Free with a sign-in.
The best 14 min of 44
How To Build The Future: Paul Graham & Amjad Masad
Amjad Masad says Replit’s years of infrastructure work positioned it to turn AI coding agents into a path toward self-driving companies.
LLightcone · Paul Graham
- Amjad Masad is co-founder and CEO of Replit.
The best minutes
- 22:32Replit is replacing some $10 million software projects for $250,0006 min
- 1:53Replit automated the copy-paste loop between ChatGPT and codePro pick 4 min
- 14:44Replit’s infrastructure gave its agent a head start4 min
Also worth watching
- 32:50Masad expects agent use to move work into the cloud3 min
- 40:13Masad uses exponential compute trends to predict what comes next4 min
- 6:18Replit's original deck proposed learning from coding sessions3 min
- 20:03Early users paid $1,000 instead of $100,000 for an app3 min
- 30:21Replit advances its strategy through customer-led increments3 min
Skip the discussion of engineers adjusting to AI and Replit's early user-research culture, plus the later tour of enterprise rollout and COBOL migration.
Full analysis · 5 min read
Real Alpha · Nov 7
Amjad Masad says firms replaced $10 million projects for $250,000
- Masad says some companies scoped internal systems at $10 million, then spent $250,000 on Replit.
- The anecdote has no project count, matching scope, maintenance costs, or outcome data.
- Treat it as demand evidence, then test total cost and controls before budgeting replacements.
Free with an account
Read the rest of the analysis
6 sections, the 4 other voices it tests Masad against, and what it means for you. Every episode we cut gets one.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.
User observation exposed the product’s real job
Replit’s agent began with a broken user workflow, not a new model breakthrough. Masad says he watched non-programmers shuttle code between ChatGPT and Replit.
For example, they pasted generated code into Replit, copied errors back, then pasted ChatGPT’s fix. The opportunity was automating the entire loop.
Handling errors and looping until a program worked mattered more than sending code once. Graham identified that terminal condition, and Masad agreed.
Masad says autocomplete could produce like 50 or 60 percent of the code. Full generation extended an existing workflow rather than appearing from nowhere.
Our read
The product insight was to remove coordination work around the model. That made task completion, not text generation, the product test.
Infrastructure helped agents finish and deploy work
Completion also needed a place where code could run. Masad says Replit could provision databases, Redis, hosting, and deployment through tools.
In 2024, he said its agent harness was 10 times more complex than today’s because models were weaker. He also said stronger models make agent engineering easier, while lowering barriers for competitors.
Replit’s collaborative IDE let an agent join as another project user. That earlier system work became an environment for agents, not just legacy engineering.
Guillermo Rauch, Vercel’s founder and CEO, said on Sourcery in October that agent-written software still needs infrastructure to run. Danila Shtan, Nebius’s CTO, told Unsupervised Learning in November that agent reliability may matter more than model size.
Those comments offer general industry context, not evidence that Replit outperforms rivals.
Guillermo Rauch · Sourcery · Oct 16AI agents drove 1.5 billion deployments in ten months- DSDanila Shtan · Unsupervised Learning · Nov 2Agent reliability may matter more than model size
Our read
An agent is useful only when its output executes safely and reliably. Replit’s accumulated systems could help, but buyers still need outcome measures.
Enterprise controls add friction and need proof
Masad says enterprise customers value security and reliability, even when guardrails make Replit slower or more expensive. That is a product trade, not a free claim of enterprise readiness.
He points to kernel operating system engineers and a transactional, reversible file system. Those details explain what Replit built, but do not quantify protection.
But this account comes from Replit’s CEO selling Replit’s platform. The episode gives no uptime, incident, audit, or comparative security data.
Jake Stauch, Serval’s CEO, told No Priors in October that his company sends strategists and engineers on site. That is a different enterprise route, with services alongside platform controls.
Our read
The control argument is credible as a product design choice, not yet independently demonstrated as a win. Enterprise teams should separate safeguards from measurable reliability.
Cost anecdotes show demand, not replacement equivalence
Masad recounts companies that scoped internal systems at $10 million, then spent $250,000 on Replit. He says adoption often starts with ambitious employees building internal tools at banks and manufacturers.
On those figures, the scope-to-spend ratio is 40 to 1, our calculation. But that comparison does not establish equal scope, finished functionality, maintenance, or risk.
Earlier customers, he said, paid $1,000 for an app rather than outsourcing it for $100,000. That cheaper alternative helps explain why an imperfect beta could still be useful.
Masad said perhaps 30 percent of the time it did not work. That shows tolerance for a cheaper option, not Replit’s current reliability.
These stories show demand for lower-cost software creation, not a verified cost curve. The episode gives no project count or follow-up outcomes.
Our read
The large numbers make the market legible, but leave the denominator hidden. Buyers need comparable delivery evidence, not just lower spend.
Flat token budgets hide uneven employee returns
Masad objects to equal token budgets because returns vary sharply across employees. He says one person could drive 100 times another’s ROI.
He then says agents widen the gap from 10 times to 100 times, even 1,000 times. Neither figure comes with a disclosed sample or method.
That makes the numbers signals of uneven returns, not sound budget multipliers. Rahul Sengottuvelu, Ramp’s CTO, told TBPN in October that Ramp is not incentivized to drive token costs.
His example is tracking coding-agent use by model and department. That complements Replit’s focus on finding valuable work, though the examples concern different governance tools.
For example, Masad describes a sales employee building a tool for the sales department. Compare token costs with hours saved, work delivered, and support required.
Then controls can follow value instead of imposing identical allowances.
Our read
The useful procurement question is not how many tokens each person gets. It is whether spend can be tied to verified business outcomes.
Customer increments discipline the self-driving vision
Masad calls Replit’s larger aim a platform for self-driving companies, not just vibe coding. The episode offers a more testable route than the slogan, building from customer attempts.
Masad says increments should come from individual customer needs, while the company keeps a broad destination. He favors that approach over announcing a sweeping transformation plan.
His history provides context, but does not certify the forecast. Masad says Replit’s 2015 or 2016 deck proposed learning from millions of coding sessions, before the Transformer paper.
He links that direction to wanting, in 2008, to program on his phone. Those details show sustained attention to access, not a record of every prediction succeeding.
Masad says compute growth led him to extrapolate from GPT-2 in 2019 toward agents and assistants. He predicts use could grow from one agent to 10, then 100, then 10,000 at some point.
He expects work eventually to move off local machines, since a MacBook cannot run 10,000 agents. That is a forecast, not a dated adoption measure.
Still, customer-led pilots can test direction before teams budget for the destination.
Our read
A long-range vision can guide product choices without serving as a budget forecast. Customer evidence is the discipline that keeps the two separate.
What it means for you
- Match scope first. For each proposed replacement, list the old system’s scope, integrations, maintenance, security controls, and delivery time. Compare them with the new result before repeating the 40-to-1 headline.
- Measure workflow ROI. Track tool spend beside work completed and time saved for each use case. Avoid assuming equal token budgets create equal value, the distortion Masad describes.
- Track agent reliability. Measure completion rates, error loops, rework, and security failures alongside speed. Replit’s pitch rests on running tasks through deployment, not simply generating code.
- Watch cloud workloads. Count how many agents teams run concurrently and what still needs a local machine. Masad’s move from one agent to 10,000 is a forecast to test, not a capacity plan.
What to watch
- Whether Masad’s $10 million scope and $250,000 spend comparison recurs with completion and maintenance costs disclosed.
- Whether enterprises expand from chatbot access to agents that execute internal software projects under IT controls.
- Whether agent workloads move into the cloud as concurrent use grows.
Send this to the COO and finance lead reviewing internal software budgets.
Watch the minute where Masad gives the $10 million and $250,000 comparison.
Written by Real Alpha from the episode's transcript and the other moments it links, then checked against the transcript. Quotes are the speaker's exact words.
Minute by minute
A Pro write-up, open to you here. Try Pro free for 7 days
A Pro write-up. The minute by minute is in Real Alpha Pro; the episode, its moments and the transcript stay free.
Read it with Pro, 7 days free- 0:00Agents must work to completion
Masad says Replit saw users copy code from ChatGPT into its editor, then paste errors back into ChatGPT to get fixes. Replit automated that loop, made the agent continue until the program worked, and let it provision infrastructure such as databases. He says model companies later asked how Replit had made it work.
- 5:15Replit pursued accessible programming early
Masad says Replit’s original 2015 or 2016 pitch deck anticipated using machine learning to predict code, before the Transformer paper. He traces the idea to wanting to program on his phone around 2008. Graham describes feeling that AI-generated code threatened the craft he had spent decades practicing.
- 10:13Watching users shaped the product
Masad says users had already experienced AI completing coding tasks, while OpenClaw helped others understand the appeal of agents and the importance of their form factor. He credits Haya with pushing Replit to watch users closely and build a culture around their needs. He says Replit’s distinctive design also helped attract teenage hackers.
- 15:01Replit's infrastructure enabled agents
Masad says Replit’s infrastructure and guardrails help it win enterprise customers, who care about reliability and security. He points to kernel operating system engineers and a transactional, reversible file system as examples of work beneath the simple interface. Because Replit already supported collaboration, the agent could join a project as another user.
- 20:03Users accepted imperfect agents
Masad says the agent failed roughly 30% of the time at launch, but users valued being able to build software even when they could not code. He says early customers could pay Replit $1,000 instead of $100,000 for an outsourced app. He also describes enterprise projects scoped at $10 million that customers completed for $250,000 using Replit.
- 24:54Enterprise buyers need outcome controls
Masad says ambitious employees often bring Replit into banks, manufacturers, and other large organizations, while IT and security teams need administrative controls. He disagrees with giving everyone a fixed monthly token allowance, arguing that returns vary widely and one employee may drive 100 times the impact of another. He says Replit needs to help companies identify where the strongest returns are coming from.
- 28:01Replit is targeting self-driving companies
Masad says Replit’s goal has expanded beyond vibe coding to a platform for “self-driving companies,” where software creation changes how firms work. He calls Replit its own test case for learning how to identify valuable uses of AI. He says companies must keep reinventing themselves as AI commoditizes what they build, and that Replit plans to move toward its vision through customer-led steps.
- 34:04AI adoption still trails Silicon Valley
Masad predicts companies will move from one agent to many, eventually making local computers inadequate for the work. He says a telecom hackathon survey found 80% of participants were anxious about AI before using Replit for an hour, while 80% were optimistic afterward. He argues that many organizations have not yet experienced the agent tools already in use in Silicon Valley.
- 38:12Exponential thinking informs forecasts
Masad says many companies still rely on older software and that some are beginning to rewrite COBOL systems with AI. He traces his forecasts to studying compute growth and says he could imagine chatbots, agents, and personal assistants from seeing GPT-2 in 2019. He argues that people often assume the world is static, while he expects change and builds toward it.
Published Nov 7, 2026.
Transcript · 34 min read
Lightcone, hosted by Paul Graham, with Amjad Masad.
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
I think with AI, everyone will have to go through this period of denial and then depression and then eventually acceptance. Replit's like this iceberg where there's this visible bit above the water that's like, look, you just tell it in English what to do. And it writes your code and then underneath there's all this stuff for deployment in sandbox. I don't think we describe ourselves as vibe coding now. We think Replit is the platform for the transformation for it to self-driving companies. More than other companies though, So I think you guys have been living in the future. If you had to say, like, what's the real DNA of Replit?
It is living in the future, for sure. So you guys basically invented Vibe Coding, right? And when you did it, like in 2024, this was, which it sounds, it's like only two years ago but it seems it's so long ago seems like technologically how did you even know this was going to work because it's obvious now that it did work right but at the time it seemed crazy it seemed crazy to me well for us it was incremental it wasn't like for the rest of the world i felt like a big departure for how from how people did things but we've always you were already writing more and more of the code well no it was it was sort of like the YC kind of ethos of looking at what your users are doing and and and just that will give you ideas on what to build because you were already doing autocomplete we're doing autocomplete so how much of the code were you writing before you went to like full vibe coding where you tell the thing in English what to do yeah where you're just generating code with autocomplete how much of the code were you already generating then I mean you could you could probably say like 50 60 percent really so that's interesting yeah even before vibe coding you were already 60 percent vibe coding yeah yeah and that was how you knew it would work but the user thought we
can make the 60 why not make the rest well kind of i think that the thing that was um very
Read the whole transcript
“Replit would produce some error. They would copy the error, put it into ChatGPT, and then copy the fix.”
I didn't like it at all. How do you feel about it right now? Because aesthetically, it was like very offensive for you. Yeah, I know. I know. Because the way I program is the exact opposite of this. I like, I program, you know, I spend three years writing a thousand lines of code. And if I can get it down to 980 in the next month, so much the better. But you give me the principle for all of this, which is when you first invest in a replo, when you first backed Replit, you said there's a super linear relationship between how easy it is to program and how many programmers you get.
Yes, that's not just true for programming. That's true for anything consumers use. If you make it 10% easier, you don't just get 10% more people. You get twice as many. Yes. Yeah, I remember that. I didn't mean make it this easy. It's like you went too far. I mean, just talk in English and it spits out all this code. Well, so the realization I had is for Repl.it to be truly successful, we needed to go the whole way. Because you did know,
“if you build a platform where there's like millions of people coding, you can like train a machine learning model that predicts.”
When you called me a Bond villain. well you seem like a bond villain you know um that's what i felt like that i was being shown this frightening thing because i remember i still remember what you said don't look at that don't look at the code that's object code yeah right and i'm like oh no the stuff i the whole stuff that i make the whole stuff i've spent decades making is just object code and now everyone's realizing that uh like i think uh yeah i've been seeing like like engineers that were also repulsed by that. And remember, your tweet was very controversial when you tweeted that.
That guy who generated vast amounts of code using AI. Yeah. Well, even that tweet about me being the one villain and object code and all 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 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 you felt that too yeah okay so you knew what i was feeling i knew what you were when you showed me that demo yes uh but you had to feel it so that you can let go 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. It's always surprising to me what is the trigger for these kind of things. Because, again, it seemed obvious to me. And... Well, for Replit users, I think they already knew. For Replit users... But the thing that convinced the rest of the world might have been Cloud Code, right? Don't you think? I think OpenClaw. Okay. I think ClawedCode partly, but OpenClaw, I think, was the first time people experienced this idea of LMs running into completion, where you give it a task, and it's able to go all the way until it's- Which you had been doing for a year and a half at that point.
But I think the form factor of OpenClaw as well being you could text it and all of that, I think, I sort of underestimate the power of a form factor, especially for everyday people. Did it influence what you've done with Replit? Did you learn anything from it? Not yet. I've just learned that lesson. Yeah, I think it will influence things in the future, which is like, I just assume that I go out and like put out a vision and like build a product and Like, oh, yeah, that's obvious. But I think in reality, I think you need to tinker a lot with how people experience it. Do you still watch users using Replit and sort of see what they're doing and are surprised by it?
Yes, I actually didn't like to do that. But Haya kind of forced me to do it early on. and i would like be kicking and screaming because i would like feel like intense pain when like users are confused by the software or hit a bug or whatever yeah i would be like hesitant like i would actually be be you know physically it's mortifying it is to watch users use your stuff yeah it really is but you know why it's mortifying because till now you were in denial about the horrors that were being perpetrated daily out there. But our company culture became a lot about being very close to our users, primarily thanks to Haya.
Having that mentality of you just need to constantly talk to them. She also made you very cool looking. That was one of the really unusual things about Replit early on. Amazing, mind-bending graphic design. Right. Like way better than some nerdy program online kind of company really needs to have. And it helped a lot with our initial sort of audience, which are like a lot of them were like teenage hackers. I thought it was really cool. It made Replit different. Yes. You still have Replit still, there's something very distinctive about it. You know, it has a very, Replit, like a lot of very successful companies, it has a very strong sort of character of itself, which sounds meaningless.
You know, you know what I mean though? It's like a very sort of a strong solution. It's not a weak, like in the chemical sense. It's not a weak solution. Replit is a strong solution. Our conceptual understanding of our, what we build in our product is we always think about like engines and like, you know, hard mechanical things. It is not. That's a lot of what that graphic design was about. Yeah. Yeah. It's like, there's something that's sort of like hardcore about it but it's also paradoxical in that you know we we slave away so much to make the ux so simple for users but on the back end just like really ugly engineering is happening systems programming yes like real systems programming right right that's part of your dna that's part of our dna but it was it's been always sort of challenging because a lot of programmers look at Replit and sort of like dismiss it because it looks very simple yeah but then um even people who end up coming to work at Replit was like oh my god I didn't know there's so much stuff behind the scenes yes yeah you you do system administration for people's sins right yeah I mean we do it so
“We have kernel operating system engineers. We have, like we built our own like file system from scratch to be like transactional and fully reversible.”
You mean having it go back and fix mistakes? Yeah. So there was no native edit tool. Now there's the native edit tool every foundation model give you. but back then we had to do our we had to create our own like sort of edit tool when you give it like a chunk of text how where to replace you couldn't rely on line numbers because they would hallucinate line numbers there's all of these different things that you had to make it ridiculously robust in the early days yes the LLMs were so bad yeah they would also be lazy so they would write like a function and be like get lazy in the middle is like to do finish it later which is not executable yeah there was that was it is now that was insane and that was both open ai and anthropic had the same bug um and i think it's because of the chat model i think they always wanted the chat to kind of like halt early like not like produce mountains of text so that was like an artifact of that and so we had to kind of detect that and like loop over so even editing text was a lot of work and now there's like a native edit tool.
Is there anything like an IDE for agents though? So in the initial days, there was a lot of research on- No, I mean now, the way agents work now. No, no, no. Okay. Now the agents just like generate the code and that's- Now you can just give them a shell. Okay. You just give them a shell and they'll just write shell command. They can do everything with like cat and basic shell commands. Hmm. Hmm. They're very good at shout. But you needed to have the IDE early on because the agents were bad. Yes. Right? Yes. And so that was one of the things that enabled you to be so early. Yes. Could you have done it any earlier?
“So now you either, like, pay, like, $100,000 to get an app done for, like, an engineer somewhere, or you pay Replit, like, $1,000.”
Because ultimately people were just asking agents for the entire thing. And not even use, not even go through the teenage kids asking the agents on their behalf. Yeah. Okay.
“it would have been $10 million. We spent $250,000 on Replit.”
Yeah, yeah, I don't think we're, I don't think we describe ourselves as vibe coding. Now, we wrote this blog post called the self driving company. And so the idea behind it is at Replit, we're early adopters of technology broadly, but definitely AI. We're seeing that we're working differently. Like there's fundamental shift in how we work. You mean Replit as a company?
Replit as a company, which is customer zero of Replit. Yeah, yeah, yeah. So maybe you should figure this out yourself first. How do you identify people doing valuable things? Yes. And so Replit, as a customer zero, is the lab. We're trying to figure out that. So the self-driving company is sort of this idea of the future of work, the future of how firms are organized. And we think Replit is the platform for the transformation for self-driving companies. And so that's sort of the expanded vision of where Replit is headed. Because ultimately, the thing that I didn't really appreciate and realize is I've always known that like making software creation more accessible would have transformational impact I didn't really appreciate that it would fundamentally change how people work like because it seemed we were programmers it seems like what programmers write programs right it's not like this the core of the way companies work doesn't consist of programming yes but maybe now it does it does yeah wow you know by the way incidentally this is an instance of what happens with the very best startups um the very best startups you know you come up with something and you think you're coming up with some idea you suggested to them they've already done it right
you've already got this big new transformation plan because you're that's also that's another thing about the best startups the best startups they're already big and then they think of this new thing as if they were a brand new company right and like they get 100x bigger yeah you know Well, I think it's now actually essential to do that because AI, like we talked about earlier, is a massive commoditization force. So I think where it used to be optional to have to reinvent yourself, it's now mandatory almost.
“And I'm just like, oh, this is what we're doing now. It's like, no, let the customers tell the story.”
“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.”
Because that was why you seemed like a Bond villain when you were demonstrating what later came to be called vibe coding. And here I was sitting there looking at it with dismay. And you're like, yes, Mr. Bond. Accept it now. Exactly. This is the future, Mr. Bond. I think you might even have been wearing a black turtleneck. I was. Not turtleneck. It was a black shirt. Okay. Yeah. The black shirt, the bald head, the beard. Meet me sitting there like, oh, no. What's happened to my beloved programming? And you're like, I already knew. I've been thinking about this for years, Mr. Bond. You know, I do think I have.
and ability to like kill the sacred cow uh in a way like i think i've always had this sort of when computers got really good at something i got all right it's done let's like let's move to the to the next thing like i'm not like so tied to to what i do like i'm fine with like throwing away things and like starting over like in my career i've like started over many times and so i remember you were always rewriting things within yeah rewriting things and so but but i think most people they just like they're sort of like rigid about this is my identity this is what i do i mean they should read keep your identity small uh one of your essays but but i you know i i think um i think this moment requires a ton of adaptability a lot of people ask me oh how do you raise children and what do you teach them and it feels like the main thing.
What do you teach them about AI? It's like flexibility and adaptability and like being a lifelong learner. What I tell them, I mean, is that if you have this tidal wave approaching the beach, the safest place to be is surfing on it, right? Not standing on the beach, holding onto a palm tree. It'll just get washed away. And so that's what your users are doing. Your users, like to the extent anybody can see the future and be protected against bad things happening, your users are it because they're all using AI. Like if you want to use AI, Replit is the way to use it, right? And if you're using it as the best way not to be squashed by it.
Right. And this is why we feel our mission is there's like a moral angle to it almost where you want to help these companies transform. Get those people off the beach and get them all onto surfboards. Right, exactly, exactly. We did this hackathon with a very large telecommunication company, and they did this thing where they did a survey before they used Replit, and a survey after they used Replit, and the survey asked them, like, how anxious are you about AI, how worried, and so on. And so pre-using Replit, just for like an hour, like 80% were anxious and were worried about their jobs or whatever.
And after they used Replit, it flipped. Like 80% were optimistic and 20% were something. For one hour. One hour, yeah. And I think people in Silicon Valley do not constantly underestimate how the rest of the world have not experienced what we've experienced just yet. And it's just going to take a long time. Because things are moving so fast, right? Things are moving so fast here. No, but that's what I'm saying. If they're just two years behind, that's forever behind, right? Because two years ago, Vibe Coding didn't exist. Yeah. I mean, most banks are using like ChatGPT, like 2023, basically. Like the thing that is rolled out at most banks, like other than engineering, I think they're like- ChatGPT3 or something like that.
Yeah. It's just an LLM. It's like barely connected to anything. And so it's not that useful. And so like, there's like a real lag of like how, what people, what people use. And, like, Silicon Valley already, like, went to, like, you know, these agents that are doing things for you. But most companies are still kind of trying to figure out what to do with all the sort of legacy software that they have that they want to rewrite and so on. And they have to. They have to. They can't just ignore it. Right? That's the landscape. That's the software landscape within the company. Maybe all this COBOL code will finally get rewritten.
It is starting to get rewritten. Really? Yes. The COBOL is finally going away. COBOL would eat up a lot of tokens. That is not very powerful. Early on in our AI journey, a lot of companies would be like, oh, this like replica thing is cool, but why don't you use it and come rewrite on COBOL? I was like, we're not going to write your code for you. We'll teach you how to use. this thing. But now there's a bunch of AI companies that do that. That will come and rewrite your COBOL. So there is one last question I want to ask you. This whole replete, for its entire history, the reason it has worked is because you're living in the future to such an extent.
You, personally, are living in the future. How do you do that? Why are you living in the future?
“I actively try to predict things and I actively revise my predictions and look at them.”
That's why someone who's like assumes dramatic change. Imagine who would be perfect for this world where everything is changing under our feet every six months.
The person who's assumes dramatic change, right? You're just like perfectly suited to this moment. That's why Replit's doing so well. You
