Of Course There’s an AI Bubble

Summary

Paul and Rich confront the elephant in the room: the AI bubble. Sparked by a LinkedIn post from Cross Data co-founder Chris Pasarski showing that vibe coding tool Lovable’s web traffic has dropped 50% in just three months, they dissect why vibe coding tools are failing to deliver on their promise. Rich lays out two fundamental reasons: first, good software requires extreme explicit detail that AI never asks for — it just makes assumptions and goes. Second, AI produces “language without thought” — it doesn’t know why it’s writing code, doesn’t understand that a restaurant website needs reservations that lead to people eating at tables, and creates “long endless strings of liability.”

The conversation pivots to the broader AI bubble question. Paul makes a nuanced argument: we’re in a “financial industry focused AI bubble,” not a technology bubble — the tech is real but severely misallocated. Currently three companies are worth half a trillion dollars each, when what the industry actually needs is 20,000 startups building domain-specific products. They compare the coming crash to the 2001 dot-com bust, reminiscing about the messy, wonderful post-crash era of bodega beer meetups and XML demos that gave birth to Web 2.0. Their prescription for the vibe coding startups: verticalize now while you still have money, because the big LLM providers can squeeze you any time they want.

Highlights

”AI does not know why it’s doing anything”

yt-dlp --download-sections "*14:30-15:30" "https://www.youtube.com/watch?v=eDwNbCGHcTA" --force-keyframes-at-cuts --merge-output-format mp4 -o "eDwNbCGHcTA-14m30s.mp4"

“AI does not know why it’s doing anything. It is not driven by purpose or goal or even a subset of goals. It is just expressing code. It does not distinguish code from a poem, from an essay. It’s another set of tokens.” — Rich Zotti, 14:30

”Language without thought”

yt-dlp --download-sections "*16:00-17:00" "https://www.youtube.com/watch?v=eDwNbCGHcTA" --force-keyframes-at-cuts --merge-output-format mp4 -o "eDwNbCGHcTA-16m00s.mp4"

“There’s a formulation the writer Rusty Foster came up with which is ‘language without thought.’ We’ve never had language without thought before and it’s really confusing because we just expect the thought to be there.” — Paul Ford, 16:00

”We are in a bubble”

yt-dlp --download-sections "*23:40-24:40" "https://www.youtube.com/watch?v=eDwNbCGHcTA" --force-keyframes-at-cuts --merge-output-format mp4 -o "eDwNbCGHcTA-23m40s.mp4"

“Are we in a bubble? Yeah, we are in a bubble. I just got an email from Bloomberg this weekend. They’re like, hey, what would an AI crash look like?” — Paul Ford, 23:40

”I believe we are in a financial industry focused AI bubble”

yt-dlp --download-sections "*24:50-26:00" "https://www.youtube.com/watch?v=eDwNbCGHcTA" --force-keyframes-at-cuts --merge-output-format mp4 -o "eDwNbCGHcTA-24m50s.mp4"

“I believe we are in a financial industry focused AI bubble. I do not believe that AI as a technology is severely overvalued. I think that everything is misallocated.” — Paul Ford, 24:50

”Give me 20,000 startups, not five”

yt-dlp --download-sections "*28:50-29:50" "https://www.youtube.com/watch?v=eDwNbCGHcTA" --force-keyframes-at-cuts --merge-output-format mp4 -o "eDwNbCGHcTA-28m50s.mp4"

“Give me 20,000 startups, not five. Get me there and then you’ll find that we’ll be back here after the crash, but it’ll be sustainable.” — Paul Ford, 28:50

”Most software projects never complete”

yt-dlp --download-sections "*12:20-13:10" "https://www.youtube.com/watch?v=eDwNbCGHcTA" --force-keyframes-at-cuts --merge-output-format mp4 -o "eDwNbCGHcTA-12m20s.mp4"

“Most software projects never complete. And so like whatever the LLM learned, I doubt it learned how to finish code because there are no precedents.” — Paul Ford, 12:20

Key Points

• Lovable Traffic Down 50% (2:14) - Chris Pasarski of Cross Data posted that Lovable and other vibe coding tools have seen traffic halve in three months
• Benioff Goes MAGA (2:21) - Salesforce’s Marc Benioff suggested Trump should bring the National Guard to San Francisco
• Vibe Coding’s Two Use Cases (3:40) - These tools either host your code or let you eject and continue working with it
• Number Two AI Use Case (4:40) - After chatbots/question-answering, vibe coding tools are the second biggest AI use case
• Paywall Squeeze (5:30) - Vibe coding tools are showing less value for free and pushing users to pay
• Dependency Vulnerability (6:00) - These startups are repackaging output from OpenAI or Anthropic, who can squeeze them anytime
• First Mile Magic (6:40) - AI is extraordinary at starting projects but terrible at finishing them
• Problem One: No Questions (10:30) - AI never tosses work back to you with questions; it makes assumptions and builds
• Problem Two: No Intentionality (14:30) - AI doesn’t know why it’s coding; it’s language without thought
• One-User Auth Bug (11:20) - Someone’s AI-built app only allowed one user to log in because they didn’t specify otherwise
• Replit’s 20-Minute Fix (15:40) - Vibe coding tools trying to solve quality by running the LLM over the code repeatedly
• Financial AI Bubble (24:50) - Paul argues it’s a financial bubble with misallocated capital, not a technology bubble
• 20,000 Startups Needed (28:50) - The industry needs 20,000 domain-specific startups, not five giant companies
• Post-Crash Nostalgia (29:00) - After the dot-com crash, people met at houses to demo web things over bodega beer
• The Dollar Test (32:00) - Whatever AI does for you should cost no more than a dollar of LLM time in five minutes
• Aboard’s Approach (32:20) - Their product builds a business application in about 10 minutes for a couple dollars in LLM time

Mentions

Companies

• Lovable (2:30) - Vibe coding tool that saw dramatic traffic decline
• Replit (3:30) - Vibe coding platform trying to fix quality by doubling back over
• Bolt (3:30) - Another vibe coding tool mentioned
• V0 (3:30) - Vercel’s vibe coding tool
• Retool (3:30) - Development platform mentioned
• Base44 (3:35) - Another vibe coding tool
• OpenAI (6:00) - LLM provider that vibe coding tools depend on
• Anthropic (6:00) - LLM provider; launched Claude Code which competes with vibe coding tools
• Salesforce (2:21) - Owns Slack; Benioff went political
• Cross Data (2:14) - Y Combinator company that tracks web traffic data
• Bloomberg (23:50) - Asked Paul what an AI crash would look like
• Sequoia (27:20) - Referenced for their famous “batten the hatches” memos during downturns
• Google (27:40) - Startup heaven is getting acquired by Google
• Aboard (1:15) - Their product builds business apps in 10 minutes for a couple dollars

Products & Technologies

• Cursor (31:10) - AI coding tool from the big providers
• Claude Code (6:10) - Anthropic’s built-in vibe coding tool that comes with subscription
• Expression Engine (15:00) - Old PHP platform that was like a for-pay Drupal

People

• Paul Ford (0:00) - Co-host
• Rich Zotti (0:02) - Co-host
• Chris Pasarski (2:14) - Co-founder of Cross Data, wrote “Is Lovable dying?” post
• Marc Benioff (2:25) - Salesforce CEO who made political comments about San Francisco
• Rusty Foster (16:00) - Writer who coined the phrase “language without thought”
• Aaron Ross Sorkin (24:10) - Financial journalist on 60 Minutes discussing bubble, wrote a book called “1929”
• Sam Altman (25:30) - OpenAI CEO referenced with sovereign wealth meetings

Surprising Quotes

“I guess when I wanted it to do authentication, I should have realized I needed to tell it that more people would be logging in through Google than me.” — Reddit user via Paul Ford, 11:20

“We created an agency and were successful in the past simply because we would finish things. And people would come to us and say, ‘You’re so expensive and I hate you and I hate my life that I have to talk to you.’” — Rich Zotti, 12:50

“Whatever the LLM learned, I doubt it learned how to finish code because there are no precedents.” — Paul Ford, 12:20

“You cannot universalize a shortcut. There is an incredible shortcut that comes with AI. But to think that magic trick can just now be applied everywhere — that is driving the bubble that exists today.” — Rich Zotti, 21:40

“After the dot-com crash, I used to go and meet at people’s houses to talk about tech. You’d bring bodega beer. You’d be like, ‘Hey, look, I got some Newcastle.’ And then everybody would do a demo of their new web thing. It was wonderful.” — Paul Ford, 29:10

Transcript

0:00 Hi, I’m Paul Ford. And I’m Rich Zotti. And this is the Aboard Podcast. The podcast about how AI is changing the world of software. And you know, Rich, there’s lots to talk about, but are we — is this a day of reckoning? Are we in a bubble?

0:44 Okay, before we talk about whether the entire AI industry is going to die and how vibe coding is losing all of its moxie, we should talk about our AI powered better than vibe coding solution that we’re building. Yes, trust us. Everything’s going to be great. Aboard uses AI to accelerate the entire software development life cycle. And to be clear, we’re a product. We make a thing. You use it with us.

1:28 Yes, we ship enterprise software. We ship big software. And big software cannot be automated away. We send people out to help you. But boy, we do it 10 times faster. We are your partner for AI acceleration. So check out aboard.com. Keep an eye on it because there’s some new wonderful things that are about to happen.

2:07 Okay. I like to read the internet. Yeah. Lots of stuff on there. Incognito, baby. So, one of the things that was just shared in the Slack here at Aboard. So, you know, Slack, Salesforce owns Slack. Benioff just went full MAGA. He did. He’s like, Trump should bring the National Guard to San Francisco. Yeah, it doesn’t need the National Guard. That’s the last thing it needs.

2:40 Anyway. So our head of design shared a post. New person to me, Chris Pasarski, co-founder of Cross Data, which is a Y Combinator company, real time B2B data kind of thing, tracks the web. Big old LinkedIn post from Chris and the first three words are, “Is Lovable dying?” Dying. Lovable is a vibe coding tool.

3:10 Let’s workshop this headline a bit. Is Lovable less lovable? So what is Lovable? Lovable is a vibe coding tool. And there’s a few — there’s Replit, Bolt, V0, Retool, and us too in a way. Base44. You go and you type I want a website that, or I want an app that. You make software with AI is the promise. Generally spins up a coding environment and writes a lot of code for you.

3:50 And then there are two kind of end states. One is maybe it will host that code that it wrote for you and that will talk to a database and all sorts of things that they manage for you. Or you’ll kind of hit eject, take that code with you and continue to work on it like any code.

4:10 Now there’s a tricky thing happening. According to Chris Pasarski, there was a real viral wave over the summer around vibe coding. Everybody got real excited. It’s worth noting besides the self-service answer-questions, make-an-image uses of AI which dominate, the number two use of AI has been this explosion in tools that are going to make apps by coding.

4:40 It’s one of the ones that’s most professionally transformative. So Lovable is one of those tools. Using web traffic as a proxy, this person is claiming that vibe coding is half as exciting to the world as it was in the summer. And this actually makes sense. Whoa, hold on. Summer. That’s three months ago. That’s not just a flattening out of the chart. Down 50% in three months.

5:10 In particular, Lovable really peaked, had a big viral pop. Everything with a grain of salt. But what I have noticed from the vibe coding tools — there’s less value available at the prompt for free with all of them. You hit this wall really quickly where they’re like, gotta pay. It’s expensive for them by the way because they have to pay one of the large AI providers like OpenAI.

5:50 I feel for them because they found something really neat, it got everybody really excited, but they have this external dependency that can squeeze them as hard as it wants. And Anthropic and OpenAI have launched really good vibe coding tools that come with your subscription. Built in. You’re touching on vulnerability number one here.

6:20 These startups, they’re big, they have millions of users, they’re very well funded, they’re valued sometimes in the billion-plus space. They are repackaging output from one of the large LLM providers. That is what they are doing. Issue number one is you can repackage and doll up something that someone else delivered and that only goes so far.

6:50 There’s another thing too. The way I would describe AI is that the first mile of a project is extraordinary because it seems so human and it fills in the blank page. It writes you all this code and to be fair getting started is the hardest part in software. So is finishing. And so that last mile is really long. Those are the two hardest parts. Getting started it’s really good. Finishing — it doesn’t work.

7:30 Code is deeply complex. Code coming together at a large scale is really really hard. It doesn’t work. To be fair, most software projects never complete. And so like whatever the LLM learned, I doubt it learned how to finish code because there are no precedents. We created an agency and were successful in the past simply because we would finish things.

8:00 People would come to us and say, “You’re so expensive and I hate you and I hate my life that I have to talk to you.” And we’d say, “I know, but we will ship your code.” And we did. And so there’s no place for the LLM to learn how to get something done. And so people are learning this because they’ve had this explosive moment where it put the app on the screen with the buttons. Holy crap. That was kind of what I thought it would look like.

8:30 And then you go to Reddit and there are all these stories. I can’t believe how bad this is. I read one the other day — I guess when I wanted it to do authentication, I should have realized I needed to tell it that more people would be logging in through Google than me. It created an authentication system for only the one user. And this is normal by the way. A human being would never have made a one person auth system.

9:10 I am going to become famous because I’m going to tell you the two reasons why vibe coding as it exists today doesn’t work. Straight up. Number one. You just backed into it. Good software is a product of extreme explicit detail. If you don’t have explicit detail — for example, anyone can sign up to the service via Google OAuth and log in later. If they forget a password, it will email them back. It supports two-factor auth.

9:50 Engineers are people who actually live inside of this detail. And as they get more senior, that knowledge is embedded in their craft. So if I’m an engineer and I say I need to authenticate through Apple, Google, and possibly Facebook, I know that there are 70 or 80 things that are implied by that statement.

10:10 And so the thing that AI does not do almost ever is toss work back to you. It just goes. A good product manager, like a good physical architect, like a good industrial designer says, “I got a lot of questions for you, because I got to fill in those blanks. If I don’t fill in those blanks, I’m going to fail you.”

10:40 And today the posture of AI is to go. It does feel like magic at first because you wrote one sentence and it just starts unraveling all this code at you. But what it’s doing is making a host of assumptions about what you want without asking. And that’s problem one.

11:00 Do you want problem two? And then I become famous. Problem two is much more nuanced. AI does not know why it’s doing anything. It lacks consciousness. It is an expression engine. It will express itself all day. It is not driven by purpose or goal or even a subset of goals. It is just expressing code.

11:30 It literally does not distinguish code which is actually going to be sent into a machine to be run, turned into working tools. It does not distinguish code from a poem, from an essay. It’s another set of tokens that it produces. Now you say, “When I say make me a website for my restaurant, it kind of made me a restaurant website.” And the reason it did that is because it found code that is related to those similar words because other restaurant websites have been created.

12:10 It does not know it’s a restaurant website. It doesn’t know that you want to make reservations on it. And so what happens is this chase occurs with vibe coding. I need to book reservations on my site and then it goes and does the same thing again. It has no understanding that reservations lead to people eating at tables in a place where food is given to them. It doesn’t know anything.

12:40 And so as a result it starts to create what I call long endless strings of liability. Code that looks like code to us but is in fact oftentimes voluminous because it doesn’t mind, it doesn’t try. There’s a formulation the writer Rusty Foster came up with which is “language without thought.” We’ve never had language without thought before and it’s really confusing because we just expect the thought to be there.

13:10 There’s no intentionality. And what drives me bananas is that’s amazing. That’s the weirdest database ever and we should be having a really good time with it and talking about it and figuring out where it begins and ends. But we decided that we were going to have Star Baby Jesus instead.

13:40 This “language without thought” is such a profound message. We just did a podcast on Friend, the pendant you wear because you’re lonely. And language without thought in that context is even more depressing. It faked loving you. It faked caring about you. It told you how great you were only to manipulate you the whole time. The behavior we typically associate with that is sociopathy.

14:20 So let’s talk about the most recent trend. The vibe coding startups are on to this challenge. What I’m seeing them do is they’ve decided to solve it by doubling back over. I hit Replit last week. I needed 20 minutes. And you got to get your credit card out. What that is is an attempt to take all that chaos that unfurled without thought and somehow keep sanding it down to maybe till it gets to a point.

15:00 I’m going to tell you my hypothesis. The future state is not throw the LLM at it over and over again. Because what we have learned is that actually never quite works because the intentionality gets divorced from the product. Your brain can’t stay on top of it enough. It’s impossible.

15:30 It just ends up spewing a bunch of stuff and you keep trying and eventually sort of trial and error gets something together, but you’re going to keep finding bugs like it only lets one person log in. Now it can get smarter. God bless us all. That’s the boil-the-ocean approach. We should point out how these things are useful. Great at starting, great at prototyping, great at exploring problem space.

16:00 Walk it over to engineering. They’re wonderful at expressing alternative algorithms. These are really good acceleration tools. Good at looking through code and finding bugs if guided. Like there’s a million ways to use them. But that kind of one-shot got my app ready for market — here’s where I think we’re headed.

16:30 It works really well when there’s really well-defined scaffolding and blueprint. What’s going to happen is I don’t think you’re going to be able to have a one-size-fits-all that writes every kind of code. I’m not going to have one tool that lets me write a 3D game that also lets me write an ERP tool that lets me write a CMS.

17:00 I think if you look at the future of this stuff, instead of solving every single problem in the world, you’re going to end up with situations where people choose a relatively narrow but still broader than ever before domain and then let it code. Scaffold-driven development. AI-assisted development. It helps you get the plan together. It’s good at boiling stuff down.

17:30 We’re going to be back to a world of applications that are specific to domains as opposed to the one-does-everything super app. If there’s one takeaway — you cannot universalize a shortcut. There is an incredible shortcut that comes with AI. But to think that magic trick can just now be applied everywhere — we’ve seen that before. And I think that is driving the bubble that exists today.

18:10 Because everyone assumes that I can apply this — the cures for diseases are coming. You don’t need lawyers anymore. No. Product is reasserting itself and instead of one meta product, we’re getting lots of little products.

18:30 So let me ask you this. Are we in a bubble? Yeah, we are in a bubble. I just got an email from Bloomberg this weekend. They’re like, hey, what would an AI crash look like? There are people — it’s in the conversation right now. I’m getting a lot of bubble DMs.

19:00 Aaron Ross Sorkin was just on 60 Minutes talking about a bubble. He just wrote a book called “1929.” Thanks, Aaron, for really cheering us on. He thinks we’re in a bubble. We are in a bubble. This is the classic overenthusiasm.

19:30 I’m going to tell you what I really believe here. I believe we are in a financial industry focused AI bubble. I do not believe that AI as a technology or that tech in general is severely overvalued. I think that everything is misallocated. Explain.

20:00 Right now we’ve got these giant companies that are worth half a trillion dollars at a pop or more. Insane. Let me counter just real quick. This is the fastest growth and adoption of a technology ever. No doubt. It is fundamental. People’s credit cards are out. But it’s not at a $400 billion investment level. There’s so much money pouring in. Also, 20 bucks a month, even if you get everybody on earth to sign up, arguably doesn’t fully justify the value.

20:40 Let me talk about resilient companies. One of our best agency relationships that continues is with the insurance industry because hurricanes don’t stop. Insurance companies have vast enormous roadmaps that they need to service. They’re also bureaucratic. If you’re a 27-year-old Nordic guy and you want to do a startup, you don’t want to go towards insurance. But when the recession hits, incredibly resilient.

21:20 Insurance is going to keep needing digital products all day long. Another one that’s really wild — nonprofits. They’re getting hit by federal funding cuts. They’re losing money right now. They are actually being forced to be more resilient. So one of the first things they do is look at their IT spend.

21:50 You’re saying this valuation is fantasy. When the squeeze comes — of course it is because people keep pouring money in. Everybody grabs you by the shoulder like, “Hey, we’re in a bubble, but I need to put another couple million dollars into this.” No one knows when the ride ends. It’s — everybody is like, “I’m not the greater fool.” That’s how bubbles happen.

22:20 If we’re really in a bubble and a crash is coming — what that means is all the people who confidently say our VC thesis is that only companies in this tiny hot zone are worth investing in. And never should anyone talk to a human being. And everyone will pay for this forever. Every time this happens, you get that same deck from Sequoia — “batten the hatches, boys.”

22:50 And a large percentage of startups go to startup heaven, which is literally just getting acquired by Google. Free coffee. Heaven’s nice. Got a ball pit. When is it going to pop? I mean, literally any day. Why wouldn’t it? Because of that last mile problem, everyone’s been lying about where this thing can get you.

23:20 There’s a ton of product work to be done and there’s probably like 10 or 20,000 companies that need to be created in order to actually materialize the value in this technology. I see the value. I believe in it. But I need 20,000 companies to exist before we can really get there. It’s not just going to be faster agents or bigger LLMs. It is like a zillion little products landing.

23:50 OpenAI can’t capture all this value. Anthropic can’t capture it. You need like half a million people messing with this stuff. Are you surprised that things consolidated to like three companies so quickly? No, because that’s the way the valley works. The valley wants to create that tight, super hot market. Sovereign wealth shows up. It’s just a picture of Sam Altman in Qatar.

24:20 Is that bad? I’m done with bad and good. I like messy and I like after the crash because everybody has to figure out what it’s good for. After the dot-com crash, I used to go and meet at people’s houses to talk about tech. Because we just loved it and everybody else went home.

24:50 Everybody went home and I started my business two years after the crash. You’d bring bodega beer. You’d be like, “Hey, look, I got some Newcastle.” And then everybody would do a demo of their new web thing. It was wonderful. People would go gather and talk about XML, right? And out of that comes Web 2.0. And out of that comes all the new wonderful things and mobile starts.

25:20 The Valley basically likes to kind of own the narrative. And they’re smarter about this than we are. They win a lot. But give me 20,000 startups, not five. Get me there and then you’ll find that we’ll be back here after the crash, but it’ll be sustainable. Yes, there is value here somewhere.

25:50 This was extremely New York City. Today’s podcast is brought to you by New York City. It’s bad. And we will absolutely put a socialist in as mayor. One of the most expensive cities in the world. One of the toughest cities to survive in. To have a family in. Everything is a nightmare.

26:20 If you want to know how to come out of a bubble, go on YouTube. Search for Eyewitness News WABC 1978. You will see a city on its knees. Bankrupt. They can’t fix parking meters. Everything’s so bad all the time.

26:50 Let’s close it with a question. These startups that are riding this wave right now, you pointed out 50%. They’re all declining. Will they survive? I mean, frankly, I hope so because they all innovated real early and they took a real big swing. I think what they need to figure out — we focus on business applications, partnerships instead of doubling down on consumer product.

27:20 I don’t know if we can have 12 of them really big and also have Claude Code and Cursor and — the big guys have the toys as well. That’s a rough moat to continually maintain. Because they’re riding on top of them. Where they can thrive is by verticalizing. They actually have good platforms. Figure out like five markets and chase them because you got a lot of money right now.

27:50 That’s not interesting to a lot of that world. What they want is to be the next trillion dollar company. But I would actually pause at this moment. I wouldn’t want to keep competing on “it will spit out the code, I promise, this time just give it three hours.”

28:20 That feels like a pretty big bet, but they’re all making it. And everything with this stuff has been so weird that three months from now, I could look back on this and be like, “Boy, they really sailed through.” We’ve been right so far. There’s no research horizon where it gets unlocked. There’s no 10% optimization.

28:50 Here’s the future everybody should aim for. Whatever it can do for you, it should do it for no more than a dollar of LLM time in five minutes. That’s a challenge. Yes, it’s very hard. We don’t always hit it. But that five minutes should get down to one minute. For a dollar I should be able to do the thing I’m going to do for you.

29:20 That’s a constraint. We had a podcast about DeepSeek and constraints and how constraints force you to be creative. It’s too much endless coins right now. Our product builds a business application for you in about 10 minutes. And it does it for a couple dollars in LLM time max. And that’s because we put the constraints in.

29:50 A little jab there at the end. A little sly pitch. Or a fact. Check us out at aboard.com. Paul, this was a lively conversation. You think we’re in a bubble, too? I do. Yeah, me too. Have a lovely week. Good luck to your portfolios. Bye.