

Or even worse, the project manager will say, hey. We can every do everything with ChatGPT, and then you need people to fix that. Yeah. Yeah. Yeah. Yeah.

Like, oh, I saw a TikTok reel. It said, you can make a website with ChatGPT. Why would I need you? Well,

not the case.

But just like I said earlier, I think there will be a demand for other IT roles, which is which could be the AI engineer who's gonna work and implement with, with all these large language models. That are there will be software engineers, front end engineers who might need to implement and integrate large language models or fine tune data, create a data ingestion pipeline. I think there are more possibilities rising right now due to the AI AI hype and boom.

Yeah. For sure. I mean, it's it's definitely going to, you know, create as many jobs, I think, and it as it takes. They're just gonna look a little bit different. And, honestly, as developers, we're poised in a very good position to take a lot of those positions because of the knowledge that we have.

Yeah. Great. I think that's a a really good, a good end word here for everybody who might be concerned that, as we said before, nobody will lose their jobs. It might just be different. And maybe there will also be some new roles out there as well. Patrick, Daniel, thank you so much for joining this lovely panel. Could you very briefly just like in one or two sentences say where can people find you or follow you just so people know that?

So I'm, on Twitter at Daniel Kelly IO. GitHub is, I might mention there's an underscore between Daniel Kelly and IO. It's on my tag there. But, yeah, that's mostly, where I'm at. Yeah. And, of course, I am my my apologies, Patrick. Of course, I'm with, Vue School and, so VueSchool.io. Definitely, come there for, we've got some AI training and stuff there as well. Awesome.

Yeah. You can, find me on x or Twitter under reg xparser. Or if that's too hard, you can look for Patrick, the LLM engineer. And I'm also on LinkedIn, with my full name. You can find me there. There's only one Patrick van Everdingen, so you'll find it.

The only one.

The only one.

Perfect. Thank you so much. And, that's almost the end of the show. Thank you so much, VGS Nation, for giving us the chance to talk a bit about AI, and Vue and how the future might look like. Thanks once again to our wonderful panelists. Also make sure to subscribe to join to next year's event, right? Vue.js Nation 2026 over at vuejsnation.com. And Michael, you got something else for us, which comes

Yeah, there's also Front End Nation, which, is coming up in later this year. June. Yeah.

Everybody. June.

Oh, you can go you can go sign up for that. There's, the website is frontendnation.com. You can go check that out.

Yeah. And, if you wanna hear more about, Deja Vue, check out the latest episodes. We have them on YouTube and on your favorite podcast platform, Apple Podcasts, Spotify. Google Podcasts is that, so not there anymore. Wherever you wanna listen to, for example, we talked about, our twenty twenty five predictions with Daniel Roe, who speak later today. So shout out to him and also with Justin Schrader about FormKit and a bit more AI. Definitely check it out.