The Casino enters the arena + Guillermo Rauch of Vercel on AI factory builders

I think we need to break down the logic of why this week people decided that the AI bubble has burst. And there were some key things that happened that I think we can deconstruct and decide whether or not they're actually worth, you know, being a hater over. The first one was Sam Altman had a dinner that I was not invited to with reporters where he discussed, among other things, and it was all on the record, uh, you know, GPT-5 and some of the disappointments around that, which are very real. And also at some point said something to the effect of, I don't have the quote in front of me, basically like, yeah, I think we're probably, you know, there's some exuberance or we're in a bubble right now where some stocks are, are very overvalued because of AI. But then in the next sentence also was like, but also AI is transformative.

It's a classic power move. Elon does this too. It's like you're in this super hypey thing. There's almost nothing to lose by being like, yeah, I don't know, people are really think I'm hot shit, maybe too much, you know?

Yeah, yeah, yeah. My, my, my lovers love me too much. My haters don't get it. But if there's something wrong, it's because of something other people did, right? I just want to disregard almost everything Sam says these days. I mean, it's just, he's on, he says so much and it's so contradictory that it just doesn't even, even though it probably did play some part in the sell-off of AI, you know, AI-driven stocks, the idea that, oh, Sam Altman has declared a bubble, so we must be near some sort of a top. And then there was this MIT survey that came out and I think that probably had the most effect on a lot of people, which headline number was that 95% of enterprise AI generative AI projects fail. Just have zero impact on a company's efficiency and bottom line. And that is, that is a bad number. The problem, of course, is that a headline number versus what was actually in the survey is like pretty different. And this was not like a survey that was put out by a fully disinterested organization. Like it's effectively a pro-AI organization.

It's always funny when somebody like runs away with one line that you said., but then like you're like, oh, but I'm sort of trying to make a pro-AI point here.

Yeah. Yeah. Like it's almost like a ratfuck in that sense. Like if you're like one of these people who decides to buy the headline and then you actually like distribute the survey and then anyone actually reads it, sees that like the key points that it makes is that AI is not incapable of transforming a business, just the implementation of AI and its inability to learn and adapt to people's workflow is what's causing it to be unsuccessful. My takeaway from actually reading through this thing quickly, quickly was that like the 5% of companies that have been successful with it are like gonna win. Like it's really about like the fact that they implemented it better and were able to use this technology is why like it is actually a transformative tech. And like one of the co-authors of it is a product manager at Microsoft. I was just looking him up. So like these people that are writing this have a they have like a dog in this fight and it's not like long-term, it's gonna be a good dog.

I think the flow of the argument is basically everybody sees the potential of ChatGPT and Claude. They try it out, but then they almost like, and they just don't do enough to actually get their data integrated into whatever AI tools they're bringing into their companies. And so then it ends up not being more useful than just like consumer-grade stuff. Right. And like, here's what you actually need to do to be better than just like using ChatGPT.

Right.

Like that was the key point. Here's what you need to be doing to make this successful. And that's gonna end up probably being the talking point from all of the salespeople at the cloud software companies that are selling this stuff is like, oh yeah, you read that report. They made these mistakes. You won't make this mistake because you're gonna buy it and use it in this way. So it ends up the whole thing was an op, I've decided.

I will say I, I texted a lot of investors this week, just like, what's your read on the bubble. And like, you know, people who are like, you know, investing in these companies often, I'm interested in what, you know, yeah, people who are trying to make money, like, you know, it's like if it's an early bubble, it's like, oh, I need to keep going. But you know, they're, they're trying to time the bubble for financial reasons. And I did get some, you know, somebody was like full bubble now trying to decide whether this is the start of 2021 or late 2021. Obviously everything unwinds from the pandemic at the end of 2021, there's still, upside to be seen at the beginning.

And as we talked about—

Is this person selling?

Are they offload?

They are. I am in tread lightly mode right now.

Haha.

Somebody else was like, I'm watching the Fed. A lot of people are like, I don't know, what do you think? Yeah. I mean, so nobody's like, what an absurd question. Andreessen Horowitz, who sort of dispositionally is like, why would you ever sort of position yourself as a bear? Our job is to fucking be professional bulls. Just came out with a piece that was, you know, knocking down basically, I thought, Tom, not trying to fight with your article about Cursor. I think the sophisticated critical take on this model—

Well, they're just mad that I have an MBA, so they just view me as the enemy.

The sophisticated critical take, I think, as we're about to say tomorrow, it seems like foundation models, revenue's still strong, but application companies like all the coding companies have pretty negative margins. They're built on top of money-losing businesses and themselves lose money. How's that all going to work out? And Andreessen and co., I think it's I assume it's Martine, um, Cassato and I think Sarah Wang sort of, you know, push back against your take. What do you think about that as somebody who is worried about app company margins?

I think the margins are less the issue than like, I think there's an interesting dynamic, which I tried to get to in that story, which is the relationship between the model providers and the application companies. And it's clear with Cursor's growth that you can build a huge user base on an interesting application, maybe has some brand name loyalty to it, but like, especially with developers, like they will just switch to the thing that provides them the most bang for their buck. That's going to be the best quality product. And like the growth of Claude Code over like 6 months, which I know wasn't impressive to Ed, but, but is to me, you know, getting to like a $400 million ARR shows that like people will gladly switch to a new thing that seems like it'll work well. And Anthropic has I mean, they're money losing too, but they're going to have better margins than Cursor will because they're paying wholesale access to the model versus the retail prices that, that, that Cursor has to pay. So I'm not incredibly bullish on application layer companies in a world where model companies build the same software. If they decide they don't need to, that they really think they can make enough money through their API business or like a consumer, a pure consumer play, then maybe there's a world for application companies, but like. I think it's a very, on this particular sector, it's a very narrow moat. And I mean, I was talking to an—

I sort of agree with the Andreessen people. These companies are not idiots. They're going to figure out like some of the stuff, like the very expensive stuff, they're going to offload to cheaper models. I agree with you that like the big business question is just like, how sticky are these products? I think that it's like, how sticky are they? I don't, and I don't, I don't really have an answer, but I don't think the bubble is going to swing either way. It's like, oh, they're not sticky though, because the people are moving to another AI product. You know, at the end of the day, coders want these coding tools. The margins will be figured out. They're gonna consume some AI product, either foundation models or companies built on top of it. Either way, it's good for the AI hype train.

Well, and I, to defend the piece, at least, like I wasn't, yeah, no, no, I don't, I'm just saying to Andreessen, 'cause apparently they wrote it at me. Look, I think, you know, that, that the cursor is not running outta money. Like the, the story was very clear that, you know, Yeah, negative gross margins are bad, um, but they have like a billion dollars or something, right? And they're losing, you know, apparently their burn is like in the single-digit millions. So it's really an issue of like, you have plenty of time to figure these things out. And I mean, you mentioned Uber, like I covered Netflix during its most cash-burning days where I'll never not bring this up on the show. Like I worked for an editor who like made a career for a time writing columns saying Netflix is a house of cards, right?

Have you written a mea culpa?

No, we never do. We move on. We're like the, we're like the Uber haters and the AI haters.

I don't move on. I just played it. I was the factual, I provided—

Where's your mea culpa about Uber?

I was like, I was like the gunrunner of the Uber skepticism era. You know what I mean? I was like giving all the haters the like ammunition they needed to like shit on Uber's like money losing. But I was always like, I don't know. And then I always got annoyed that the people who took my stories and like, oh, they lose a lot of money, did all these like insane math formulas based on them. And I'm like, there's no way. My numbers are like so like, there's such impressionistic, you know, like, Uber was giving very, like, rough numbers to people when they were trying to back out their economics out of it. I was like, this is lunacy.

Yeah, well, the problem with the hater argument there is it relies on the investors who see more than we ever get to really just being dumb, right? I mean, it's the idea that, like, they don't understand it, but we as the skeptic and the hater do. And believe me, I don't like giving investors credit. Like, it's not fun, it makes my job worse. But, like, they do kind of do it for a living. Like, that's the one thing they should be able to understand, right? Is like the basic economics. One of the things we should get haters on here too.

I want to get some more haters, right? No, no, we like the ad episode. One of the things the YouTubers— somebody got mad at me for saying, do you believe in capitalism? You know, like it was like some sort of shibboleth, but it was sort of a— it was a sincere question. Like some of these people, I do want to understand like what premises do we share? Because like to me, part of believing in capitalism is like, investors are fairly rational. They're making different bets. They have their own, you know, it's just like, I, it's useful to know how much people think like the whole capitalist thing doesn't ultimately direct us towards likely outcomes. Like to me, markets are great truth mechanisms and like we play a part in that by, you know, point poking at things and interrogating them and then people change their ideas and then they eventually move their money. But when people have commentary that acts like investors won't listen to a good idea and like change their strategy if it's in their incentive, Yeah, it feels like they don't think the market—

the truth mechanism is over time. I mean, like, the periods of irrationality in the midst of a bubble can give people the wrong impression of, like, the, you know, solidity of a business. Like, I think that's what Ed was arguing. It wasn't like capitalism is failing because these companies exist and keep raising more money. It's just like it hasn't reached the point yet where, like, you know, whatever the Warren Buffett line about, like, the shore— the shoreline going out and, you know, finding out who's naked and who's not.

Right. And I mean, one thing we really believe in, and I think smart business journalists are like, there are people who can be behaving according to their incentives and driving us off a cliff, right? I mean, this is sort of the like Tiger Globals and the Softbanks. Like often I've made the point in the newsletter that it makes sense career-wise to be the like biggest bull of the biggest mania, right? You're like, oh, at least they were the most important person, you know, like in that moment, like The Tesla longs and stuff like what's her name, Cathie Wood. Like, even if you have a terrible performance, you sort of like win fans just for being like really extremely bullish. And so even if you like believe in capitalism like I do, certainly I think there are all these actors who are like, you know, almost like taking advantage of, you know, human psychology rather than, you know, trying to make the most money they can. They're just— they're like, oh, I'm going to be able to— if I'm the best bull in the world, I'll be able to raise the most money again in the next bull run. And my game is just to like Run as hard at the bull market as possible every time. And you know, like that's a, that's a game people play with, with great success.

Speaking of taking advantage of human psychology, Chamath is back out there again.

Welcome back, scam in the arena.

Scam in the arena's back. No crying at the casino, Chamath.

Right.

I know.

What a, what a, oh, he's so good at it. He's so good at it. He's like, like he, he's basically calling himself like the casino and it's like, sure.

The casino, the, the, the random outcome generator.

I understand. He's literally like, don't invest in me if you're not willing to lose all your money. And if it's like trivial to you, like, yeah, who invests like that? Even if you're rich, no rich person wants to lose their money either. Like everybody is investing to get a return. Like I know it's just like a bullshit thing you say, but this is the unique twist of the Trump era, right?

Is that like he has discovered there is a whole cohort of Americans who like being scammed, who like, uh, well, that's true.

I mean, sports betting is the biggest thing in the world. Like, yeah, exactly.

Like Yeah, I lost a lot of money on sports betting over the weekend. Yeah.

Yeah.

Oh God.

Yeah. I hope you just played it. Was it just for the fun of—

it was just for the fun of it. Yeah. There, I just went down to the sportsbook and I put $50 on a, on a horse called Duck Duck Go.

And at least it was in person. You got some, you got an experience out of it. My, my issue with the sports betting apps is like, what are you even getting? You know, like I think it's—

it raises the stakes of watching a game. I mean, if you're not watching the game, you're putting money on it, then that's a real, that's real sick.

Right.

I did invest also. I put a bet on Australian rules football because I thought the spread was so ridiculous, but turns out it was not. That team lost by like 150 points, which I don't know what that actually means. Anyway, Chamath is, yeah, I think he absolutely represents the peak of Trump era. We like to be scammed mentality where he is actively calling what he's doing a casino, that there will be crying and you will not make any money. I will make money from it. He made money last time. I mean, you like wrote, I thought, a great piece about like his scam the last time through. Like, what are you expecting this time around?

I've written two pieces on Chamath. I mean, one thing that makes me so sad about the world we live in is just how much your internet brand and your real person brand can be so far apart. Like, real people in Silicon Valley think he's like a huckster. Like, people who work with him, it ends poorly for them. Many of his professional relationships have gone poorly. He didn't make people money on the SPACs. He made money through fees. Like he's looking out for himself. And to some degree, you still want somebody who's like looking out, who cares about— caring about your reputation is good because it means that the people you partner with and work for and invest their money for you, you care about your long-term reputation. You want to be good by them. And this is someone who hasn't minded his long-term reputation.

It hasn't., you know, not his reputation as an investor, but if you look at the way he positions himself, he wears his, like, I don't know, the fucking brand name suits.

He loves to be rich.

Yeah. Yeah. And so like, again, not to make everything just about Trump, but this is all part of the same thing. People in the real estate world knew that he was an absolute dud as an investor. Like all of his projects would go bankrupt. He was a brand name to, to, you know, add a modicum of flash on top of an otherwise shitty project.

Here you're like, you're like, bring back the WASPs. Where were the rich people who care about—

I mean, what old money?

Just people who inherited their wealth. No, bless oblige. You know, we need it. We need like people who are not—

we just need the Carnegies. I am watching Gilded Age right now, so I have a very strong attachment to that era of wealth. But yeah, those guys look tremendous compared to the Chamath. I mean, they've done nothing, or he's done nothing, but he's built an image that I'm sure he will have no problem getting people signing up for his SPACs.

Well, it's that sort of fatalism that makes it possible, right? It's this sort of like, people are like, oh, he'll raise the money, we should do it.

Yeah. And I guess maybe the last point I talked about—

I am not cheering for him.

I don't know what to say. I would be very sad if good people lose money on what are certainly going to be failed IPOs or whatever you call them, SPACs.

I do think, to be fair, you know, or business journalists, like, I do think he changed his structure this time. I do think he's slightly more aligned now with the actual outcome of the SPACs, though knowing Chamath, he's found some other way to make money as well.

But burden's on him to prove that he's not scamming people.

Sorry.

This— well, that you believe in capitalism, don't you? There's infinite numbers of shots in a free market.

Sure. Yeah. But I mean, this is always— this is one of these, I don't know, to me, psychological weaknesses. The elites, once you're in the club, you're like, oh, you're in the club, you should get to do it again. It's like, There are tons of human beings, like, try it with somebody who hasn't lost people a lot of money. Like, to me, this sort of desire to, like, keep giving money back to the type of people, to people who just, like, gamble the money rather than saying, maybe we should try it with somebody new, that to me is like a human psychological weakness of just, like, loving fame rather than success.

It's never gone away. I mean, like, Michael Milken in the '80s, the junk bond king, still, like, runs a large financial conference in Los Angeles, the Milken Conference. Nothing, nothing tends to really happen to these people for the long term. I mean, some of them do get like banned from trading and, but that's not even close to what's gonna happen with Chamath.

Well, not even more. Yeah, exactly.

That's it. Like, I'm not accusing Chamath of anything of that level. He just was an opportunist and fucked over a lot of people who didn't—

anyway, we're looking at fights. So Chamath, come on the show. We're, yeah, I don't think he is. That would be a big get for lots of reasons. Probably the person who's written the most negatively about him. Not betting on that one, but if you're, if you're a hater, positively, negatively, we're ready to argue with you. We like this arguing format. So we want to have arguments, not quite an argument. I like him too much, but after this, we're going to have Guillermo, Guillermo Rauch of Vercel is coming on the show. I think he had, I thought he had a really good prediction about where things are going. We talk about agents. He obviously runs a no-code company and I continue not to, I have not built my no-code startup empire. If it's so easy to no-code, like where, where are the apps that I'm using?

But anyway, I get into fights with Lovable.

I know, dude. I don't know if people know what you're talking about.

We don't. Yeah. That, that can be deep lore for newcomer fans. You can look up Eric's tweets trying to get his money back.

Waste your time on the forum. I, I lost it on Sunday just because I, I got charged twice from Lovable after trying to cancel. They, they have extreme dark patterns that they say they're gonna try to reform. Literally, you click a big red button that says cancel, and then you still are not canceled from Lovable. They're like, oh, that was canceling within Lovable. Then we had to send you to Stripe, and then you have to obviously run through the Stripe. But anyway, they swear they're changing, they're gonna fix it.

Maybe cancel in Swedish means something else. Maybe it means like, I'm still down. I like to give it another shot.

I like the company. I would've come, I, it was just like, I'm not using it this second. I would've come back. Now, I don't know, I'm pretty angry with them, but hopefully they're reformed.

We're big supporters of Europe tech over here on Newcomers, so We long-term, long-term bulls on Lovable and every European tech company.

All right.

We should probably cut on over to your interview with Guillermo.

All right. I'm here with the founder and CEO of Vercel, Guillermo Rauch. Welcome to the Newcomer Podcast.

Thanks for having me.

Last time we were hanging out, I feel like you convinced me that Vercel would be every company in the world. So we'll see. We'll see what you brainwashed me into by the end of this conversation.

Where, where were we?

I think it was HumanX, if I'm remembering. I think we had like lunch off that and just like, yeah, the great founder visionary that I was like, oh, what won't this company do? You're obviously touching a lot in AI, you know, like you've got Veo now with sort of your own vibe coding product, but you're also sort of powering a lot of like the vibe coding that everyone's talking about and sort of some of the infrastructure to make it possible. For people to sort of spin up sort of websites and applications quickly using artificial intelligence. But also you're just like a fun founder who's willing to sort of like say what's going on and, you know, not, not just let the venture capitalists dominate the chattering among the Silicon Valley classes.

And I'm permanently online, so you can ask me about anything that's happened on X 5 minutes ago and I'll know.

So I just want to start off with, yeah, like the existential question about like the mood. I mean, I've been, Honestly, been texting everybody today about like sort of animal spirits or like, how's, how's the market? Like, you know, what, what's your read on sort of the AI mania? Like, do you feel like there's still a lot of room to run here or are you getting signs that there's like some, some paranoia about the euphoria?

You mentioned Vercel is in a very specific and special position that we see every, almost everything that's being built on AI, right? Has to have some kind of user interface to the world. Right. And so traditionally we've been supporting founders in building these AI portals, right? AI applications, now increasingly agents. And one of the things that I've been noticing is that there's basically 3 chapters of AI so far. Number 1 was assistants. Number 2, and it's the arc that I think we're currently in, is agents. I believe that there's a 3rd phase. I think what's really interesting being so close to developers is that I've always kind of called out that us developers, we're very self-serving. So if we get AI, we're going to apply it to our job first. We're going to make our own life easy. And so what I'm seeing with developers is that the next generation is going to be teams of agents or multi-agent architecture. We've gone from assistant to agent to multiple agents. The alpha developers today, what they're doing is they're, instead of just limiting themselves to like one agentic session, they're spinning up 20 in parallel. Or 10 or 5 or whatever. And that gives you a glimpse of what I think is gonna be the next phase of AI. You can think of the current phase of AI as you have a virtual coworker. You can think of the next phase of AI as I have a virtual team of coworkers and I'm the manager.

Because the challenge with just having the virtual coworker is in some ways the human gets the lame part of the job where it's like I have to sit there and be like, sure, yeah, keep doing the hard work.

That's right. Yeah, keep doing it.

Just hallucinating. Yeah.

And it's tedious and perhaps you don't get that 10x effect. Right. You know, I do see a lot of debate, healthy debate sometimes of like, is AI a 2x or is it a 10x for my job? I saw an article that a programmer put out the other day saying like, look, in certain areas of my job, it definitely feels like a 10x because the example that this gentleman was giving was like, I don't write compilers every day. And so like the fact that I was just able to tell the AI, can you write me this like compiler transformation? It's like, whoa, like I feel like a superhuman. But then he was saying like, in other areas of my job, like, I don't know, sometimes it like kind of gets in the way and whatever.

To sort of really, I guess, go directly at the current moment. I mean, yeah. What's your reaction been to like ChatGPT-5, I guess, like to GPT-5? Because in some ways, I guess, I think if that were amazing, the lead here would be like, oh, the models are gonna keep coming up with new ideas and that's where we're gonna make progress. And just by the fact that you're saying, oh, we're going to find ways to use the existing models and agents to like get more out of them by having them collaborate, to me is sort of a sign that we're not having the like, oh my God, ChatGPT or GPT-5 has like blown all our minds. Do you agree with that or what's your read on it?

I agree with you that analyzing the GPT phenomenon is really important to the industry because you're right, like it came out and it didn't come with a short proof of the Riemann hypothesis.

This is exactly where's our—

exactly.

And then expectations are very high.

Yeah, right. But something that was really interesting was that OpenAI is taking the coding problems really seriously. Prior to its release, they came to us and they're like, can you please run this on v0? Can you please run your evals with it? They were asking us like, what are the vibes from your team? What are the vibes from your tests? Is it good at design? So there was a lot of attention being paid to the topic of coding. And the simplistic view of that is that the market is becoming hypercompetitive on the enterprise side, and that Anthropic is making a ton of progress with coding. There's amazing Chinese models coming out like Qwen, et cetera. They're really good at coding. And so the simplistic view is like, oh, coding is a hot market, so that's what they're going into it. My view is actually that because agents are the future, not assistants and not one-off, one-shot things. The future to superior and higher intelligence and more productivity will be agents writing code to solve problems. And so the thing to pay attention to is not a consumer observation. In fact, it almost backfired for consumers, right? Because we were like, can you please bring back my boyfriend or girlfriend or—

Right, we don't want it to be smart, we just want it to be, yeah, likable.

That was the main thing. Right. Part of me had almost anticipated that this was gonna happen because I did the following thought experiment. If we could revive Albert Einstein and we all get a digital copy, how many people would be jazzed? I know a lot of people here in Silicon Valley, you and I would be like, holy crap, right? Einstein in a box.

Like, you're a crazy guy. You don't want to talk about like reality television or whatever. Like, why, why would I want to talk to you all this? I see the point, right?

Yeah. The average person would be getting too much horsepower. They're like, hey Albert, like, what should we watch on Netflix tonight? And he's like, I don't know, like, I don't watch Netflix. Let's talk about like general relativity. Right. And so I think GPT-5 had a little bit of that, of like, it's becoming intelligent in a way that's going to enable this like emergence of intelligence that you can't easily probe for. The future of intelligence will be, and we're seeing this with like test-time models, like going in a loop, spending a lot of energy writing code, testing it, exploiting multiple branches of the search algorithm. And these are things that a consumer cannot possibly ascertain. Like the most that a consumer can do is, you know, I have friends that keep questions in their head that they know like AIs are not very good at. And so like they do their own little benchmark of like, oh, let's see if GPT-5 gets this right now. Testing the intelligence or the IQ of AI in itself is becoming really, really difficult. That's why OpenAI was so interested in like, what's the v0 point of view? You're creating a coding agent that, you know, is specializing in, in vibe coding, et cetera. And so that means you have a lot of tests. That means you have a lot of benchmarks. Like, you know, what does that mean for you? I believe every company should be concerned with creating these benchmarks of their own institutional knowledge and intelligence. Well, some of it, you know, it's like, for me, it's like I put in like, oh, is it good?

How good is it now at writing a story like me? How, how's, you know, good is it keeping track of?

Yes.

Proofread problem. You know, you, you end up having your specific challenges. So when you're in an industry, it's not that hard to figure out like, oh, is it doing a better job for me or not?

You could also make the case that you should— you could be working on a writing agent that is more than just one forward pass of like— and this is another thing that kind of backfired, but it's also very interesting, is that they try to ship in a model router, right? Like the default thing was auto and then it's gonna choose if it's going to think hard or if it's gonna give you an immediate response. Another thing we've learned about AI that's fascinating to me is that there is this metaphor of like thinking fast and slow for which there is a book and like people have been talking about for years, like prior to like AI existing and it's, It's proven true. There is two modes of thinking. There is fast, fast twitch muscle fibers where like you just want like, hey, like can you proofread this really quick and find me typos? And then there is like, no, no, no, like let's analyze like are you repeating words too much? And that requires like a slow thinking. And so OpenAI tried to, because they're in the pursuit of creating a good consumer experience, they were like, okay, we're gonna try to route automatically. That appears to me to be an extremely hard problem, if not impossible. It's hard to know from the query if it's a hard problem or a slow problem. We have a related problem in computer science called the halting problem. Like by just reading some code, we don't know for how long it's gonna run. It could run until the heat death of the universe or it could run really quickly, but we have to run it to actually know. And so there's almost like an uncertainty of like, should I spend a lot of energy thinking here or should I respond really fast?

In, in some ways, this is the argument for the role of startups in AI and need in sort of narrowly focused companies. It's like really you're saying the model companies should almost be an API layer. They're not gonna be great at deciding in every sector whether it should be a fast or slow thinker and what content to bring in or how many agents to use, et cetera.

And so where I was gonna go with your case is that It'd be really cool to be crafting the ideal writing agent.

I know most people listening to this will have an intuitive sense of the answer to this question, but like, I'm curious to hear you define it. What is an agent to you? Like, does it need to take action? Like, I mean, deep research itself is an interesting case where it's like, oh, I mean, a consultant is like an, you know, taking actions in a certain way in deep research. Behaves like a consultant. On the other hand, and people have probably heard me say this before on this podcast, like, to me, the gauntlet for an agent is like, can you go spend my money? Like, will I let you really make sort of permanent sort of decisions that, that have real consequences in the world? I don't know. How do you think about what, what sort of the minimum requirement for what makes something—

the minimum requirement is that it produces an artifact an output that is generally longer than an answer. So it, right, like if I make that distinction between like an assistant and an agent. So if I'm asking like, you know, what's the province, main pro— the largest province in Argentina versus build me an application. Funny enough for us, that was like one of the biggest jumps in how people have perceived the intelligence of v0. v0 started being what we call one-shot. You would tell it, give me a user interface for a document collaboration system. And the feedback that we got was like some people literally have, because this was so early, it was basically maybe a few months after ChatGPT launched, some people would tell us, all right, there's not going to be any jobs anymore. I just saw God, holy crap, what do you guys do? And some people would tell us this is the worst piece of crap ever. Because in the one-shot world, quality is so divergent. It's almost like playing the lottery. In fact, some naysayers of AI for coding are saying that like what drives the revenue is not the quality, is the lottery effect, which I completely disagree with, but I wanted to tell you, this is funny.

It's a dangerous idea, Vizito. Yeah, it's a grabby idea.

Interesting. The next big jump for us was when we made Vizito really work like an agent. And so what it does is that it makes that one shot, but then it looks at it. Visio asks itself, does it work? Does it compile? The most recent version of our agent even physically looks at it, like using computer vision. So as an example, say I want to make a website for my daughter's birthday, please theme it like, I don't know, Disney princess. It'll then look at what it cooked and say, does it look like what I was asked? And so an agent is taking a multi-step approach, much like a human would, in order to produce a higher quality output. One of the key ingredients there is this idea of reflection. Like, the agent is asking itself questions, it's looking at artifacts, it is using tools, One tool that we have is it can look at a screenshot, it can look at designs for inspiration. So sometimes the model doesn't know what you mean by like, build me a podcasting app like Riverside. And so it needs to go and research.

Do you like the term vibe code, by the way? Do you embrace that term?

I personally do. It's funny enough because like our product's becoming so successful, like the Fortune 500, some people are like, Do you really wanna say vibe coding? I think vibe coding is, it evokes this idea of you're not actually coding, which is really important. The term came out because, uh, Andrej Karpathy and a bunch of our people were realizing AI was progressing at a clip where your own micromanagement of the agent was producing diminishing returns. AI, once upon a time, like in AI times, like 6 months ago, once upon a time, it was so bad that you literally had to look at every token it output. As the models get better and the world becomes more agentic and self-healing, you realize that you are looking at it, it's just wasting your own energy and your time. You can start trusting and even seeing because we can render that output, we can run the application, you can see it come to life. And so there has to be a term for that. The winning term right now is vibe coding, but it's basically— sometimes I've called it instead of being code first, like a product like Cursor, it's code last. You might occasionally, maybe at the end of the process, you look at the code, but the code is not the important thing here. And I believe that most agents are going— going back to like why GPT-5 is obsessed about coding, even though people are obsessed about 4.0 being nice to them. I believe that agents will have to write code to get to that next frontier of value for society and for intelligence. And I believe that a lot of people will not even realize that these agents are writing code.

What's happening to your employee salaries? I mean, you know, the, the AI engineers, obviously we've seen all these stories of Meta poaching everyone from the research labs. Like, how's this playing out for sort of the typical software engineer in Silicon Valley? Or what?

Yeah.

What are you seeing in terms of trends in comp? Are the, the $100 million salaries as widespread as chattered about?

We are in a little bit of a microcosm here. A metaphor that someone gave me the other day is like, look, like it seems like if you extrapolate what you're saying, And because I know that v0 hires developers to build v0, we're transitioning into a world where the job of an engineer is to build the factories of code rather than being the line worker. And so the engineers that are capable to building the factories of code, the— by coding platforms, the agents, the AI interfaces, the assistants, those people are highly empowered today. And so it becomes that highly competitive space. It's like, you know, how many soccer players like Lionel Messi are there that have that specific set of skills? And then like, you know, there's lots of soccer players, but a handful that make a ton of money. And so what I would recommend to people is that they position themselves in the, in world, right? Like become the person that understands how to bring those two worlds that I talked about earlier, the theoretical AI potential and its applied AI presentation layer. That's kind of the, our thesis of the Vercel AI Cloud is that as this model layer keeps getting smarter, better, more accessible, you should be focused on applying that AI to a specific domain, a specific business, a specific vertical. A specific kind of agent, but you're going to need a cloud that has all of the like right patterns, tools, defaults, etc. for building those. And that's, that's our thesis with, you know, we're not going to build a cloud that is one-to-one exactly like AWS. There's no, no advantage there. But we want to build a cloud that is perfect for that next generation of companies and developers.

I'm trying to understand. I mean, this is, we're in an insane situation where you're right at the high end. If you're building the AI factories, you could command insane salaries. On the other hand, if you're sort of the sort of, I don't know, worker, more like line coder, software engineer, I guess you're worried like, oh, you see v0 and you're like, oh, that's coming for my job. So my question I think is coming at you from two ends, which is like at once like, oh man, we see these insane salaries. On the other hand, I feel like there are the sort of, I don't know, Wall Street Journal stories saying don't study computer science because there aren't any jobs. So I guess take one at a time because I think they're getting mixed together. Talk about each end of the equation here.

So for the other end, I do have a complete different take on universities and colleges and that is a can of worms. I think we should address it. Okay. On the line coder, I also think that the demand for code from those people has actually never relented. Meaning let's say that you were able to produce like 1,000 lines of code per hour, per day, whatever. Your boss probably wanted you to produce even more. Like there's more features, more bugs, there's more innovations, more products, et cetera. And so if you use AI, you're go— that's why I was saying that it was, it's so exciting to be seeing people go into this multi-agent ways of working. And so the person that uses AI is gonna run laps around the person that doesn't.

But I do think— Do you think we'll have more like frontline engineers in a decade than we do today?

It's very hard to estimate in, in like in total numbers, right? But I do think we'll definitely have more software. Right?

Like, right. Well, yeah, I think we all agree on that.

The definition of an engineer will likely expand in such a way that I can tell you definitively yes. Right? Like, because we're putting software generations in the hands of everybody, as a whole, there will be, you know, hundreds of millions, billions of people creating software. The specific engineer as we understand them today, I can't even, you know, I had a tweet a couple days ago that Elon replied to saying, I was saying like, I can't even explain just how different software engineering will look in 5 years because the perception that people have today is of a code editor. Like that's it. Like if you watch a hacker movie, what is the idea of a programmer or they're looking at a code editor. But now I'm looking at my fellow factory builders and I don't see them looking at a code editor anymore. I'm increasingly seeing them talk to all these agents. And so that to me already, if I was able to transport myself to 5 years from now, I don't think software engineering looks anything like what we do today. And so it's very hard to say, are we calling those people that are managing the fleets of agents, engineers. You know, most likely I would say it's more on the consumerization or democratization of software creation.

To a direct point you were touching on earlier, do you think people should study computer science right now?

I think people really need to understand how AI systems work, period. They need a deep understanding of the technology stack, the limitations of the technology, what makes them safe, what makes them not safe, what makes them make mistakes, when are they good, when are they not good. I do have a little bit of apprehension about do the universities today themselves know? And I would say most likely not, right? Like think about how fast the industry's moving.

Right. That you're like, come work at Vercel and figure it out here and skip.

Skip calling. Our VIZOR roadmap is moving week by week. Right. It's very hard to say in Q3 2026, VIZOR is going to look this way. We're going to need to find a way to get the universities up to speed on AI, on vibe coding, or start to blur the lines between university and industry. We've been running very successful internship program side of Vercel where I personally spend a lot of time with the interns. Like I'm picking their brain, they're picking mine. Like this intern is like, you know, hanging out with CEOs and CTOs 'cause like I do think that people are getting up to speed really fast if they're in the right room. I do think there's a little bit of a get rid of your ego, forget about everything you are good at and start from first principles. If you thought you were really good at TypeScript, if you thought you were really good at C, well, yes, but also someone that never knew how that technology even worked will probably be generating lots of it just like you very soon. They might even be learning faster than you because they're able to try more things. I'm very contrary in the sense of like people say the junior developer job is dead. I would argue, Maybe there's nothing that's ever been more important than infusing your company with quote unquote junior developers that are more AI native and can show you new approaches truly to learning, to iterating, to figuring out how to solve a problem when you don't know all of the science behind it. When you ask about computer science, computer science is a very specific version of like how they teach you data structures and algorithms, et cetera, if they're leaving out this like pragmatic AI side of engineering. I really liked what Elon said about like, we're nuking the term researchers at Groq, at xAI. We're all engineers. I do think it's a very healthy mindset to have. We're all engineers and we're all figuring it out as we go. And we're all using AI to figure it out as we go.

Taste has become, you know, a keyword in how people think about about the role of humans going forward in AI. And I, I want to use that as a lens to sort of untangle some of the coding, uh, stuff we're talking about. Like, you know, obviously I spend a lot of time thinking about writing. With writing, like, you can, everybody can have taste about stories because we can all like read and assess them. And obviously there are sort of the practitioners who sort of know what goes into making it, but then there have always been sort of critics of, of, you know, novels and writing that sort of— you can imagine with AI, the sort of line between the writer and sort of this, uh, sophisticated critic of writing sort of is collapsing because you can sort of use AI to write over time. I guess with code, my, my question is like, there's sort of the— there's the product, there's like what, what is the thing outputted that you could have taste on? And that's something that seems like it's being democratized, like potentially writing, but won't this sort of code base itself still matter? Or like, will we need engineers that have like the taste for like what is well-written code and like how systems are set up? Or are we just gonna like yield that piece of it to agents? Like, is the taste only about whether the product sort of itself looks good, works well, or is it about like how the actual codebase is set up?

When I first came to Silicon Valley, one of the things that was in vogue was there were a handful of people in the Valley that could create deep tech software systems and also design beautiful products. And I could count them with like one hand. There were a handful of people that I knew and I aspire to become like them. I've cared always a lot about the design and the engineering. To me, that's what was so special about the Steve Jobs and Apple mythology is that they literally had a slide in one of their presentations of like, Apple at its best is the intersection between computer science and liberal arts. And they had a sign of like the intersection of two streets, one called computer science, one called liberal arts. I think the future will be more of that. I would argue that the liberal arts is gonna be overweight because the computer science part, well, the agents will figure out all the data structures, algorithms, optimizations, blah, blah. And so your ability to infuse creativity, taste, culture, design will probably matter even more in the future.

You know, I'm happy to hear the argument for like the liberal arts major, but I still think, you know, tech companies still need to, build things today. And it's not just about like clever billboards and clever marketing.

Yes, but there is a discovery element to technology. Like there's just so much technology that is possible and yet we don't know it's available to us. And so the marketing—

I hear people say this stuff, I'm like, oh, I have a distribution platform. This is what everybody seems to want. I just need to do what you guys are saying is the easy part. Oh, just cobble up some tech thing to sell people. It's like, oh, I have people's attention. It's like, all right. Clearly I'm deficient in figuring out the, you know, tech thing to just like cobble on. But, but to me, from where I sit, it seems harder than, uh, some of the vibe coding takes. It's like, oh, you can't just spin up, you can't be like, oh, I have attention, let me just spin up, uh, software to sell people.

Well, there is a, there is a, a point to there that you're pointing out, which is that even if you figure out a huge platform, you need to create something that's great for that platform. And so if software is getting democratized, like Maybe it's when I mean, when I mean that everybody can cook, a lot of it is like solving your own problems and creating tools for you and your colleagues and things like that matter. Creating software at scale, like if you get the biggest like Times Square sign and you're going to get every eyeball on the planet and you want to give them something of high value, it's still really hard.

Yeah, you can solve your own problems, you know what you need, but solve some sort of general problem for a bunch of humans is— is a hard skill.

That's product building. That is very hard. My argument was there's something really interesting about the Steve Jobs story because he wasn't an engineer, or at the very least he wasn't an engineer in the sense that he's— Wozniak was an engineer. And so there is this huge skillset to figuring out, one, what are the boundaries of the technology? What are the limits of the technology? Number two, how do you best present that technology? Number 3, how do you capture people's attention and distribute it? And so I think Steve Jobs was really good at these things. And then he quote unquote used a team of agents, the people that he recruited to Apple.

Those human beings.

Right. Yeah. But you can kind of see where this is going, right? There will likely be a bunch of Steve Jobs that figured out that trifecta and are offloading the hard engineering work to agents. But also there might be teams constituted almost entirely of Steve Jobs people. Like we're all tastemakers in this room. We all bounce ideas around on like what kind of products we wanna see in the world. We do a bunch of research, we talk to customers, but we're basically like, I would say it's almost like the return of the product manager. I don't know if you remember that iconic moment in culture of the PMs by the pool.

The intersection of engineers resenting product managers and general Twitter loathing of women. Yes.

I'll tell you, like, I don't think people will have seen this coming, right? But there is a return to, you know, well, do you need more than the PM? Which is the person sort of like understanding market customers' requirements, taste, et cetera. And so it's almost like a pendulum swing back on some levels at least. But yeah, I think the reality is likely to be a lot more nuanced than any of this predictions. But my personal prediction is in 5 years, this whole industry does look extremely different, extremely, extremely different.

One 6-month prediction. 6 months from now, what will be different?

6 months, I do think there's gonna be a, maybe even sooner than this, I think people are underestimating GPT-5. I think they're judging it by, again, judging a model is becoming an important industry. Right. There's companies like LMRE.

I know people wanna have quick takes. I'm like, ah, you really need to sit with a model and sort of really understand. That's right.

And that's an emergent thing. I, I don't remember technologies where like you really had to sit with them for like, like getting to know them like a person almost, right?

It's like, yeah, you want to assess their soul like on the first meeting. Yeah, exactly. We have to sit with it and sort of see what it can do.

It's like, uh, it's like your 30, 60, 90-day plan for a new hire or a new executive. Like, you have to treat it—

I think people are underestimating GPT-5.

Yeah, I think people are. I really think people are. I also think you are not realizing that this vertical agents. There's probably gonna be some agents, maybe some like agents for video editing or creating video or creating ads or, you know, I can't tell you exactly which. There's gonna be one that we're all gonna be talking about.

Interesting. Yeah, yeah. We need to break out ChatGPT as so much to find the use case. We need sort of a vertical niche where it's like, it solves this problem, do this thing.

You know, calendar, something along this lines. Like it seems so op, But how do we not do it as soon as ChatGPT came out? Well, we didn't do it because we didn't have the infrastructure, we didn't have the best practices, we didn't have the AI gateways, we didn't have the agent frameworks. We've built a lot of this stuff. And so there's gonna be an obvious in retrospect application with a huge user base. That's my perspective.

You're saying right now that Gmail is not solving, like that the model is not in my email yet to me is the craziest. I mean, Superhuman wants to do it, but to me like Gmail, it's like sitting there. There's a lot of risk, obviously. I would be a bad reporter if I didn't ask. You can dodge or not. The information says you're getting approaches at $9 billion. You have anything to say about that?

No comment. No comment.

Speculation. All right, Guillermo, thank you so much for coming on the show. We have to do this again. This was a lot of fun.

Thank you, Eric. Always a pleasure.