Vibe Coding and The Rise of AI Agents with Amjad Masad and Yohei Nakajima

28:33And so curious for you to weigh in. I'll share one sort of person who I think highly of, John Donahoe, who was just recently CEO of Nike. But when he was CEO of ServiceNow, we did a Village Global event with him at ServiceNow HQ. And he talked about how, and John was formerly CEO of eBay and has been around the Valley a long time. And he talked about Silicon Valley is such an interesting place because on the one hand, there's all this cutting edge, new ideas, new thinking, new technologies. And yet the timeless principles of business remain, he said. And figuring out what are the timeless principles of investing, of company building, what have you, and then what actually is new that needs to be rethought, relearned or emphases shifted.

29:19That's a tricky balance to get right. Right. And you know, hey, you and I were chatting about when evaluating early stage founders building an AI, what do we look for? And you talked about, you know, closeness to the customer and obsession with the customer, which, again, is almost almost cliche. I agree. I'm just like, it's hard to have a moat because everything's so much easier to build. So like how like if you can get close to your customer, understand their problem deeply, learn to solve it. And I think we as we're still learning how to leverage AI and agents. So if you can stay on top of the best practices and figure out how to apply those to real problems, then I think you're in a good place.

29:55And you can figure out how to scale from there. Yeah. And I guess, Yohei, just based on the discussion so far in the session, I'd love to get your perspective on shifting a little bit back to building applications and agents. You know, we had a question from David Blake at Degreed about how to get in for all the non-engineers on this call and the engineers like who want to lean into Vibe Coding, which I know is one of Amjad's favorite phrases. What advice do you have for people who are building for the first time on Replit or other products to sort of get going firsthand using and building AI agents?

30:35I've had a lot of people ask me this. And my suggestion to everyone is just start playing around. Have fun with it. Think of something that interests you. Is there an API you want to try? Is there something personal that would be fun? A lot of my projects are actually for my kids because it's really easy to do these little simple apps. And it's just a great place to start. And then as you get familiar with it, you can try to tackle more and more complicated things. Got it. And so what are some examples? So Amjad mentioned you building Google Reader. By the way, our little segment on Google Reader is generally, I've got multiple texts in real time here about the nostalgia people have over that.

31:10So who knew that it's that emotional pressure point? So rebuilding Google Reader, but like, and I'm just a two-foot-foot-a-way in, like inside the enterprise, say you're running an investment firm, an endowment, or a 50-person startup. Like, what are the kinds of things that a non-technical person can vibe code into existence? Yeah, I actually have a great example. Everybody on our team now uses Replit Agent. I think every Monday I'd show up to our team meetings and say, look what I built this weekend. And eventually they were like, can I subscribe to Replit? I was just, yes, everybody on my team gets an agent.

31:43And one of my office person who has no technical background, who was managing all of our data on Notion, when Replit came up with the Notion integration, I said, there's a Notion integration. Why don't you try building it up on top of the data? A week later, she showed me my executive dashboard that's built on top of our Notion data. and she did it with no technical background. So you mean, so created a dashboard to be built a replication to crawl through all the data you had in. Yeah, custom dashboard that pulls in all the data from different parts of our notion, like into like all the stuff that I need to see in one place.

32:16Because she noticed that like our notions pretty well organized, but I struggled to like click through and find the stuff. So she just pulled all the stuff I want into a single page and built a RepliDat. And so that was a really great internal tool, right? There's not much risk because I'm the only user. And it was like a great place to start, I think. What are you guys building, Amjad, inside Replot itself? My very example, similar to Yohi's example, is in HR, we needed org charts. And a lot of the org charts, you know, software in the market is too expensive and doesn't have the features that we want.

32:52We wanted something to like connect to like ADP and allows us to like sort of continuously pull from there, be able to like interactively look at the org chart, be able to like have versions so we can look at it over time. And our sort of HR, you know, ops person decided to just like go build it. And she spent three days in Replit, never coded, built like an amazing org chart software that actually we can go out and sell it. It's better than a lot of the SaaS software that charges you like 15, $20 ,000 a year for this kind of software. So there's a lot of different examples like that from our customers.

33:32Yesterday, I was just looking at, you know, I call it an arbitrage opportunity. Someone, their company was quoted from like NetSuite$150 ,000 to build like a sort of a NetSuite extension of sorts. And he decided to go build it in Replit. It costed him$400 and he sold it to his employer for$32 ,000. And so maybe I'll link the tweet. And so this is like, there's like, you know, two to three orders of magnitude savings that can come from using Vype coding in the enterprise. another tip i give people that i think is relevant to this point is that um i think it's a great it's hard to replace a core product feature internal workflow with ai especially if you're not familiar with it so a great place to start with for work is start with things that you wish you had time for but just don't right that kind of nice to have that's a good place to start because it's low risk you're not doing it anyways right it could be i want to track all the publications and you know all the publications in my industry, grab the writers' names, and then tag them by what they wrote about.

34:43That would be helpful for any company. You probably don't have time to do it, but that seems like a really great standalone tool that one might build. So I often kind of ask that question, and it tends to trigger some ideas on where to start. Got it. And what do either of you think about in terms of Max from Happyverse asks, what new capabilities do you expect to be unlocked by upcoming models, you know, Sonic 5, GPT-6, et cetera. Like, what is that? And also, Amjad, you've said separately that as founders, as builders, when we get too focused on the existing capabilities of the foundation models, we actually get trapped in a kind of a local maxima thinking.

35:21So just given how quickly the underlying capabilities are changing, whether it's founders who are building or lay people who are vibe coding, how do we think about the ground moving from under us, you know, in the months ahead? That's a really tough question. Actually, an interesting anecdote here is, you know, Cursor had like Cursor Composer, which is like a really, you know, advanced co-gen tool for its time, like in 2023, 2024. And then when agents came on the scene, starting with Replit and then others, they felt like their system was like old and they couldn't kind of their users liked it.

36:02And so they had an innovators dilemma where, you know, they don't want to miss the new thing. And so they started adding checkboxes. It's like, oh, checkbox, use agents. And so they went through a phase of kind of struggling to kind of like really go into the new thing. I think they figured it out now and they're at the forefront. But you could see that being a lot bigger problem. If you have a big user base, a lot of enterprise customers, they don't want change. And so again, how do you deal with that? I think that's a timeless question of innovators dilemma. And the answer is, I think, ultimately, is being bold.

36:41Is really being bold and be willing to chew glass for a moment of time. see revenue go down, see customers leave in order to sprint towards the future? I think one of the ways that I try to keep track of how fast the underlying models move is to just use it all the time. And I feel like I have a pretty good sense of how far you can vibe code, right? I mean, baby AGI two years ago was probably with vibe coded with chat GPT, copy pasted. Probably my most complicated project I've built again on Replit Agent was VCpedia, which creeps Twitter, grabs funding data. It's like a cheap, you know, poor man's crunch base kind of, but it, you know, it tracks funding data.

37:24That was probably, you know, one of the more complicated things I've built with the Repled Asian. Sorry, just because we have so many investors in the call, describe again what you built. Do you build something that tracks funding announcements? DCpedia.com. I have a couple of Twitter queries that run on a schedule and then an LLM decides if there's funding data in that tweet. And then it extracts funding data from that tweet, converts it into tables of funding, startups, investors, and then enriches with EXA. And then I'm still working on the daily newsletter. So, I mean, it's, again, like, is it better than Crunchbase?

37:57No. Did I build it over a weekend by myself? Yes. Wow, fascinating. We'll put a link to that in the chat. Let's have Dennis come on stage from XMOL and ask your question, Dennis. Welcome. Question for Amjad. I was wondering about how you build software at Replit. do your agents build all the code from scratch or can they use sort of rebuilt components and i'm asking that because when we build agents um internally it was a lot about how much we automate before we hand that over to the agent that should they can handle the complexity absolutely like i think you you you um yeah you perhaps paradoxically you want as much determinism as possible You want to leave as little freedom to the agent as possible.

38:49And the reason is because these models are stochastic. And so there's a lot of things in Replit agent that are deterministic. Like when we start the server, when we restart the server, when we install a package, when we show you when to deploy, all these things, the way we manage database, you want to restrict the points for which the agent is making a decision to the thing that you can't deterministically do. And this is where their power comes in. So absolutely, yeah, do as many things in classical programming before you plug in an agent. Cool. Thank you. Yeah, thanks, Anas. And Yohei, do you want to, should we screen share a little bit and get into the product?

39:35I mean, I think the reason I think this is interesting is I've just done this with so many people where when people see me just start a project, the next time I see them, they've already started one. And it's really, I mean, I've seen the screen so many times, but you go to Replit, you log in. And I think we had talked about an idea of using, looking at some stock. So I'm going to do the little cooking show trick, right? I want an industry sentiment tracker leveraging Wi-Fi finance data across major industry where each line on the chart represents an industry, five to 10 tickers each, with the ability to manage the industries and tickers that go in each, with filters, et cetera, on the chart designed elegantly and modern.

40:12So it's pretty straightforward. I specified that I wanted Wi-Fi Nance because I know it's a free API. So, and I just type in start chat and that's essentially how you do it. Now I'm going to do the cooking show trick, which is where I already have one that I started a little bit earlier. Now we're going to just pull that up because it's going to do, you know, 10, 15 minutes of work for me. I mean, we can't wait for that long, but this is the exact same prompt above. It worked for, you know, it worked for a while. and then I think it ran into a little error. So I said, I'm seeing an error fetching the data.

40:42It tried to use a different API, but I said, no, I want to use Y Finance. And then I said, how's the sentiment calculated? Data feels kind of loud. Those are the only things I typed in. And the result I got from this is this. And what you're seeing here is a handful of industries. I can go in here and see the different tickers that it's basing it on. I can add more industries. I can update the tickers. and it's giving me a daily sentiment analysis. So again, it was a single problem. And is that a public link that you can share? I can deploy it and make it shareable. So again, if you want to actually share a product, but this is what I love about Replit is, if I press run, I can just test the app.

41:23But if I want to deploy it, I can say, I'll just do one machine, sentiment, stock sentiment.replit.app. And I'll click deploy. and then in a minute we can share the link and people can look at it. I mean, that's how easy it is to just get something out there. And if you want a custom domain, you can go into the settings and add a custom domain and link it kind of like you would with WordPress essentially. I mean, so again, this was an idea you texted me about, what, probably like 30 minutes before this call. And so, I mean, this is kind of the process, right? So if you have an idea, it's really as simple as just logging into Replit, typing in the question.

41:59And there are certain types of problems that I've seen people run into that do get a little bit tricky. And I think that's where some technical background helps on my end. But yeah. That's great. Okay, that's awesome. Thank you for sharing that. So for everyone on the call, if you haven't built something, get started today and let us know what you create. And I think that one click to deploy is pretty magical once you see it happen that way. I'm curious, Yohei and Amjad, when you're thinking about as we get to towards the end of our time together, thinking about the landscape of different agents and agent companies.

42:36Yohei, you mentioned that you're seeing a lot of founders build general autonomous agents and there's not much differentiation among them. Unpack that a little bit. And where do you see people going wrong when they think about building or investing in agent companies? I mean, it's hard to say because it's moving so fast. But two years ago when BBAGI came out, everyone started building general agents on top. It was just way too early. We didn't have any of the tools, any of the know-hows. Two years fast forward, I feel like we're seeing another resurgence of these general agents. But I still think there's just too many edge cases for them to be truly general, right?

43:13I think kind of an EA specialist agent is actually more of a specialist than in my book because a general agent should really be able to do anything. So I think what we'll find, my guess, is that some of today's stuff will kind of identify an ICP and try to make sure they can handle all the educations for that ICP, which eventually kind of turns them into specialized agents is kind of my gut feeling, although it's a weekly health conviction. What's your sense, Amjad, in terms of general agents for specialists? I think specialists is still the way to go. there's this interesting thing it's like AGI is such a profound disruption that it's almost it's almost not useful to prepare for it so you want to prepare for the world where we continue to have jobs and we continue to be useful as humans and so you know true general agents is AGI and and so my feeling is that we'll continue to narrow agents will continue to perform better so sorry just let's go slowly because these are some profound statements here I want to make sure I understand it true general agents is AGI yeah okay and so just unpack unpack what that means so if you want an agent that can go into any domain and figure it out so you can give it a novel task like let's say you work in a sort of some kind of niche of a niche of an issue like a sort of a nuclear uh engineer that works in a new material like salt or whatever to generate nuclear energy and like models just don't have a lot of information about that, but like, but you wanted to actually go and research that and do something about that, then you would need an AGI because it's going outside of the training data.

45:14So AGI, I think is the most distilled way to kind of talk about an AGI is that it is generalizing outside of its training data. And we haven't seen evidence of that, of like agents going into, a fundamentally new area and figuring it out without having retrained knowledge. And once you have that, then almost all jobs, especially jobs in front of the computer, will be gone in a matter of months or years. And so that world is so disruptive. People call it the singularity. And by the way, by the time you have AGI, then the AGI companies will use the AGI to develop the next version of the AGI. So you're going to have this like, that world is so disruptive that I don't even try to like plan my business around it.

46:17It's like the moment of singularity. I plan my business around the trends that we're seeing with how AI is able to work longer, AI is able to use tools more effectively, AI has more computer use capabilities, we're adding more training data on more domains. So today, the way automation will happen is that the big labs will get a vertical domain and then get as much data on that domain and then train these models. By the way, the reason computer use doesn't work very well is because we don't have a lot of data. So they need to go to Scale AI and get that data. By the way, the thesis here is that Scale AI and companies like that will just get a lot more profitable because you need data for every domain if you don't have a set of general agents.

47:07So for the foreseeable future, it's much better for experts to imbue their domain expertise into a domain-specific agent. in that domain specific agent will always perform better than a general agent up until the point that we get AGI. Well, and I like Gerdain's comment here, Dr. Strangelove, or I guess he sent it just to host and panelists, Dr. Strangelove or how I learned to stop worrying and love AGI. That's a profound vision of the future. But I mean, I guess, Amjad, just with respect to your comment about scale and needing data, like there's still a huge amount of data that exists only in the heads of humans that has not yet been ingested into LLM.

47:54So there's still a long way to go, right? There's a lot of data that's not captured. I've heard someone who's telling me this example of like a M &A transaction. I think this is Andrew Ng's example of like, think about how an M &A transaction happens. There's like 150 steps. And there are a small number of humans on planet of earth that understand each of those steps and understand the sequencing and how it all comes together. And none of that is codified in a way that can be adjusted by an alum today. And there are like millions of those instances around how things get done in the real world that our models have not been trained on yet.

48:28So there's still quite some time, right? Until all this. Yeah. If you have an AGI, it'll figure it out. Obviously it'll be super intelligence and it doesn't even need to train on it. But Stan's ADI, it would need someone, one of those labs would need to go instrument that and take that and buy it and put it in the training data, which they're increasingly doing. They're spending a lot of money on the compute, but they're spending equal amount of money going around and buying data. Yeah. Yohei, do you want to come in and then we might bring someone in for a question? Oh, the one thing that my mind kept going to was the matrix helicopter scene.

49:02If you had a general agent, you should be able to go look up instructions on how to fly a plane, figure out how to control the fingers, maybe even build its own simulator, practice a little bit in the simulation, and then, you know, control a robot to fly a helicopter for you, right? I mean, hypothetically, that seems possible, but I feel like we're very far from that being reliable enough that I would jump on a helicopter with an AGI that just learned how to do it on its own. So I do feel like there's that my kind of sense of like how safe that feels is kind of a distance of how I measure. And let's let's Brian Ramal just asked, which is a provocative question.

49:36This week, Dario for Anthropic predicted 10 to 20 % unemployment rate in the next one to five years. Current unemployment is what, 3 % in the US, something like that. So even going to 8 or 9%, I think would be dramatic, let alone 20%. I mean, that's, I think, serious civil unrest. Amjad, what's your take on Dario's prediction? It's like a point of contention, I guess. the employment rate is a lot more than 3 % because of employment participation rate. So at any rate, I do think that a lot of routine jobs are within the bullseye, within reach, like within months of reach, especially when we talked about computer use, quality assurance, data entry, any sort of routine in front of the computer uh thing is is going to get automated and then you go into some of the things that require a lot of more creativity and um but but you know is it you know agents are able to do such as like i don't know sdr and entry-level you know jobs and so i i think in the next one to five years uh 10 to 20 just feels like a lot even even doubling even going from three or six to nine percent.

50:58Well, yeah, the problem with the unemployment prediction stuff is there's multiple variables. One is like, what's the progress of AI? The other is what new things can humans invent to do for themselves? Like my canonical example is massage therapy, which did not exist as a profession, you know, 40 years ago, whenever it was. And then out of nowhere, with all this abundant leisure time, like an entire industry was created and now millions of people do that work and billions of dollars are spent every year. So you can imagine several new fields like massage therapy being created that will then employ people.

51:30And so this kind of simultaneous analysis of the progress of AI and the path to AGI, along with the frontiers of human creativity and how we want to spend leisure time. Yeah. And there's other so-called limiting factors. Compute might be a limiting factor. Energy might be a limiting factor. The willingness of enterprise to adopt this technology is a limiting factor. Government might step in and sort of regulate that they're all out of these automations. So it's not entirely clear to me that it's going to be as smooth as what Dario said. Yeah. In our closing minutes, what are some hot takes on companies, founders, technologies that you're really excited about that you're paying attention to?

52:22Could be startups and formation stage stuff could be at scale that you're really either bullish on or perhaps that you think are threatened or overrated. That's a fun question. Wait, just to quickly touch on the last point, though, I do feel like historically, every time there's a new big technology that replaces jobs, everyone's always afraid of it, replacing jobs that ultimately always creates more jobs. And like, I see all the arguments for why AGI wouldn't be the case. Again, not AGI, but five to 10 years, I think we'll still have more jobs. Like, I don't know any robot mechanics, but I'm assuming there'll be plenty of those, probably more than car mechanics, right?

52:56Five to 10 years from now. So on the company side, man, there's so many exciting, fun companies. Yeah. So tick off for both of you, tick off a few that you're excited about or tracking or invested in, and we'll put the links in the chat so people can check them out. I'll drop to Neo is a company that built a fully autonomous ML engineering team. They benchmarked against 50 Kaggle competitions. I think I've scored a medal on 26 % of them. So it's essentially a Kaggle grandmaster, right? Autonomous agent. So I think there's a lot of things it can do for helping, for example, companies that want to fine tune their own models.

53:33Maybe they don't need to, or maybe they can just use an autonomous agent to do that. A recent one called Layers is a fully autonomous marketing ad agent that's baked into the IDE. So you can just download it and run it from your IDE and it'll understand your code and product and just start marketing it for you. I love that idea of like doing everything from your IDE. I can imagine a future where you can, you know, register your company from your IDE and just like hire people from your IDE and a developer can just sit in their IDE and just like run the whole business. Very cool. Amjad, what about you?

54:07I think vertical agents, it's still tough, but I think they'll start to work. Obviously some companies had recent troubles like 11X. I still think it's a good idea for a company to say, okay, we're just going to focus on one role and go and try to automate it. A company I invested in recently is Basis. It's like the accounting agent. I think the URL is offbasis.com. um so uh it is um uh it is like okay we have domain expertise and we're just gonna uh we're gonna go and like put that into the agent but also there's like a white left service aspect of it it's like we're gonna go into these accounting firms and teach them about it because they're gonna have again they're just not gonna you know it's gonna be hard for them to adopt this technology.

55:02So vertical agents is very interesting. I've always been skeptical of investing in labs, but I did invest in one lab, which is Francois Choulet and Mike Noob's lab. So Mike is the founder of Zapier. Francois started Keras, the ML framework, worked at Google. They started this company called India, N-D-E-A. And it is, they're the founders of Arc AGI. So Arc AGI is this benchmark that everyone's competing on. It's like, it's widely known that it's the best sort of AGI benchmark. Their approach is to use program synthesis as the basics of how the models work. So program synthesis is like how you generate code, but it has like a more discrete way of generating.

56:03It's like more correct way of generating code. As we talked about earlier, in my opinion, software agents are a core component of what you might call AGI or general agents in the future. So it's one of those bets. It's my first bet that I'm making on like a sort of AGI company. I think it has like a very interesting approach. So we'll bring Jack on stage real quick for our last question about how to integrate sort of Vibe Coding and Replot into a software engineering existing code base. Jack, go for it. Hey, thanks, guys. I have a super practical question, which is just it's clear that if you're just trying to create something from scratch, it's just incredible what you can do super quickly.

56:45But what if you want to take an existing tool that's part of a large code base and you want to improve it or you want to add in a new tool, an existing code base that you don't have right now? That's what we're trying to do and we're finding it more challenging to make that happen in a seamless way. I'm curious if you have any best practices on that. With Replit, we're focused on greenfield applications that are built on Replit. Over time, those get more complex and you can continue working on them and we're just continuously making the system better. But we want to control the initial condition of the app such that the agent is going to do its best job.

57:27And as you see in things like Yohi has built like BCpedia and things like that, you can go pretty far. The way people are integrating Repl.it into their product development process is like as a prototyping tool. And it goes way beyond the Figma prototypes or whatever because what we're seeing is product managers and designers are actually taking it to users and testing it in front of users and maybe going through like a beta program with users to kind of learn a lot more. And it's cutting, we're hearing from customers that it's like, you know, cutting 20, 30 % of the product development life time.

58:00We want to get to a point where you can like import a GitHub repo and like run Replit agent on it. But I think that agents are not very good at it right now. I'm sure you're trying Devon and Kodaks and things like that. It's still not there with Replit. We're always trying to be at the edge of what's possible, but we don't want to overshoot. And so... So it sounds like prototyping for now, maybe full integration later. I would try Codex and Devin and these things. Thank you, Jack. I'm Jad Yohei. Thanks so much for the insights. Villagers around the world, thank you for joining us. We'll keep the conversation going.

58:32Have a great day, everyone. Take care. This was fun. Thank you. Thanks so much for listening to the Village Global Podcast. You can check us out online at villageglobal.vc. We'd love to hear from you, your feedback, your ideas, your inspirations. You can email us at hello at villageglobal.vc.