# Panel: The Future of the World and AI

Canonical: https://davidk.dev/talks/panel-the-future-of-the-world-and-ai/
Video: https://www.youtube.com/watch?v=oCLsuoL8bsg
Speaker: David Khourshid (@DavidKPiano), creator of XState
Type: Panel · Event: Agent Conf 2023 · Year: 2023 · Duration: 25:58 · Views: 97 (as of 2026-10-19)
Topics: AI agents & LLMs, Startups & open source

## Summary
AgentConf 2023 hosts a captivating panel discussion featuring esteemed experts Tejas Kumar, David Khourshid, and Jani Eväkallio. The panelists engage in an insightful conversation centered around the future of the world and the role of artificial intelligence (AI) in shaping it. This thought-provoking discussion takes place amidst the backdrop of AgentConf's mission to provide "Unique Inspiration.

## Key takeaways (quoted from the transcript)
- Fundamentally, it's not in service of the goal of the technology to begin with, which is to actually build products to help people's lives and build infrastructure. (10:00)
- Essentially today is that you take, you know, somewhere between, like a hundred and ten thousand samples of data, and you upload it into the model as a CSV file. (13:00)
- LLM is a large language model trained on the Corpus of, just all of human text for the last, however many years. (15:00)
- Always bet on open source, like if something can be done, chances are people are going to open source it. (21:00)

## Chapters
- 0:00 Introduction
- 5:00 I absolutely, absolutely do, in a way, because, yeah…
- 7:00 Where we're no longer writing assembly, we're no longer writing C…
- 9:00 You can now use TypeScript in order to build sort of…
- 19:00 At this point, I don't necessarily think that all of these…

## Transcript
[0:00] David is a good friend. Please take a seat, David's a good friend. And, um. Founder of Stately..io A.I. We'll talk a little bit about that. And we have Statelli. It's not Stately, it's Stately, and we'll have, we'll have, Yani as well coming up to the panel. Yani, yeah, give it up Biryani, nice to see you. And of course, we have ChatGPT, who deserves no applause because it's a machine. Like, there's space for chat GBT here.

[0:00] If the cable reached, I would honestly give that space to ChatGPT. But anyway, hi, these are your panelists. Yani works at a story writing company with AI, just to be equal with David, whose company I also talked about. It's good, it's good to see you guys. How are you doing? Good? Not tired at all? Yeah, not tired at all at all. I like that. Cool, Okay, so, this is a panel about AI, your both of your talks are about AI.

[0:00] That's why, that's why we're here. I have, I think. The burning question to start with, is, you know, is my job? Is our jobs going away? Let's let's start there, and we can, we can go from there. What do you think? What's your take on that? Are our jobs going away? As programmers? Yes, exactly. Are we becoming less relevant? Are we becoming like an unmaintained open source repo? No, I don't think so.

[1:00] I mean, like, when I first started programming, I thought coding was the thing. You know, coding was the activity. Writing code was what programming was. But the more that you know you work in the industry, you start to find out. There's actually like combination of like, applying your sort of accrued expert knowledge in order to solve problems. And I think at this, you know, today, and it may be in the near future.

[1:00] I don't think the types of AI models that we have available will be that helpful in that type of problem solving. Now, what it will do is it will make our job a lot quicker. Which means that, you know, companies who do hire us are probably able to get more done with less time. Which means that they might be in the short term, maybe fewer jobs. Created in certain sectors, but I do believe that the demand will meet them.

[2:00] The sort of like the new, found productivity, but that's actually more to do with, sort of, like, the overall economy and capitalism. Will we be able to create enough jobs and create enough sort of for demand? You know, for this works to be done, but no, I think there will be, less jobs, but I think overall there will be, jobs. Yeah, I completely agree with AI being able to assist you in your job rather than completely taking it over.

[2:00] Some people estimate that AI will be able to take over, like my job, within 10 years. And I could do my job within three days, so I'm already better than AI in that regard. But yeah, we see with things like co-pilot, you still need a human in the loop, right? Because, I, I'm sure that all of you have experienced. Like, once you do something with Copilot, you actually have to go back and check it.

[3:00] Because AI right now is at the stage where it is confidently wrong, not wrong all the time, but wrong enough that you have to go back and be like, are you sure? But I am also frequently confidently wrong. Like, you know, just to kind of like, you know, argue against my own point, you know, play the Devil's Advocate a little bit. I mean, I, you know, the progress in these tools is so rapid.

[3:00] And I think, you know, if you look at Co-pilot today, it is already so impressive. But it is so underpowered for what it could be even today with today's technology. Because so much of how we use these AIS and LLMs is actually nothing to do with the capabilities of the model. But it's like, what we feed them and how we feed them. So if you look at Co-pilot today, all it does it, it uploads the code from your open vs. Code tabs.

[3:00] It doesn't know anything about your overall code base, your structure, your code style. It doesn't understand sort of the goals that you have. You know, it doesn't know anything about the design of the thing that you're trying to build. But even today, Microsoft, or anyone with access to, you know? I don't know a couple of million dollars, you know, could, you know, use these off-the-shelf models to build an engine that is much, much smarter.

[4:00] So I think we need to be also, like, very open to the idea that the development here will be more rapid. Than what we can sort of predict right now. So hearing you say that, I'm curious, you said. Co-pilot uses your open tabs only, so if I open more tabs, I get better output from Copilot. Yeah, so there's a limit, of course, like, I think, the maximum amount of, tokens, or where it's.

[4:00] It's about 8000 that you can upload to it, so I don't know how it does. The the. I think it just picks the most recent ones. But yeah, it is a very primitive, you know, style of prompt engineering is that when you go and you start working on a task with co-pilot. Open the tabs that you wanted to know about. If you wanted to know about utilities in a certain file, if you wanted to know about, you know your sort of tests or implementation, have those files open in your recent tabs.

[4:00] That's actually a little Pro tip for you all. I think you made it angry, but awesome.

[5:00] So what I hear is, it's it's not taking our jobs, it's it's helping us, it's assisting us and doing our jobs. But we need this like fine-tuning mind of the human being that is irreplaceable. So I think of when I think of this, I think of like, Pilots. I'm an aviation Enthusiast, dangerous with my personality, but I like, I like flying planes. In simulators and in VR and in real life, kind of, that like that.

[5:00] There's been AI in this field for years, right? Like Autopilot, that's where Copilot even comes from. The the naming, I suspect, and that's not to say Pilots are entirely useless. But, like a lot of flying commercial airlines is. Just like, press the button, let the machine do its job, and you, you're just there in case something breaks, an engine fails or something. Do you see this as an analogy for the future of tech and code?

[6:00] I absolutely, absolutely do, in a way, because, yeah, Autopilot has been. Around for a while in the airplanes. And in fact, Autopilot is so capable, many people don't know this, but it could take off, fly to a destination and even land itself. So the pilot doesn't even have to be involved at all, yet there always has to be a pilot in a co-pilot in the plane. Why? Because the human in the loop, ensures that no one dies right in case something goes wrong.

[6:00] And so, while programming isn't as risky to the same extent, it's not like if we have a bug in our code, people may die. It's still important to have a human in the loop just to make sure that we don't make errors. So I think that for the foreseeable future, we are going to, at the very least, act. I. I think that it's not going to take over, but we're worst case scenario, we're going to be assistants to the AI, rather than AI being assistance to us.

[6:00] Wow, I see you nodding, what's your? What do you have to say to that?

[7:00] I think this really interesting, this sort of like the relationship with the human and the machine, right? Like, you know who is helping, you know who, and to what goal. I mean, you know, currently we can only really use these things as a systems because they aren't that smart. But I just saw a demo the other day where essentially somebody fed in, you know, uh. GPT3 into a rebel. And asked it to write a code.

[7:00] And then it ran it through a compiler, it read the error message, then it rewrote the code, read the error message, rewrote the code right there until it worked. Now, of course, compilers today aren't very good. Because, you know, like, they can only add very simple, you know, detect very simple syntax errors and stuff like that. But if you can also feed this into something that can generate tests, in a way that is verifiable, for example, from some kind of cucumber kind of syntax.

[7:00] Just yesterday we saw climates. Do the lightning talk about, you know, like writing, essentially a programming language for React components. So I do think that there's a possibility here where, you know, we can at some point develop. A programming language that essentially just describes the functionalities that we want in human natural language. And then is that, you know, that doesn't translate directly to JavaScript code, but it translates the requirements that then it translates into set of tests and code that verifies that behavior.

[8:00] And at that point, I don't know what the human, you know, role in that. We're writing documents, and then are we all just product managers? If this sounds terrifying. If you really enjoy writing code, it's okay you can write code in the same way as people can still write. American sports car from the 60s, you know, like, it's a passion thing, you know, it's not strictly speaking necessary, but you can still do it.

[8:00] Yeah, I actually really like that perspective, because, for many, many years, we have layered abstractions on top of abstractions. Where we're no longer writing assembly, we're no longer writing C, I mean, some of us maybe are C, C plus, but we're writing language at higher abstractions. Even JavaScript now today, like, I don't think many people in this room write just pure JavaScript. We're writing on top of, Babel, you know, we're transpiling typescripts, and so we're adding layers of abstraction on top of it.

[9:00] And we could see Copilot as yet another layer of abstraction, and so, as Yani was saying, we could have another AI flavored layer of abstraction on top of that.

[9:00] So it might get to a point where we're literally just like, build me a chat app with websockets, and it does it. And we just need a human to curate as opposed to create, right, exactly, interesting, so we're becoming more curators and creator. Do you think this would mean then engineering teams? Even though the jobs may not go away, you would get a lot smaller because, like, you don't need. I think of like the big tech companies like Spotify that have 4 000 engineers, four thousand people to build a music player.

[9:00] And do a lot of other things, of course. Like there's people I used to work there, I can say that there's there's people whose job it is, like curate playlists and so on. But about the engineering specifically, like, do you think the teams would get smaller because at this point? We're just guiding, we're not really, I mean, yes, but it. But also like, if you, if you look at, you know, any sort of like, a science fiction fantasy story about the future.

[10:00] You know? We live in this perfect Xanadu, where all the buildings are made out of white crystal and there's flying cars. And you know, what have you, right? Yeah, and in order to sort of, like, get to that, sort of like vision, we do need a lot more technology. And I think, you know, one thing that we do tend to forget, you know, in the weeds of programming, is that, like, technology is a tool to Aid, you know, Humanity?

[10:00] And I think, you know, building software has always been frustratingly difficult, complicated and expensive. And that has been great for us as engineers because we get paid a lot of money to do jobs, um. But fundamentally, it's not in service of the goal of the technology to begin with, which is to actually build products to help people's lives and build infrastructure. And, you know, build systems that make, you know, make us all sort of like, happier, more efficient.

[10:00] The real problem here is capitalism. But I'll leave that for the next question. Yeah, I wanted to ask. Like to make get really practical, because we've been in the cloud so far, I want to make it really practical day-to-day. Both of you run company. Oh well, you're a founding engineer, but you, you, you're both very like senior in companies, and I'm curious about career development. With this age, does it make sense for people to people just getting into Tech?

[11:00] There's like, ATV serial was Julian, are you here? Yesterday was 18, there he is. Happy birthday man. Yesterday was his 18th birthday. So as people, yeah, as people, sure. As people try to navigate, especially the younger people coming into the industry, try to navigate it. Would you say? It makes more sense now than ever to, like, maybe consider a career in data science and data engineering versus if somebody you know is more like, inclined to front end or back-end engineering.

[11:00] If you're starting off in Tech, then absolutely, I think that, there is a wave, you know that, you know. It's well, it's already here. And there's going to be further waves of just more and more AI developments. And so the core of AI right now, it's it's machine learning, it's neural networks, it's data science, it's all of that. And that's definitely not going away anytime soon. Whereas we could argue that our development jobs that don't have anything to do with AI, might not go away, but be further diminished.

[12:00] Yeah, I think right now it's a great opportunity to get into that, and even as an engineer, if you've been in this field for many years, at least. Learning the basics, which I'm going to talk about in my talk, is going to be extremely helpful for navigating and riding this wave. Yeah, I mean, I, I definitely agree with that, but so I think, you know, the the structure of the industry, I think, will still be kind of inverted pyramid.

[12:00] Where, like, at the bottom of it. You will have a lot of people, you know, working at, you know, in data science, in, you know, sort of like, neural networks, you know, all of these, like underlying technologies that build the models, but the direction that we're seeing, at least right now, the trends going, is for these generalized models. Such as like GPT3, for example. Where we don't actually as individuals engineers like, they don't like OpenAI was.

[13:00] How many people when they developed this, I, I don't know, and of course, Google was many people. You know, when you know Transformers was developed. It was a small team that did it. And so I think fundamentally the. There will be a few people, relatively speaking, building out the base technology, that more people will then interact with. Now it's good to understand how the technology works, especially now that we're in a transitional phase.

[13:00] But I do think that we are going more and more towards this idea. Where, like, we will act, we will interface with very large, generalized models, with almost trivial amounts of prompting to get them to do what we want. And so in that regard, I don't think we'll all be doing data science in like, a few years, right? But it does. If you wanted to, then take a generalized model and fine-tune it.

[13:00] You would still have to have some data science, well, again, like, I mean, fine-tuning process. Essentially today is that you take, you know, somewhere between, like a hundred and ten thousand samples of data, and you upload it into the model as a CSV file. I mean, how much data science does that take? Yeah, that's it. I mean, look, of course, you know, there is, you know, you need to understand, sort of like, you know, sort of how the, you know, the weight tuning, you know.

[14:00] Works. In order to sort of guide it the right direction, you need to be able to measure the results of it. So I think there's elements of like, data science and statistics there. But fundamentally, it doesn't take a lot of technical skill to do that. What you need is a lot of GPUS I that doesn't take a lot of technical skills. Sounds like I qualify. Are you? Are you doing? I would love to see a talk from you about that.

[14:00] By the way, at some point there's a just a plug since we're on the topic. There's a complimentary conference to agent Conf in September about AI tickets on sale. Agent.sh. That's going to be about AI. And you know, I want, I want to see that talk probably there. Because fine-tuning, I think fine-tuning is really important, because there's, as you said, there's plenty of generalized models. But but really, doing what I want for my thing is important.

[14:00] There's a trend in web development, you know, composable architectures. We keep hearing about this. Netlify just acquired Gatsby because of this, right? I wanted to ask if that. But for AI is a trend we're seeing. Because I think it might be where companies will not just, you know, build a product around one model, but will intelligently like, compose, and use multiple models. If you can speak to that, I'm curious how that relates, for one, if we can.

[15:00] Define composable architecture from web Dev for the people here. But then also talk to that second part about, how, you know, composing models together is done today. Yeah, yeah, so that concept is called model chaining. And so what that is, especially with LLMs right now, is that, instead of just relying on one LLM, what? You would do is you would basically take a prompt and that prompt, it would be dissected amongst many LLMs, you know?

[15:00] So maybe you're trying to research like, where is Stages from? or something like that? And you have to reach out to many sources to do that. To understand the question by Ella, let me mean large level, yeah, so sorry. LLM is a large language model trained on the Corpus of, just all of human text for the last, however many years. And so let's say I want to say, how old is Tedious, raised to the 32nd power or something like that?

[16:00] So now that involves two things, you know, that involves a plain language understanding of, like, what is the age of Tejus? And then it involves, like, mathematical reasoning. So those are two completely separate things. So model chaining would take basically two separate models, combine them together and get the result that you want, so that would be the composable architecture of language models. What he said? Also just adding there's a really interesting, kind of early stage but very powerful open source project called Lang chain.

[16:00] That enables you to already do this with very little sort of actual engineering. They've been up until now, they've been Python only, so the front-end language is scripting language has been Python. But they've just added TypeScript support in the latest version. So you can now use TypeScript in order to build sort of composable AI models. That will have things like not only chaining, but also memory if you want to build a chat thing.

[17:00] So I remember what you last said, Annie can integrate with outside third-party data sources. Wow, this. This gives me a lot of ideas. I think I saw this in practice. On Astros documentation site, where it's it's docs, it's a documentation. But you can ask it a question like, how do I do this specific thing for my Astro app? And it will give you an answer, right? Do you think that's the future of docs?

[17:00] Almost then? I do think so. I mean, it's very close to what we do. No spoilers for my talk later. And it's really funny because a year ago, when we started building the system, nothing existed we had. To build this thing essentially like handcraft, this information to travel, semantic search, sort of like machine trading system, nothing was in open source. And now there's like, I don't know, a good dozen projects and all of them are running in production in different ways.

[17:00] So, like, this is definitely the way that we're going right now. And yes, I think all docs should probably in the next year or so. Have an interactive sort of Clippy bot that you can talk to. You heard it here? First, make the Clippy Bots. I will say that I'll play Devil's Advocate for a bit. I think it is the direction that dogs go, but I still think that we're very far from an ideal.

[18:00] Because, all right, let's say that your application, your software, your library changes. And so those old docs where the LLM was fed that information is no longer valid again, LLMs could be confidently wrong. However, it could retrain on, let's say, new data. It could read the source code again and figure it out. But docs go much further than that. I'm sure that we've all read docs where it's like, Hey, copy this example and it's going to work.

[18:00] And of course, a language model would be able to read that and tell you the same thing if you asked it in natural language, but the example could be wrong. Like, we've all been through that. So there needs to be that additional layer of verification, which requires, like, I'm sure that LLMs can do it. Like, copilot is very capable, but it has to be able to understand the query and also verify.

[18:00] Am I telling the reader the correct information?

[19:00] Interesting? And that's just not so easy now, but no, it might become over time. Okay, we're almost out of time. I have, one last question that's maybe a bit pointed, but it's too early to think of anything else. It's so Meta, right? That's what makes this fun. I. I keep seeing OpenAI, like a lot of these, a lot of companies, a lot of products, just use OpenAI's APIs under the hood. Why is that not open source like it opens in the name, but OpenAI is?

[19:00] Is just like you pay for this API usage? I'm curious what that's about. Capitalism? Should it? should it? Yeah, should it be open source? No, I. At this point, I don't necessarily think that all of these models should be open source. I mean, there's sort of like parallel, so there's an image generation thing, so stability. AI has, open source, a lot of the models, including stable diffusion. Which means that they now have no control over those systems.

[20:00] They can be used to generate, you know, all kinds of things, including harmful things, whereas, like, I think, OpenAI, you know? While not a perfect company. And I'm certainly not a big fan of the sort of like, some ultimate, you know, white combinator school of like, you know, up and figure it out later. But you know, they are being like, reasonably, reasonably, sort of responsible, with these things, um. And I do think that we do in this initial period, when we are still training and like, figuring out how these things interact with Society at large, I think it is good that there are some controls over it.

[20:00] So, no, I don't think that they should be. You know, it should be available, and also none of you can afford them. On GPUS. It takes to run these things, so it's kind of a theoretical point, but this makes this makes OpenAI really like. So from a systems design language. It may makes them kind of a single point of failure, right? Because it's so much responsibility to curate these models. Can we like?

[20:00] It's a it's a lot of crust to put in, like a few people at a company. No, obviously is, but I mean, they do have a competition, right? Like, you know, like all of these, large tech companies, I mean, you know, Google is perennially almost ready to, you know, like, release theirs, not that Google is any better Steward of anything. But, you know, like I, I don't really have the answer to this.

[21:00] I mean, it's like, damned if you do, damn it if you don't, like, you know either, you, you know, like, let it out open in the world, and you know, like, let everybody do whatever they want with it. It doesn't feel good right to me either. Or then you, you know. Get a very few people very, very rich and sort of entrench the existing Capital class. You know, we've been doing that for 30 years.

[21:00] What's new? Yeah, again, I'll also play Devil's Advocate. Just like, it has this thing always sped on JavaScript, right? Always bet on open source, like if something can be done, chances are people are going to open source it. So when Dolly came out, you know, stability, AI, open source, stable diffusion, and so now it's in the hands of everyone. And so, the same thing with LLM. So maybe not to the extent of open the eye, where they have an absolutely huge data set, but I'm sure people have huge closets and huge servers.

[22:00] Well, okay, so maybe that's impractical, but you are going to see an open sourcing of large language models. Which might not be the entire domain of human knowledge that OpenAI has collected, but it might be a constrained domain where it's easy to share a few gigabytes or terabytes around. Yeah, for sure. So there, there is already a model that we use at our company, it's called Bloom. It's an open source model, you can download it from.

[22:00] Hugging face. It, it, you know, works pretty fine, it's trained with, you know, largely the same sort of, like Internet. Corpus data set as OpenAI did 175 billion parameters. What it doesn't do is that they haven't secretly, also uploaded all of the copyrighted materials. Such as, like all of human literature and like scientific journals, which OpenAI has definitely done. But don't admit to, and, and so, you know, like it is, it is actually quite powerful.

[22:00] And if you do happen to have a GPU cluster with 384 gigs of RAM running around, you can actually just run this thing on your machines. Wow, wow, okay, I didn't know that, that's, that's good. That sounds like open source is winning. All right, we're done. We're just about out of time. But I want to give some time for both of you to talk to, one after the other. You have the mic, so you start, about what you're currently thinking about and excited about in the field of, AI and machine learning, yeah, so.

[23:00] I'm thinking about the future of AI. And, in, in deep learning, there's this thing called a local maximum or local minimum, depending on how you interpret it. And so I feel like right now we are at a local maximum, where we're like, Wow, OpenAI chat. GPT can do so many things, but the amount of things it could do is really limited. However, it is completely new territory. It's a it opens a wide range of things.

[23:00] We're seeing apps, you know, being developed, just on top of this technology, however, it can only go so far. So what I'm looking at is what comes after that local Maps maximum, and that's pointing towards, you know, AGI. So I'm going to talk about in a minute and seeing, like, instead of what can AI tell me, it's more what can AI do for me? Yeah, I'm no good at predicting the future, but I mean, okay, so here's the thing, like, yes, David is absolutely right that we are going to hit a local Maxima.

[24:00] In terms of, you know, this sort of like, this type of, LLM, you know, reinforcement learning, you know, kind of model. That is absolutely true. Yet, at the same time, we haven't even really began to explore the sort of, like, the emergent properties of these systems. What's that Guillermo Roush quote about? We'll be exploring the implications and applications of React for the next 10 years, for the next 10 years. And I do feel like, I, I do feel like, you know, like, even if the LLM technology, technologically speaking, didn't get a lot better, which it also will.

[24:00] You know, we would still have, you know, like, significant new impact being discovered year on year for a long time. And for me, you know, no spoilers for my talk, but for me, the most powerful thing that I'm excited about is natural language. The fact that we can now make interacting computers more human and Humane. You know, interfacing with computers is always has been too difficult for the average person, and it's been way too difficult even for us.

[25:00] You know, people in the know to do complicated things, and I think natural language, while imprecise and sometimes hard. To sort of, you know, get it to do exactly what you want. It is the way that we communicate to each other. And making the computer more human, I think, is the sort of the greatest gift that this thing can get can give. That's so interesting to hear that this irony of leading on machines makes things more human.

[25:00] But that's, that's a great Insight. I think with that we'll wrap it up. Thank you both for being part of the panel. Let's hear David Yani. Thanks so much for being part of it. Appreciate it. Wow, so good.