Now Anyone Can Code: How AI Agents Can Build Your Whole App

1. 0:00 – 1:15

   Intro

   1. AMAmjad Masad0:00

      (instrumental music) 1984, the Mac brought personal computing to, to the masses. 2024, we have personal software.

   2. GTGarry Tan0:07

      You actually are going to be able to orchestrate this giant army of agents, and I think of Mickey Mouse in Fantasia, just like (laughs) ...

   3. AMAmjad Masad0:16

      (laughs)

   4. GTGarry Tan0:16

      ... you know, like learning this new magical sort of ability, and suddenly all the brooms are walking and talking and dancing, and it's this incredible menagerie of being able to build whatever the heck you want, whenever you want.

   5. AMAmjad Masad0:28

      Someone who had an idea for 15 years but didn't have the tools to build it, and was able to build it in 15 minutes. And he recorded his reaction. I almost shed a tear on that. (instrumental music)

   6. GTGarry Tan0:46

      Welcome back to another episode of The Light Cone. I'm Gary. This is Jared, Harj, and Diana. And collectively, we funded companies worth hundreds of billions of dollars right at the beginning, just a few people, uh, with an idea. And today, we have one of our best alumni to show off what he just launched, Replit Agent. Amjad, thanks so much for joining us today.

   7. AMAmjad Masad1:13

      My pleasure. Thank you for having me. Yeah,

2. 1:15 – 6:19

   Making an app with Replit

   1. AMAmjad Masad1:15

      so we just launched this product. It is in early access, meaning it's barely beta software, uh, but people got really excited about it. Uh, it works some of the time.

   2. GTGarry Tan1:25

      (laughs)

   3. AMAmjad Masad1:25

      So there's a lot of bugs, but we're gonna do a live demo here. And I wanted to, like, build an app, like a personal app, that could track my morning mood correlated with, like, what I've done the, the previous day, so... I want an app, uh, to log my mood in the morning, uh, and also things I've done the previous day, uh, such as the last time I had coffee or if I had alcohol and if I exercised that day. That'll send it to the agent now. We have this, like, chat interface. So you can see the agent just read the, the message and it's now thinking.

   4. GTGarry Tan2:13

      So what we're looking at here is actually how you might chat with another user? Or is this, like, specifically?

   5. AMAmjad Masad2:19

      Yeah, I mean, it's, it's similar.

   6. GTGarry Tan2:21

      Yeah, yeah.

   7. AMAmjad Masad2:21

      It's very similar to, to, like, a multiplayer experience on Replit.

   8. GTGarry Tan2:24

      Got it.

   9. AMAmjad Masad2:25

      Uh, so here, uh, it's saying I created a, um, a plan for you to log your daily mood. The app will show your mood, coffee, alcohol consumption, and exercise. And it also suggests, uh, other features. So for example, uh, it's suggesting, uh, visualization, and that sounds good. Reminders, I don't know, I'll, I'll remember. So let's just go with these two steps.

   10. JFJared Friedman2:49

       I think what was also cool, it picked the tech stack that's very quick to get started. So Flask, Vanilla JS, Postgres, like, very, very good.

   11. AMAmjad Masad2:57

       So now, we're looking at the, what we're calling the progress pane. So the progress pane is, uh, you can see what the AI is doing. Right now it's installing packages and actually wrote a lot of code, and it looks like it built, like, a database connection and all of that, and it's now installing packages, and we should be able to see our results pretty soon.

   12. JFJared Friedman3:15

       This is really cool because I think a lot of times for new software engineers, one of the annoying parts is just getting all the packages and dependencies and picking the right stuff, and this is just, does it for you, the agent.

   13. AMAmjad Masad3:27

       So here we have, we have our Mood App. Uh, I can kind of put that I'm feeling pretty good today. I did have coffee yesterday but I didn't exercise. I'll log my mood, go to history.

   14. JFJared Friedman3:40

       So we've built a complete web app with just a prompt, like, no further instruction from you?

   15. AMAmjad Masad3:45

       Yes, and, and it's, it has a backend, it has Postgres, and I can just deploy this. So this is already pretty useful. You have this rating, and you, and you have the history. Uh, and it's asking me if, if it did the right thing.

   16. GTGarry Tan3:58

       Oh, it actually is asking you to, uh, test it for them.

   17. AMAmjad Masad4:03

       Yeah, it, it actually did te- so- some testing on its own, so it took a screenshot here. And so it, it knows that at least something is presented, but it wants someone to actually go in and do a little bit of QA.

   18. JFJared Friedman4:14

       Is it using, like, computer vision to look at the screenshot?

   19. AMAmjad Masad4:17

       Yeah.

   20. JFJared Friedman4:17

       Okay.

   21. AMAmjad Masad4:17

       Yeah, yeah, yeah. A- and now all the models are multimodal, and so it's fairly straightforward.

   22. GTGarry Tan4:23

       What's on the backend right now?

   23. AMAmjad Masad4:25

       We have a, actually, a few models because you know it's, it's a multi-agent system and we found different models work for different types of agents. The main code gen one is Claude Sonnet 3.5, which is, like, just unbeatable on code gen. It is, like, the best thing. But we use GPT-4O in some, some, some cases. Uh, there's also some, like, uh, in-house, uh, models, like, we built the embedding model. It's a super fast embedding model, bi- binary embedding model. And the retrieval system and in indexing, uh, this is all built in-house. And a big part of what makes this work is, um, is the sort of retrieval system because figuring out what to edit, turns out, is the most important thing for making these agents work.

   24. JFJared Friedman5:11

       You're going a step beyond just RAG because RAG hits, hit the limit for this and you basically have to find a new way to search and find the right places to edit in the code?

   25. AMAmjad Masad5:23

       Yes.

   26. JFJared Friedman5:23

       Which is actually something that I don't think has happened yet, but I think is gonna happen, that for all these multi-agent system, people are gonna move away from RAG and start building custom orchestration-

   27. AMAmjad Masad5:35

       Mm-hmm.

   28. JFJared Friedman5:35

       ... like this. So this is very notable, this is, like, a very cool thing that you figured out.

   29. GTGarry Tan5:39

       Yeah. I- just throwing the code base in RAG is not gonna work. Y- you actually have several different representations that-

   30. AMAmjad Masad5:45

       Exactly.
3. 6:19 – 8:07

   Feel the AGI, personal software era

   1. AMAmjad Masad6:19

      And one of the things I'm really excited about is, like, this idea of personal software. 1984, the Mac brought, like, personal computing to, to the masses. 2024, we have personal software. (laughs)

   2. GTGarry Tan6:30

      I think we just experienced this. You know, Karpathy just tweeted about, uh, Replit Agent, and he said, "This is a feel the AGI moment."

   3. AMAmjad Masad6:38

      Mm-hmm.

   4. GTGarry Tan6:38

      Did you just feel the AGI?

   5. JFJared Friedman6:40

      I definitely did, and I, I did last night, I, I spent a few hours last night using Replit Agent to make a Hacker News clone.

   6. AMAmjad Masad6:46

      Nice.

   7. JFJared Friedman6:47

      There were a couple of moments where, like, I really felt the AGI.

   8. AMAmjad Masad6:50

      Mmm.

   9. JFJared Friedman6:50

      Um, the first was, it actually had, like, really good intuition about what UI to make and how to design it. And it, like, we, we saw that there, where, like, you didn't give it the idea to make the slider bar be like, uh, like, like emojis.

   10. AMAmjad Masad7:04

       Yes.

   11. JFJared Friedman7:04

       It just came up with that on its own. And then the second thing was, when I was using it, it really felt like I had a development partner, where it would ask me questions, it would ask me to, like, change things. At one point, it got, like, stuck or wasn't sure how to do something, and so it asked me how to do the thing-

   12. AMAmjad Masad7:20

       (laughs)

   13. JFJared Friedman7:20

       ... and then I told it, and then it's like, "Cool, got it." (laughs) And just, like, kept going.

   14. AMAmjad Masad7:23

       (laughs) Yeah, yeah, I guess. Yeah, it, it feels great, and, and s- sometimes you wanna give it some, some help, right? You wanna, you wanna go debug, if you know how to debug yourself, or you go ask ChatGPT about something and come back to it. Just give it more information, it'll be able to kind of react to it.

   15. SPSpeaker7:38

       You should have, it definitely feels like talking to, like, a developer.

   16. AMAmjad Masad7:42

       Yeah.

   17. SPSpeaker7:42

       You should do, like, the grok thing and have different modes. You could have, like, grouchy programming-

   18. AMAmjad Masad7:47

       (laughs)

   19. SPSpeaker7:47

       ... where it just tells you, like-

   20. JFJared Friedman7:48

       (laughs)

   21. SPSpeaker7:48

       ... ideas are bad or wants to build something else anyway.

   22. AMAmjad Masad7:51

       Oh, that would be cool.

   23. SPSpeaker7:52

       (laughs)

   24. AMAmjad Masad7:53

       Just, like, have a, like, a toggle, for example, like an over-engineer.

   25. SPSpeaker7:56

       Yeah, yeah. (laughs)

   26. AMAmjad Masad7:57

       Like, just, like, over-engineer everything.

   27. SPSpeaker7:59

       Exactly.

   28. AMAmjad Masad7:59

       So, so it added this, uh, toggle, but I don't think it works.

   29. GTGarry Tan8:03

       I don't think it connected up to the x-axis.

   30. AMAmjad Masad8:05

       Yeah.
4. 8:07 – 9:51

   Having AI code the way humans do

   1. JFJared Friedman8:07

      about all these AI programmers, which is that it's not like we created some super intelligence that somehow can just build an entire app perfectly from start to finish without making any mistakes. It actually codes the way a human does, which is it, like, writes some code, and it's, like, "Well, I think this is right, but I'm not sure. I guess I'll try it." And then it tries, like, "Oh, no, I have a bug." (laughs) It's, like, it's the same thing, yeah.

   2. AMAmjad Masad8:27

      Yeah, and, and we, again, our d- our design decision, uh, has been always, like, this is a, uh, a coworker. And you can just close this, and you can go to the code, and you can code yourself. Uh...

   3. GTGarry Tan8:41

      Just fix it yourself, yeah.

   4. AMAmjad Masad8:42

      (laughs) Fix it yourself. And again, if, if you don't know how to code, I, uh, my hope is as you are reading what the agent is doing is that you've learned a little bit of coding along the way. And by the way, this is how I think our generation learned how to code. Not through agents, but almost by doing these incremental small things, like editing your MySpace page or doing a Geocities, uh, thing. And I feel like we sort of lost that incremental learning, uh, scale, where now you need to go out and get a, like, computer science degree or go to coding bootcamp to kind of figure this out. But if we made this, like, fun thing that people can go build side projects in and get exposed to what code is, I think that would be perfect. And again, my view is that we're still far from fully automated software engineering agents, uh, and people sh- should still learn how to code. You have to do way less coding, but you will be, you, you will have to read the code, you will have to debug it in some cases. The agent will get you fairly far, but sometimes it'll get stuck and you need to go into the code and figure it out.

   5. GTGarry Tan9:50

      Yeah, I think that that's

5. 9:51 – 11:42

   You should still learn to code!

   1. GTGarry Tan9:51

      actually pretty important. I'm s- I've been meeting a lot of, you know, 18, 19-year-olds who are freshmen, and they're like, "Well, the code will write itself, right? Like, I don't have to study this stuff anymore." And I'm like, "No, that's not true at all." Like, I actually think that now it is actually more leverage, it is far-

   2. AMAmjad Masad10:07

      Yes.

   3. GTGarry Tan10:07

      ... more leverage to know how to code than ever before.

   4. AMAmjad Masad10:10

      That's exactly right.

   5. GTGarry Tan10:10

      And it's actually even more important, and it will make you way more powerful. Like, you don't have to be all the way in the weeds on everything. You actually are going to be able to, um, like, orchestrate this giant army of agents. And, uh, I think of Mickey Mouse in Fantasia, just like (laughs) -

   6. AMAmjad Masad10:27

      (laughs)

   7. GTGarry Tan10:27

      ... you know, like, learning this new magical sort of ability and, like-

   8. AMAmjad Masad10:31

      I love that.

   9. GTGarry Tan10:31

      ... you know, uh, suddenly all the brooms are, like, you know, walking and talking and dancing, and there's this incredible menagerie of being able to build whatever the heck you want whenever you want, just, like, like, literally from any computer-

   10. AMAmjad Masad10:43

       Yeah.

   11. GTGarry Tan10:43

       ... from any web browser.

   12. AMAmjad Masad10:44

       Yeah. I, I try to come up with a, like a Moore's Law type, type thing, where it's like, the return on, on learning a code is, like, doubling every six months or something like that. So, learning to code a little bit in, uh, you know, 2020, um, you know, w- was not that useful, because you would still, y- well, you'd get blocked. You wouldn't know how to deploy something, you wouldn't know how to configure something. Let's go to 2023 with ChatGPT. Learning to code just a little bit will get you fairly far, because ChatGPT can help you. And then 2024, learning to code a little bit is a massive leverage, because we have agents like this and others, and there's a lot of really cool tools out there, like Cursor and others, that will get you super far by just, like, having a little bit of coding. And, and just extend that forward, like, six months later, you're gonna have even more power. So, programmers are just on this massive trajectory of increased power.

6. 11:42 – 17:19

   The underlying tech

   1. JFJared Friedman11:42

      Okay, tell us more about the tech behind this. It's kinda fascinating.

   2. AMAmjad Masad11:45

      At, at, at the heart of it, it is sort of this, uh, uh, as I described before, it's multi-agent system. Um, you, you have this core s- sort of React-like loop. So, React is a, a, you know, an agent chain-of-thought-type, uh, prompting that's been around for a couple years now, and most agents are, are built on that. Uh, but ours is also a multi sort of agent system. We give it a ton of tools using tool calling, um, and those tools are...... the same tools, again, that are exposed to people. And by the way, you, you need to be really careful about how to expose these tools and how does the agent see them. Um, so for example, our edit tool returns, uh, errors from the language server. So, we have a language server here, a Python language server. It's like a human coding, you know? If I, if I make a mistake, um, anywhere here, it will show me, right? Similarly, when the agent is coding, it gets feedback from the language server. So again, you wanna treat it as much as you can like a, like a real user. And, and so for, for any action, it gets, it gets sort of a feedback and then it can react to that feedback. And so these are the tools. Again, this is package management, uh, editing, deployment, all... The database, all those are, are tools. Um, and then there are a lot of things that, uh, make sure that it, you know, doesn't go totally off the rails, 'cause it's very easy. You, we've all, you know, used agents that go off the rails-

   3. JFJared Friedman13:18

      Mm-hmm.

   4. AMAmjad Masad13:18

      ... and go into endless loops. This still som- sometimes does it, but we have another loop that is doing a reflection, that's always thinking, "Am I doing the right thing?" We use a lot of, uh, LangChain tools. So, LangGraph is an interesting new tool, uh, from LangChain that allows you to build agent DAGs very nicely, and they have, uh, some logging mechanism. Um, and, uh, a tool called LangSmith where you can look at the traces. Looking at the traces for, for DAGs is, is very, very, very difficult and very hard. So, debugging these things have been fairly difficult 'cause you, you want a tool to actually, like, visualize the graph and there isn't a lot of tools that do that right now. And so, so there's this reflection tool, uh, reflection agent, um, and it... A- and the other thing that we talked about earlier is, uh, retrieval is, is crucial. And again, this, this has to be, um, kind of neuro symbolic. It, it has to be able to do RAG-style embeddings retrieval, but it has to be able to look up functions and symbols inside, inside the code.

   5. JFJared Friedman14:24

      This is why I do think I may be extrapolating a bit more. Even if we get into the world of foundation models that have really, really large context windows, I mean, Gemini already is in the millions of tokens, you will still need very specialized things that do lookups like this because apply to different contexts, knowing the functions and treating it more like how it compiles at the end, like a AST graph makes a difference.

   6. AMAmjad Masad14:48

      Large context windows, uh, you can totally shoot yourself in the foot with them.

   7. JFJared Friedman14:51

      Yes.

   8. AMAmjad Masad14:52

      Because it's easy for the model to... It's actually, you know... The, the model will bias a lot more towards whatever is at the end.

   9. GTGarry Tan14:59

      Hmm, kind of like a human.

   10. AMAmjad Masad15:00

       Yes, exactly. And, and so you still need to do context management. Um, and you need to figure out what to put on, wha- how to rank memories. So, this agent, every time it does a step, uh, it, it goes into a memory bank, and then every time we go into a, uh, the next step, we would be able to pick the right memories and figure out how to put them in context. If you pick the wrong memories, for example, if you pick a memory that, that, you know, had a bug or there was an error in it, whatever, it might still think that there's a bug. But, but if you already recovered from that, you wanna make sure that me- that memory of, of having created a bug, uh, is, is either kind of augmented by another memory of fixing it or entirely removed from the context. And so memory management is, is crucial, uh, here. You, you don't want to put the entire memory in, in context. You wanna be able to pick the right memories for the right tasks.

   11. GTGarry Tan16:00

       I feel like this is a really concrete, um, rebuttal to situational awareness and that whole, like, sort of sci-fi, uh, you know, AGI is gonna kill us tomorrow kind of argument, simply because that all is predicated on larger context window, um, more parameters, throw GPUs at it and it's gonna work. Like, you can't just scale it up. Like, you're not going to get what you want from just scaling it up. There is actually a lot of utility in having these agents work one... With one another, with, uh, being actually smart about, uh, what is the intermediate representation and being able to pull back, you know, sort of model what a human would do.

   12. AMAmjad Masad16:41

       Mm-hmm.

   13. GTGarry Tan16:42

       I mean, thi- this is sort of like the, the case study and like, oh, yeah, you can't just, you know, scale up everything by 50X and have it work the way that, uh, they think it will.

   14. AMAmjad Masad16:53

       Yeah. In many ways, like, building a system like that sort of humbles you. You know, it sets, sets your expectations, uh, ab- about AI and the progress in AI in, in a sort of a different way because, yeah, the systems are very fragile. They're really still not great at following instructions. People talk a lot about the halluci- hallucination problem. I think the bigger problem is, like, just following orders. Uh, it's so hard to get them to actually do the right

7. 17:19 – 19:41

   The path to AGI

   1. AMAmjad Masad17:19

      thing.

   2. JFJared Friedman17:19

      What do you think is the path to AGI?

   3. AMAmjad Masad17:22

      So, so my view on AGI is that maybe we'll get to something called... We can call functional AGI, which is, um, we, uh, automate all those sort of economically useful tasks. I think that's fairly within reach. Uh, I think it's, it's almost like a brute force problem. It's sort of the bitter lesson, right?

   4. JFJared Friedman17:43

      Do you think it involves doing a lot of work like what you guys did? Like, basically building... (clears throat) Like, carefully fine-tuning orchestrations of groups of agents for each task? So, doing what you did for programming, but doing it for customer support and for sales and for every... Accounting every function?

   5. AMAmjad Masad17:57

      Yeah, I, I think so. And maybe you can eventually put it all into one model. The history of, of machine learning has been, um, we create the systems, we grow these systems around these models, and eventually the model will eat th- those systems. So, hopefully, like, everything that we did, at some- someday there's, like, an end-to-end system, uh, machine learning system that could do it. You know, Tesla, you know, famously, you know, had all these logic and, and whatever, and now, like, you know, I think after V13, they, they... It's just end-to-end training. Um, a- a- and so, you know, eventually we'll, we'll get there. Um-But, but I wouldn't consider it true AGI w- because, uh, you throw something out of distribution at it and it wouldn't be able to, uh, to, to handle it. Um, I think true AGI would require efficient learning, being able to be thrown in an environment with no information at all, being able to understand environments by examining it, and learning a skill required to navigate that environment, and LLMs are not that. Maybe they're a component of that, but they're not efficient learners at all.

   6. JFJared Friedman19:09

      You actually demonstrated this because the way you describe LLMs are intuition machines-

   7. AMAmjad Masad19:14

      Mm-hmm.

   8. JFJared Friedman19:14

      ... and in order to get them to work in programming tasks, you had to add this layer with, uh, symbolic representation-

   9. AMAmjad Masad19:21

      Mm-hmm.

   10. JFJared Friedman19:21

       ... like in programming and ASTs. Like, a lot of concepts in programming-

   11. AMAmjad Masad19:25

       Mm-hmm.

   12. JFJared Friedman19:26

       ... and how computation works, like Turing-complete with DAGs and all that, right?

   13. AMAmjad Masad19:29

       Yes, exactly.

   14. JFJared Friedman19:30

       Those are, like, very explicit classical computer science-

   15. AMAmjad Masad19:33

       Classical AIs, yeah. That we do backtracking-

   16. GTGarry Tan19:36

       That's not generalized.

   17. AMAmjad Masad19:36

       ... and all of that, yes.

   18. GTGarry Tan19:37

       That's not generalized, that's specialized. I mean, incredibly-

   19. AMAmjad Masad19:39

       Yeah. (laughs)

   20. GTGarry Tan19:39

       ... useful specialized.

   21. AMAmjad Masad19:40

       Yes.

8. 19:41 – 25:56

   What users made with Replit

   1. JFJared Friedman19:41

      So, it's only been live for four days-

   2. AMAmjad Masad19:44

      Yeah.

   3. JFJared Friedman19:44

      ... but already people have done a bunch of, like, really interesting and impressive stuff with it. Do y- do you want to talk about some of the things that you've seen people do with it-

   4. AMAmjad Masad19:50

      Yeah.

   5. JFJared Friedman19:50

      ... that are most, like, surprising and interesting?

   6. AMAmjad Masad19:53

      Yeah. One of my favorite thing that I saw w- was someone who had an idea for 15 years, but didn't have the tools to build it, and was able to build it in 15 minutes, and he recorded his reaction. And it's like a personal app. He b- r- he built an app where he can put memories on a map and attach files and audio files to it, memories about his life, "I went to school here," and, like, add a picture or whatever. When the app showed up and he tested it and he was like...

   7. JFJared Friedman20:21

      (laughs)

   8. AMAmjad Masad20:21

      He was so surprised. I almost shed a tear on that. I was like, you being able to unlock people's creativity is, is, is so rewarding.

   9. GTGarry Tan20:31

      And then I want an integration with, uh, Apple Photos or (laughs) to use it to actually build a, a, a export tool.

   10. AMAmjad Masad20:38

       Yes. And a- another user, Mackay, built, um, uh, sort of a Stripe coupon tool.

   11. JFJared Friedman20:45

       Hmm.

   12. AMAmjad Masad20:45

       So, he has a course, he runs it on Stripe, and he wants, like, to be able to, like, send people coupons. And so he built it in, like, you know, five, 10 minutes.

   13. JFJared Friedman20:53

       Hmm.

   14. AMAmjad Masad20:53

       And actually, I don't think you would be able to build something like that in no-code. You would struggle really hard. You would probably use two or three no-code tools. People use, like, Bubble on the front end and Zapier on the back end and, and what have you. Y- sometimes I'm surprised, the no-code people are actually quite, quite smart and quite, uh, hardworking because they figure out how to create these systems using no code, but it's just actually a lot easier to just generate the code for it.

   15. GTGarry Tan21:20

       Yep.

   16. JFJared Friedman21:20

       It's a coding tool for the no-codes. (laughs)

   17. AMAmjad Masad21:22

       Yes, yes. And so, yeah, w- we, we're seeing a lot of traction there.

   18. SPSpeaker21:26

       W- which is actually a challenge I think the no-code tools have in general is straddling this line between they start very much no-code and then they find that people keep pushing the limits of what they want to build in these tools and, and-

   19. JFJared Friedman21:36

       And then, and then the, the frustrating part with no-code tools is that if you hit the limits, you're just stuck. Like, you, you just, you can't solve it. And the cool thing is if, as you were saying earlier, if you can get the no-code people to switch to Replit, maybe initially they don't program at all, all they know how to do is, like, prompt it.

   20. SPSpeaker21:51

       Yeah.

   21. JFJared Friedman21:51

       But then at some point they're gonna, like, look at the code and they'll realize that they can just edit it and, like, it isn't that hard. And then that's how they, like, gradually become programmers.

   22. SPSpeaker22:00

       Yeah, that's interesting. I played around with it to build just, like, a simple recruiting CRM, which is actually the kind of thing I- you would have used Airtable for.

   23. AMAmjad Masad22:07

       Mm-hmm.

   24. SPSpeaker22:07

       And one of the suggested... When it told me the plans, one of the, "Oh, would you like this feature?" was exactly that.

   25. AMAmjad Masad22:12

       Mm-hmm.

   26. SPSpeaker22:12

       It was just, like, role-based permissions and off-

   27. AMAmjad Masad22:14

       Yeah.

   28. SPSpeaker22:14

       ... it's like, oh, that's pretty, like, a sophisticated prompt or, like, suggestion off the bat.

   29. GTGarry Tan22:19

       Yeah. Yeah, that's a $10,000 a month enterprise-

   30. SPSpeaker22:21

       (laughs)
9. 25:56 – 33:29

   Challenges in resetting the org

   1. GTGarry Tan25:56

      had a, I guess, sort of mini Chesky moment earlier this year, then. We're all blown away by this demo and sort of, you know, you've been working hard on sort of remaking the way, um, all software is deployed and written for some time. I mean, what, what did it take to, you know, get to this moment? Um, you know, you did have to do a layoff and reset your org, you know? What happened?

   2. AMAmjad Masad26:21

      Yeah, so, so last year, we, we raised a, we raised a big round. Um, we, we felt we were making fast progress, and there, there was a lot of energy. And I, I felt like I needed to, okay, grow the company. You know, for, for a long time, Jared knows, for a long time, Replit was, like, tiny.

   3. JFJared Friedman26:40

      It was actually run out of your apartment.

   4. AMAmjad Masad26:42

      Yes. (laughs)

   5. JFJared Friedman26:42

      For how many years?

   6. AMAmjad Masad26:43

      For many years, sort of, like, three or four years. And we were, like, four or five people for, like, s- many years. So we started growing in 2021.

   7. JFJared Friedman26:51

      Even when you had a lot of users.

   8. AMAmjad Masad26:52

      Yes.

   9. JFJared Friedman26:52

      Like, you were four or five employees when you had, like, millions of users.

   10. AMAmjad Masad26:55

       Yes, that's right. And so we were always kind of lean, but I thought last year, "Okay, we have, we have really big ambitions. We gotta go hire people. I gotta hire executives. I gotta create a, like, a management structure. I gotta, like, grow up." (laughs)

   11. SPSpeaker27:06

       Is that what, uh, investors were telling you, is like, "Oh, you gotta hire people"

   12. AMAmjad Masad27:09

       No, actually. I, I was-

   13. SPSpeaker27:11

       Or were you thought you needed-

   14. AMAmjad Masad27:11

       ... done on my own. (laughs)

   15. JFJared Friedman27:14

       (laughs)

   16. AMAmjad Masad27:15

       Uh ...

   17. SPSpeaker27:16

       Or you thought-

   18. JFJared Friedman27:16

       But, but it definitely was the prevalent advice.

   19. AMAmjad Masad27:18

       Yeah.

   20. JFJared Friedman27:18

       I mean, you were, you were absorbing this advice from sort of like-

   21. AMAmjad Masad27:21

       Yes.

   22. JFJared Friedman27:21

       ... the, the world that was, that ordinarily advises startups to do exactly that.

   23. AMAmjad Masad27:25

       That's right. That's right. And it just got really miserable. Uh, we had, like, you know, multiple layers. We had different meetings where I'm trying to, like, run the company from. We had, like, a executive meeting, staff meeting, whatever. We had roadmaps, we had planning sessions, and I just couldn't shake the feeling that it was all LARPing.

   24. JFJared Friedman27:45

       Hmm.

   25. AMAmjad Masad27:47

       (laughs)

   26. SPSpeaker27:47

       (laughs)

   27. JFJared Friedman27:47

       (laughs)

   28. AMAmjad Masad27:47

       It was not work.

   29. SPSpeaker27:49

       (laughs) .

   30. AMAmjad Masad27:49

       It was LARPing. And but, right now, we don't have a roadmap. Right now, literally, we work on, like, three or four things. I'm involved in all of them, and I know what's going on there. I know what people are working on, and I think we got a lot more productive by getting smaller, by, you know, flattening the organization.
10. 33:29 – 36:12

    Future plans

    1. SPSpeaker33:29

       coming next with the agent? Like, what's... Um, what do you want to add to it? What, what do you think are gonna be the big next leap forwards for it?

    2. AMAmjad Masad33:35

       Reliability. I think the, the most important thing right now is reliability and making sure it's not spinning, making sure it's not breaking, uh, and then, uh, expanding it to support any stack you would want. So right now, we don't really listen to the user (laughs) when they give us a stack.

    3. SPSpeaker33:51

       Ah.

    4. AMAmjad Masad33:52

       We, we push back. The agent pushes back. It's like, "Ah, I'm just gonna do it in Python or whatever." Uh, but if you really want-

    5. SPSpeaker33:57

       Crafty engineer mode.

    6. AMAmjad Masad33:58

       Yeah. (laughs)

    7. SPSpeaker34:01

       (laughs)

    8. AMAmjad Masad34:01

       (laughs) So we want to be able to accept, uh, user requirements with regards to stack.

    9. SPSpeaker34:06

       Should have the Paul Graham mode where I only write it in Lisp.

    10. AMAmjad Masad34:09

        Yes.

    11. JFJared Friedman34:09

        (laughs)

    12. AMAmjad Masad34:09

        Yes. (laughs)

    13. GTGarry Tan34:09

        (laughs)

    14. SPSpeaker34:09

        (laughs) To do it in anything else.

    15. AMAmjad Masad34:10

        You know, this modes thing is a really... Like a April Fool thing.

    16. SPSpeaker34:14

        (laughs)

    17. AMAmjad Masad34:14

        Like, Paul Graham, overengineer.

    18. SPSpeaker34:17

        (laughs)

    19. AMAmjad Masad34:18

        (laughs)

    20. GTGarry Tan34:18

        (laughs)

    21. JFJared Friedman34:19

        (laughs)

    22. SPSpeaker34:20

        Bad UI. Doesn't care about UI.

    23. AMAmjad Masad34:22

        (laughs)

    24. SPSpeaker34:22

        Everything's confu- literally correct, but very c- confusing.

    25. AMAmjad Masad34:26

        (laughs)

    26. SPSpeaker34:27

        How about just the interaction? I mean, you mentioned, like Licklider and the whole-

    27. AMAmjad Masad34:31

        Yeah.

    28. SPSpeaker34:31

        ... human-computer symbiosis theory. Like, is text, like, as far as it goes? Are there other ways that people you think will want to interact with their AI agent?

    29. AMAmjad Masad34:40

        You should be able to, like, draw, uh, in the UI and communicate with the, with the AI by drawing, right? You should be able to say, "Hey, like, this button's not working. Maybe move this here, or this file, you know, is bro- is not, you know, refactor this file," whatever. So, you know, if the whole thing is a canvas that you can draw on, you can communicate a- a lot more expressively with the agent. And of course, you're talking, you know, as opposed to typing, being able to talk and draw, it's... I- imagine you're on the iPad too. We have an iPad app. It could get really, really fun and creative.

    30. JFJared Friedman35:14

        Kind of like a full UI mockup that you would do in Figma, you could kind of hand sketch it and get it to-
11. 36:12 – 37:13

    Outro

    1. GTGarry Tan36:12

       (laughs)

    2. AMAmjad Masad36:12

       If you're, if you're, if you're, uh, brave and you want to test it and give us feedback, uh, go to Replit, sign up for our core plan 'cause this thing is expensive. We can't give it away for free. (laughs) And, uh, and you'll be able to see that module on the homepage that says, "What do you want to build today?" Uh, and then you can go through that and start working with the agent. Just have an idea in your mind. Just write a couple sentences. Do- don't make it too complicated, um, or too technical, and, uh, and get started. You'll get a feel of how to work with the agent pretty quickly. It should be pretty intuitive. And share with us what you're building. Happy to kind of reshare, retweet whatever p- people are buil- building with the agent.

    3. GTGarry Tan36:55

       Amazing. Well, it's time to feel the AGI.

    4. SPSpeaker36:58

       (laughs)

    5. AMAmjad Masad36:58

       (laughs)

    6. GTGarry Tan36:58

       We'll see you guys next week.

    7. SPSpeaker37:00

       (laughs) (instrumental music plays)

Episode duration: 37:13