#56 - Amjad Masad

0:00I think that humans cannot really express a good future. Like it's hard to find even like the best sci-fi authors, it's hard to find any expression of a good future. Religion tries, right? But even religion, when you like think really what does it mean to be in heaven living in eternity, it also sounds boring.

0:31That is Amjad Massad. He's the co-founder and CEO of Replit, a programming environment for everyone. Whether you've never written a line of code before or have been developing software your whole life, the goal is that Replit can give you whatever level of control you're looking for. Amjad talks a lot about what we should expect of technology. The tough thing to grapple with is that technology seems to, by definition, make us less resilient. That humanity's history with it has been one of creating more and more abstraction. But Amjad thinks control is worth fighting for. And ultimately, what makes technology good is a question of servitude.

1:11Almost a priori, like there's no app that's going to fully satisfy you. Steve Jobs had this insight, the way to make computers more mainstream is like we had to you know introduce this layer this graphical user interface on top of that. And since then computing did go mainstream I think his insight was correct but at the same time actually computers have become less expressive and the main use case for computers are not that different from televisions if think about what people you know predominantly use computers for is to watch things. Replit has found itself riding a wave of AI that's made its mission of bringing programming to everyone closer and closer to reality.

1:54But at the same time, what it means to build software is changing very quickly, and the second-order effects of that raise a lot of questions. In a world where software costs nothing and can be created at the speed of thought, what holds its value? What does the world look like with lots of on-demand disposable solutions? Do companies just do more with less or do they bring more competencies under one umbrella? I think SaaS is definitely in trouble. I do think that there's like still going to be a place for more low-level handcrafted engineering. For example, you don't want to generate JavaScript to run like SpaceX Falcon, right?

2:37And so I think there's a place for that, but I think that that the generalist is generating software on the fly, maybe some of that software sticks around, but a lot of the time it's like one use, discarded sort of software. Amjad also points out that more broadly, these new superpowers have a significant effect on not just how we build companies and how many people are needed to build them or how specific their job functions are, but at a much, much higher level, how we organize ourselves across society. There's like one primitive or one model that civilization organized around at any given era.

3:18But I think the 20th century and maybe perhaps the 19th century onwards, the factory model created specialization and the way companies work, product management, design, engineering, release, customer success, sales, what have you, right? And I think that model is going to start to break down with AI because a single salesperson could be the SDR, could be the account executive, could be the customer success because they're, again, massively leveraged by these tools. We spoke about science fiction books, expressiveness in computing, the shape and future of software, societal primitives, the nation state and stagnation, resilience, the generalist employee, what becomes valuable in a post-AI world, and the Unabomber.

4:08This is the La Bossier podcast.

4:19We deliver right from the desktop. We're going to reinvent the phone. The world takes it again.

4:31Thank you for having me. Of course. It's really, really good to be here. My pleasure. Okay, I wanted to start by sort of framing the rest of the conversation that we're going to have by starting at the end, which is what a world looks like post-AI or post-AGI. You've said that The City and the Stars is one of your favorite sci-fi books. It describes this future where AI automates all labor, leaving humans to live forever and create art. How much of that book's vision do you think we're actually heading toward? And if that happens, what's left for humans to do? What's interesting in that book is that basically all humans are recycled.

5:09Right. Like there's no new humans, at least in that in that city, in that fully automated, you know, communism city that is like run by like a large computer. And basically, once you've like lived for a thousand years, like the premise in the book is that you just like have a lot of memories. and you have like scars and whatever. So what you do is like you go into like sort of this machine or this experience where you edit your memory and then restart your life. And then when you restart your life, you go through like almost like a school where you go through a simulation to kind of get you started and then you leave that and you go back to your life and you start over.

5:59and then somehow a new person came out of that. They don't actually say how that happened, almost like some kind of glitch, but this person came out with totally fresh memory. They didn't have any experience of what has happened before. It didn't look like they actually lived before. And that person being new thought that the life they're living is like bland and sort of lacking of adventure, danger. You know, he was just like yearning for something else. and then you know without sort of you know going to spoilers he would be the one to actually like get out of the city and try to explore the stars and and there's like all these religious undertones to this to this book I think Ian Banks wrote it but but you know that I forgot who but like I read a bunch of his books and they're all kind of the same way and what's sort of interesting about this book is it basically says this like utopian vision is actually quite dystopian in a way because one of the things that happened when he exits he actually finds people out and they're like exactly the opposite of the city and the stars They are living in this like steady state sort of almost villager life and they had children.

7:45And he was like, for the first time I heard the sound of children. And that was like an extremely joyful, interesting event. And so the vision that Silicon Valley typically presents of AGI is one of this utopia. People talk about UBI and things being automated and things like that. And it is not clear that this is actually utopian. And perhaps, you know, perhaps, you know, there's some like horseshoe type theory where like utopia is dystopia. Totally. Generally speaking. It's a fine line. Yeah. And actually, I think that humans cannot really express a good future. Like it's hard to find even like the best sci-fi authors.

8:45It's hard to find any expression of a good future. Religion tries, right? But even religion, when you like think really, try to think about what does it mean to be in heaven living in eternity, it also sounds boring. Yeah. And so the question is like what happens post-AGI? Well, I mean, yeah, again, like it's hard to express a sort of a good future post-scarcity. My view, however, is that it's unclear whether a true, you know, sense of AGI will be achieved and perhaps ever be achieved. And this is like a metaphysical view. I think functional AGI is actually quite close. And what I mean by functional AGI is that you are able to automate a lot of current jobs.

9:52Anything you can collect data on and maybe generalize a little bit over that data, you know, AIs will be able to do these jobs, especially now that we know how to, you know, start from a foundation model and then like be able to scale this like test time compute where even if the foundation model is not, you know, able to produce a strong answer from the start, you can get to a strong answer. So that's like a huge insight and I think it's a huge unlock for AGI. And again, when I say AGI, I think of it as a functional AGI. I think true AGI, you would need to have a theory of consciousness and what does it mean to be human.

10:48And there's a question for me, and perhaps it's not a very scientific question, is that it doesn't feel like humans are this fully closed system. Whereas when you think about foundation models, the current iteration of AI, they're like fully closed systems. Like whatever you train them on is whatever they're capable of. Even the most recent reasoning models are still working within the distribution of the pre-training, whereas it doesn't appear to me to be true about humans. Interesting. Okay, so a lot of question marks in that end state. Let's talk about present day. I mean, let's talk about Replit, what you do on a day-to-day basis, and then we can kind of get into the steps to getting to that end state.

11:40You said that Replit is not just an IDE. It's not strictly a development environment, even though that's how I discovered it. that's why I started using it as a CS student. It seems like there's a lot more than meets the eye right now and sort of vision-wise. How do you describe Replit today? How do you describe what you guys do? Yeah, I've sort of like been perpetually dissatisfied with the state of the product and maybe perhaps the perception of the product. So what we set out to do is like, it's like a very, you know, like in the seed deck was like computing for everyone or like the like series a deck like making programming more accessible and now we our mission is like to to empower a billion people it's all like different iterations on the same idea of like and by the way the idea is like has been in the air in silicon valley for a long time which is like how do you make it so that um we're actually using um computers to uh you know uh to their ultimate potential and I think we have as an industry kind of failed to do that.

12:46If you look back to the early computing pioneers, they thought that everyone would learn programming and this is how you would use a machine and you know and programming is the best way to use these machines, right? Because otherwise you're always working within the limits of whatever the app developer has created, right? This idea of Turing completeness. Turing completeness mean it's like a general purpose machine. Any app is not Turing complete, any app you sort of click on is not general. And therefore, it will always not satisfy you, right? Almost a priority, like there's no app that's going to fully satisfy you unless it's like a programming app like Replit because it can make whatever you want.

13:48And so, you know, you had Steve Jobs had this insight that the way to make programming, the way to make computers more mainstream is like we had to introduce this layer, this graphical user interface on top of that. And since then, computing did go mainstream. I think his insight was correct. But at the same time, actually computers have become less expressive and the main use case for computers are not that different from televisions. If you think about what people predominantly use computers for is to watch things. Which just like doesn't feel like this is what you need computers for, right?

14:31Like TikTok and Instagram and YouTube. So, you know, an IDE was sort of like one way to approach this problem, but we try to make it a lot simpler. We take care of the setting of the development environment. You just go to a URL in your browser and you start coding. We take care of package installation. We take care of deployment and hosting. And you just go down the line of everything you need to do as a program, and we sort of take care of it. it still wasn't satisfying because it felt like we're still asking too much of people. And so it was popular with students and teachers, which is really great.

15:35And I think we had a big impact there. But again, it wasn't like breaking into the general population, which is what I wanted to do. And, you know, ultimately, like the ultimate bottleneck is the code. You can solve all this problem, but the code is the thing that is going to always be a bottleneck. For example, we introduced this like 100 days of code sort of program. And it's sort of so absurd looking back at it because what you're asking people is to invest 100 days to use this app. There's no other application in the world where you tell people, okay, you have to go through 100 days. 100 days, yeah.

16:12Imagine Facebook like makes you go through 100 days of friending people. And so, you know, and so like with AI, it's like it felt like it was the key answer towards making programming more accessible. Programming languages were not going to do that. Programming interface were not going to do that. ultimately it was going to be AI and the ability to generate code or synthesize programs. And so I think we're now approaching what I think is the ultimate expression of what Replit was supposed to be and it is like a programming environment for everyone. How do you reconcile those last two like really big pieces you talked about?

17:04Because like to your point there's this huge wave that you guys have sort of found yourself riding in this post-LLM era, but what it means to create software has changed fundamentally. So on one hand, you want to give people the tools and the, I guess, granularity of expression in terms of like having control over what that end product is. But at the same time, to the Steve Jobs point, you don't want to abstract them away so much that they don't really have any control over what they're creating at all. How do you reconcile those two opposites in a product. You know what I mean? Yeah, so like one way, one thing I learned from my time at Facebook working on React and associated sort of product libraries

17:53was, you know, this idea of escape hatches. And so with React, we wanted everyone to like follow a certain sort of, we called them the pit of success. Like you would literally fall into success. And whenever you sort of run into a problem, this like closed box that we talked about, you need an escape hatch. And so React had like a bunch of escape hatches to kind of peer under the hood. For example, getting the DOM element was something that React didn't want you to do. But actually, there's a way to do it. Because sometimes you need to see underneath the hood. And so Replit is sort of a layer cake.

18:44So if you, first of all, like any code you make on Replit, you can download it and run it on your machine. That just works. So you can, like, right now, you're talking to an AI. You're seeing the app. You can click on the top left, you open the file system, you can click on the files, you can look at the files, right? So that's the first layer. The second layer is the package management and associated libraries. You can go look into your Node modules or Python modules and spec them, do whatever you want with them. The lower layer is the runtime. What is it actually running on? And so we run on Linux and we run on Nix.

19:26is the underlying sort of operating system. And so you can go peek under that and look at the config files, because we put everything in your repository. And so all the way down, you sort of can understand what's going on. And this is how we're different than any other code generation tool, is by, yes, you have this higher level interface, but you can just keep drilling down. Give people control at every layer of that cake. Okay, let's shift gears for a second and talk about, I guess, the future of software. Tech has created winners of a size that a lot of people thought kind of impossible just a couple decades ago.

20:05I mean, Amazon, Google, Apple, Tesla, the idea of a trillion-dollar company, let alone, I think, 10 of them today, seems impossible not too long ago. Some of it is inflation. Yeah, to be totally fair, yeah. At a company level, do you think AI is more of a sustaining or disruptive technology? Like on net, does it consolidate power further or does it create more opportunity for challengers? Yeah. That's like a question that was pertinent even before AI. It's a question about the internet in general. So think of YouTube. Is YouTube a centralizing or decentralizing technology or disruptive technology, like you say?

20:49Well, in one end, you can argue that it is actually centralizing more power in the hands of Google. It is controlling the content. It is able to do censorship and so on and so forth. You can argue the other side as well. You can say, well, before YouTube, you had to go through the bureaucracy of the television studios and channels, which is almost impossible to get on TV as like a person, as like a kid. And so, well, in that regard, YouTube is extremely radical, extremely disruptive. It disrupted the entertainment industry, it disrupted the television industry, and so on and so forth. And so, yeah, I don't even, like, I think these categories tend to be wrong.

21:47But maybe, like, another way to ask the question about AI is that, like, sort of net-net, is it increasing the power in the sort of hands of individuals versus does it, like, centralize power more in the hands of, like, corporations and governments. So far it's trending towards the former and the reason is trending towards the former is turns out any major advancement in AI is quite easily replicable efficiently, you know, even by individuals and so you know we've seen that with deep seek where you know you know it takes it took the big labs hundreds of millions of dollars not just in training but also in like these ablation experiments and all these things to arrive at gpt4 or clod but it took deep seek you know five six million dollars perhaps if you add all the other stuff like 10 10 20 million dollars to to get to deep Seek v3, which is a state-of-the-art model.

23:06And then on top of that, they caught up to O1. Again, this is all very new, but by just making a few observations about how OpenAI was doing it. More recently, actually just this past weekend, there's a paper that came out that was able to match a one R1 on a 32 billion parameter open source model by dumbest thing. So as the model is generating the chain of thought, whenever it stops, they add wait. They add a token that says wait. And then the model reflects. So it says, oh, I arrived at this answer. Wait. Well, maybe that was wrong. And let me reflect on that. So continuously add weight within a budget, and then they close the, and then they get the answer.

23:59They close the thinking loop. And they match the performance of the top labs. And, like, it cost them$500 in training. And so, you know, today, from my vintage point, it seems like AI is, like, a radically empowering, decentralizing technology. you can run R1 at home for like$2 ,000, again, state-of-the-art model.

24:28But the reason the big labs are trying to make it like a capital game, So, you know, springing up a$500 billion data center, like perhaps the theory is like that's how you create moats. So it becomes like a capital moat. So that's like the big question mark. But like today, my answer is that it is an empowering technology. it. AI is also at a company level or an individual level making it theoretically possible for a single person to build the equivalent of what just a few years ago would have been an entire high-growth tech company. What does this mean for traditional startup teams? I mean, do companies just run with few or no employees?

25:22Do companies bring more competencies under one umbrella? which way does this net out? I do think that you need less employees. And we're seeing that all over the place with companies reaching enormous levels of revenue with relatively few employees. And I think you – it's a combination of – like people think that the reason you need less people is because people are purely more productive. But part of the reason you need less people is that the power of the systems you're building on has increased tremendously. So you have massive leverage. So, you know, think about it as sort of like a platform.

26:18platform, let's say like building on the iFrame platform is like here, building on the Windows platform is here, building on the cloud platform is here, building on the AI platforms all the way to the top, right? Like you're massively leveraged because there's like insane power in the model that you can unlock. And like a lot of companies would write like a few good prompts and scale really fast. Now TBD, how defensible that is. So it's not just that the employees are more productive, which they are, but I think more importantly is that you're unlocking these latent capabilities in these models.

27:05I think the more interesting trend, which I think we're headed towards, which is still not entirely true today, is the generalist employee. And I think the way things are headed is that, you know, we lived in this world where, you know, the company is modeled after like factories, right? Like basically, I wrote this essay called Civilizational Primitives. And the idea is that, you know, there's like one primitive or one model that civilization organized around at any given era. So, you know, they had the agricultural era, which is like everything was a pyramid. You know, they even built pyramids.

27:49But like, they're like, oh, you had the king and you had like this thing and like they created the hierarchy. And so that was the thing, like, you know, everything should resemble this kind of hierarchy, which we still inherited, like it doesn't go away. But I think the 20th century and maybe perhaps the 19th century onwards, the factory model where you have the way the production works is you have one team or department or person work on one thing and then it goes to the other team and it goes to the other theme and at the end you get a final product, which is a huge innovation, right? create specialization and you can think of the supply chain as an extension of that.

28:35You can think of education as like a factory model, like first grade, second grade, third grade, high school, college. You can think of the way companies work. Of course, there's a hierarchy in company, but like there's these departments like the sort of product management, design, engineering, release, customer success, sales, what have you. Right? So, and I think that model is going to start to break down with AI because like a single salesperson could be the SDR, could be the account executive, could be the customer success because they're, again, massively leveraged by these tools. And so companies will start to gravitate towards hiring more journalists.

29:20I think tools like Replit will allow these journalists to be also software developers. So they'll be able to build those tools that's going to help them succeed. So what does that world look like? I mean, you were mentioning YouTube earlier. I draw this really interesting comparison that software until recently sort of mirrored pre-internet media in the sense that it was really expensive to create. it was pretty expensive to distribute. If AI drives the cost of software creation to near zero, as we saw with media, like what happens, what becomes defensible? And this is kind of investor lingo, but like, where does value actually accrue?

29:56Does that make sense? Yeah, good question. I don't know is the short answer. I can sort of like speculate. So like, if we take the media analogy, how do you think, like, what happened to media? Where did the value accrue? I mean, value accrued mostly to the long tail, right? Like your Mr. Beasts of the world ended up soaking up most of the value that was created. And I don't think that means that accessibility is for nothing, right? Like the, I would say base waterline of quality goes up. But you still see these super long tail effects playing out. I think what I wonder about in the context of software is what the shape of software kind of turns into.

30:47So not just the value question, but if you do want to hold value if a switching cost goes to zero from one thing to another, because I can just spin something up at one prompt, do you start to see huge suites of software? Does everything look like Salesforce that has any value or is it just like a ton of point solutions on demand? Yeah. You know? I mean, if you imagine this machine that you can program entirely in natural language, that's obviously AGI, right? So I think my definition of AGI, for any given problem, there's a program that solves it.

31:34and you could imagine you have infinite time and infinite compute and you can generate random strings of JavaScript and you get a problem and you just do that forever and eventually evaluate a program that actually solves the problem, right? So you can build a GI with a wild loop and a random number generator.