Masad was explaining why companies need an outside platform to govern AI agents, building on the idea that agents behave like insider threats. It is his pitch for Replit as a controlled home for agents.
Masters of Scale · 2026-11-17
Permalink →
Listen →
Masters of Scale
Around 12:18 into the episode
Amjad Masad
…startup legal documents all in one guided process. And once your company is formed, which for most people happens in just one day, you're ready to begin accepting payments on day one. More than one in four new Delaware corporations is formed with Stripe Atlas. Learn more at stripe.com slash Atlas.
Bob Safian
…they could build a version of what Replit is doing. Some of them are trying to. How do you create a moat that protects against that? You know, what like is your incentive? We were talking about. This is a little bit earlier, sort of the difference between your incentive and maybe their incentives.
Speaker 4
Yeah. Yeah. So when you're bringing AI into your organization, actually, Satya wrote a very good essay. He said, like, AI agents are insider threats. The way I like to visualize it is: imagine inviting an employee into your office. That employee is very intelligent, very productive, but it's slightly insane and can go off the rails and might be an actual security threat to you and the other employees. Do you rely on him for or her for the security? Would you say, oh, go secure yourself? Probably not. You would want like an external security company or like an internal platform. So the way we think about Replit is it's like the agent habitat. It is the place for agents to be productive, but also secure, governed. Like the way we architected our system is based on this zero trust principles and security. We spent a lot of time working on the sandbox and the operating system underlying the sandbox. One interesting stat is there was like 100 zero day vulnerabilities over the past year. Replit was affected by about zero of them because we actually build sort of a special version of the kernel that is…
Bob Safian
that I'm using those tokens
Speaker 4
most effectively. In a most effective way, efficient way, and so on. And I think the labs lack those kind of incentives.
Bob Safian
I mean, another sort of day after tomorrow topic would be that some people claim that VIBE coding can break down when it's deployed at scale. Is that a legitimate concern? It
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
Asked whether vibe coding breaks down at scale, Masad said that was true until recently, but agents can now write in safer languages and do formal verification humans found too laborious. It cuts against the story that AI-written code is making software less safe.
Masters of Scale · 2026-11-17
Permalink →
Listen →
Masters of Scale
Around 15:35 into the episode
Speaker 4
most effectively. In a most effective way, efficient way, and so on. And I think the labs lack those kind of incentives.
Bob Safian
I mean, another sort of day after tomorrow topic would be that some people claim that VIBE coding can break down when it's deployed at scale. Is that a legitimate concern? It
Speaker 4
…NASA space shuttle and you're writing the software for that, oftentimes you might spend a year trying to formally verify, meaning make mathematical proofs about the program you're writing. And you would do it in the most extreme life or death scenarios because it was very laborious, hard work. Now we're having AI do formal proofs over applications because it's become very cheap. So I will say we've crossed the threshold to the point where automated software engineering is safer than human engineering.
Bob Safian
So do you have a vision about what the day after tomorrow looks like? You know, in this extraordinary dynamic moment in history, you're at the center of it. You see, like, can you see what the future is going to be?
Bob Safian
Or is it too soon? Or, you know, is there a clear horizon?
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
In the same answer on AI risk, Masad turned to the week's security news, saying his team had been in the office overnight mitigating new vulnerabilities. It speaks to the wave of AI-found bugs and shrinking disclosure windows.
Masters of Scale · 2026-11-17
Permalink →
Listen →
Masters of Scale
Around 09:01 into the episode
Speaker 4
…with Steve Jobs' kind of reality distortion field. Really, every founder needs to do some reality distortion. You're selling a dream and suddenly you're like, we're running out of money. The dream is kind of immediately kind of dissipates. But I'll bet those people who quit now, they're saying, oh,
Bob Safian
can I come back? Yeah, some of them came back, actually,
Speaker 4
which we were always welcoming people to come back. But there was this juxtaposition where this cool, dark office where everyone quit, we had a war room. I saw your war room kind of whenever we have a big idea that we want to do, we always kind of take this like YC demo day mindset of we're going to just put, you know, five, 10, 12 people. In a room, and we're just going to work really hard towards a milestone. We're going to do it in a few months. And there we were inventing agents. So I would go into this room, enormous energy, excitement. People are seeing the future on their screens for the first time. We see this entity coming to life that's like able to code and run the code and deploy. And then I would go out, and then the rest of the team felt like total despair. And I think the lesson from that is there's always hope. There's always kind of something you could do to turn things around. And there's a group of people at your company that are going to be motivated to do that. Well, let's go to the next movie we picked here, which is Twister. From the outside, it can seem like from the time you were at a couple of million dollars to a…
Bob Safian
The fact that Amjad can refer to anything that's happened at Replit in the past two years as slow and steady is kind of remarkable. And it underscores what's so different about the pace of a true AI-native business. So is he at all worried that the big AI players will copy what Replit's doing? And…
Amjad Masad
When you've built substantial wealth through your business, it's often tied up in a single equity position. The upside is real, but so is the risk. And knowing when to act isn't always obvious. Creative Planning works with business owners to build a strategy around concentrated equity, when to…
Bob Safian
Before the break, Replit's Amjad Masad talked about how Replit went from a $3 billion valuation to $9 billion in just six months. Now he talks about the competitive risks ahead posed by the biggest AI firms, how vibe coding is changing the work world and more. Let's jump back in. Let's go to the…
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
Asked whether fear of rogue AI agents is itself a kind of "AI psychosis", Masad said the real near-term risk is cybersecurity, and that talk of machine consciousness and extinction distracts from it.
Masters of Scale · 2026-11-17
Permalink →
Listen →
Masters of Scale
Around 09:01 into the episode
Speaker 4
…with Steve Jobs' kind of reality distortion field. Really, every founder needs to do some reality distortion. You're selling a dream and suddenly you're like, we're running out of money. The dream is kind of immediately kind of dissipates. But I'll bet those people who quit now, they're saying, oh,
Bob Safian
can I come back? Yeah, some of them came back, actually,
Speaker 4
…Yeah. In the Twist neighborhood, you've decried what you call AI psychosis. You've called out Dario Amodi specifically at Anthropic for overheated predictions about AI-fed job losses. Today's fear over AI agents run amok. Is that also psychosis? Or is that a different AI risk? No, in fact, that's. That is really the thing that I've been talking about, the risk that I've been talking about. I've been on the cybersecurity risk for two years now. I think everything else is a distraction from that. Like when they go out and talk about how AI is conscious and they go meet the Pope and do all this random stuff, it's really distracting and also just in many ways destructive and bad for humanity for us to go care about the well-being of like these machines when we haven't really figured out the well-being of humans. And also, it distracts from the real question of cybersecurity. Just in the…
Bob Safian
The fact that Amjad can refer to anything that's happened at Replit in the past two years as slow and steady is kind of remarkable. And it underscores what's so different about the pace of a true AI-native business. So is he at all worried that the big AI players will copy what Replit's doing? And…
Amjad Masad
When you've built substantial wealth through your business, it's often tied up in a single equity position. The upside is real, but so is the risk. And knowing when to act isn't always obvious. Creative Planning works with business owners to build a strategy around concentrated equity, when to…
Bob Safian
Before the break, Replit's Amjad Masad talked about how Replit went from a $3 billion valuation to $9 billion in just six months. Now he talks about the competitive risks ahead posed by the biggest AI firms, how vibe coding is changing the work world and more. Let's jump back in. Let's go to the…
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
Asked how Replit can defend itself when Anthropic, OpenAI, Google and Microsoft could build something similar, Masad argued that its sandboxing is the moat and that the labs lack both the security expertise and the incentive. It is a pointed claim about AI labs profiting from the security problems their own models create.
Masters of Scale · 2026-11-17
Permalink →
Listen →
Masters of Scale
Around 12:18 into the episode
Amjad Masad
…startup legal documents all in one guided process. And once your company is formed, which for most people happens in just one day, you're ready to begin accepting payments on day one. More than one in four new Delaware corporations is formed with Stripe Atlas. Learn more at stripe.com slash Atlas.
Bob Safian
…they could build a version of what Replit is doing. Some of them are trying to. How do you create a moat that protects against that? You know, what like is your incentive? We were talking about. This is a little bit earlier, sort of the difference between your incentive and maybe their incentives.
Speaker 4
Yeah. Yeah. So when you're bringing AI into your organization, actually, Satya wrote a very good essay. He said, like, AI agents are insider threats. The way I like to visualize it is: imagine inviting an employee into your office. That employee is very intelligent, very productive, but it's slightly insane and can go off the rails and might be an actual security threat to you and the other employees. Do you rely on him for or her for the security? Would you say, oh, go secure yourself? Probably not. You would want like an external security company or like an internal platform. So the way we think about Replit is it's like the agent habitat. It is the place for agents to be productive, but also secure, governed. Like the way we architected our system is based on this zero trust principles and security. We spent a lot of time working on the sandbox and the operating system underlying the sandbox. One interesting stat is there was like 100 zero day vulnerabilities over the past year. Replit was affected by about zero of them because we actually build sort of a special version of the kernel that is stripped down so it has less surface area to attack. So we do a lot of work there.…
Bob Safian
that I'm using those tokens
Speaker 4
most effectively. In a most effective way, efficient way, and so on. And I think the labs lack those kind of incentives.
Bob Safian
I mean, another sort of day after tomorrow topic would be that some people claim that VIBE coding can break down when it's deployed at scale. Is that a legitimate concern? It
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
Graham was recalling his own horrified reaction to Masad's early agent demo. Masad says all engineers will go through stages of grief over AI writing code, and that he went through them early.
Y Combinator Startup Podcast · 2026-11-07
Permalink →
Listen →
Y Combinator Startup Podcast
Around 12:15 into the episode
Amjad Masad
No, everyone was realizing that. Like, I think I've been seeing like engineers that were also repulsed by that. I remember your tweet was very controversial when I tweeted that.
Paul Graham
That guy who generated vast amounts of code using AI. Yeah.
Amjad Masad
Well, even that tweet about me being the one. Bill and object code and all of that was also controversial. But now all engineers are, you know, I think with AI, everyone will have to go through this period of, you know, denial and then, and then sort of later on, you know, depression and then, you know, eventually acceptance. Yeah. I've had it like in like 2020, you know, really. And so, and so it just makes it easier for me. Like, I had the grief of the craft that I love is gone. I've already grieved. You
Paul Graham
felt that too. Yeah. Okay. So you knew what I was feeling. I
Amjad Masad
knew what you were feeling. When you showed me that demo. Yes. But you had to feel it so that you can let go.
Paul Graham
I think everybody got convinced at the end of last year. At the end of 2025 is when everybody realized, okay, AI is going to write all the code.
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
Graham asked whether Replit is 'living in the future'. Masad answers with a forecast about agent counts per person, arguing that this scale forces work off the laptop and into the cloud.
Y Combinator Startup Podcast · 2026-11-07
Permalink →
Listen →
Y Combinator Startup Podcast
Around 37:46 into the episode
Amjad Masad
the future for sure.
Paul Graham
More even than AI specifically.
Amjad Masad
It's often frustrating. Like, I think that. Because stuff doesn't work. Well, stuff doesn't work and people change so slowly. So, for example, cloud development. Are people convinced about that now? Oh, yes. Still. How are they going to use AI? It's crazy. Exactly. So our prediction internally is that you go from everyone using one agent to 10 agents to 100 agents to 10,000 agents at some point. You can't run 10,000 agents on your MacBook. There's no way you can do it. Yeah,
Paul Graham
Robert proved that back in the 1980s.
Amjad Masad
Yes. So eventually we're going to get away from local work in general. But there's like people still predominantly do work on their desktop.
Paul Graham
Maybe this is what will finally convince them. Yes.
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
Graham asked whether the shrinking agent harness makes it easier for competitors to catch Replit. Masad says yes, and describes the treadmill of constantly moving on to the next thing.
Y Combinator Startup Podcast · 2026-11-07
Permalink →
Listen →
Y Combinator Startup Podcast
Around 05:27 into the episode
Amjad Masad
…you make it work? And the crazy thing, Paul, is that the harness circuit 24 is 10 times more complex than the harness today because the models were less capable. Yeah. So as models get more capable, like what the kind of the engineering you have to do to make them become agents is less and less so.
Paul Graham
Does that mean it's easier for people to compete with you?
Amjad Masad
Yes. But we have to do, well, that's the thing about the AI business today is that you have to just accept that everything that you're doing is like getting commoditized constantly. And you have to like zoom in. Be doing
Paul Graham
the next thing. Be
Amjad Masad
doing the next thing, right? But the other big thing that we demonstrated is that AI can provision infrastructure. So up until that point,
Paul Graham
you actually have infrastructure to provision. Yes. That's the thing that's different between you and all these competitors, right?
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
Paul Graham asked whether Replit's enterprise customers are mostly non-tech companies. Masad describes banks and manufacturers where a few internal 'vibe coders' start replacing expensive legacy software.
Y Combinator Startup Podcast · 2026-11-07
Permalink →
Listen →
Y Combinator Startup Podcast
Around 26:53 into the episode
Amjad Masad
…corporations, banks, you know, things like that. And they have a lot of like SaaS software they pay $10 million for or like old software. And they find like a few Vibe coders inside the company that have adopted Replit and they're going and like replacing all these different systems. And
Paul Graham
it starts with internal stuff.
Amjad Masad
It starts with internal stuff. Yes. And previously they would tell us that we've actually scoped this project out and it would have been $10 million. We spent $250,000 on Replit. And so it's happening all over the place. And I think really the outsourcing industry is sort of dead at this
Paul Graham
point. So it's, and it's, is it mostly not very technical companies that you're talking to? Like companies that don't specialize in technology. They wouldn't call themselves tech companies.
Amjad Masad
We have a lot of tech companies as customers. I think that that's not where we have the biggest impact because I think they can fish for their own, but the sort of the banks, the, you know, the manufacturing or ordering companies, they need something like Replit. Otherwise, they'll shoot themselves…
Paul Graham
And really, it's driven mostly by people internally who are sort of ambitious. Like some guy who's in charge of sales and wants to write some tool for his sales department.
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.
Masad and Alex Atallah were talking about why enterprises may come to prefer small, specialized models whose outputs can be tightly controlled. Atallah had just said that structured-output models leave much less room for misbehavior. Masad's reply is a programmer's nostalgia for software that did only what it was told.
a16z Podcast · 2026-10-03
Permalink →
Listen →
a16z Podcast
Around 37:01 into the episode
Amjad Masad
Would they be willing to pay 10X to get that much?
Alex Atallah
…and only machines are like dealing with the outputs. And it's not writing code that can execute. Those tasks feel like probably underrepresented in the ones that people talk about and in the things that enterprises are dealing with. So I expect enterprises to get a lot more interested in them.
Amjad Masad
Yeah, I feel like we're going to slowly realize how good we've had it with deterministic code. Like, oh my God, remember the days when computers did exactly what we told them to do. And I think things like Jev, I think, hint at more of a need for not only specialized models, but models whose output domain is more controllable. And maybe you could do a lot more than we thought you'd need by using a bunch of specialized models, specialized output models.
Alex Atallah
Have you guys done any work? Workloads internally with it.
Amjad Masad
You know, I've been training a lot of small models. I mean, I said this glib thing when it first came out because I was like, sort of, I gave this Hacker News comment-like comment, which I felt disgusted with myself afterwards. But I've been taking a lot of like Quinn 8B and like asking, honestly…
Alex Atallah
Yeah. Like I definitely, it also feels like less model debt. Like something that I still hear from companies is that they're like worried about fine-tuning models for like unstructured outputs because you're just like always, you got to redo it again in like two months and everybody just feels the…
Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.