00:22:30 · Haya Odeh
[00:22:30] S1: So for most part, people like to think about persona and not to think about persona is very difficult because you need to de attach yourself from these personalities and start to see common traits between all of them and see and notice behaviors. And what makes all of these people across all of their domain and expertise, the behaviors are all the same, which is, I wanna be heard, I wanna do something, I notice a problem, and I want to do something about this problem now. And these people are very stubborn. And because of all of that, too, they want to win, and they want to prove that they're right, they're not wrong. So whether they're designers, they're PMs, they are small, medium business owners, they are an enterprise, they all have these traits and qualities. And it's very similar from the consumer to enterprise.
00:23:32 · Aditya Agarwal
[00:23:32] S2: I'm curious, Amjad, I mean, you more than you and Haya have more than almost anyone kind of a front row window into the evolution of the models and especially kind of the accelerating progress and capability of these models over the last three or four months, right? And I think that you're laying out a version of the world, which I subscribe to as well, which is that these models have seen, particularly kind of Claude, has seen a tremendous kind of rate of increase in terms of capabilities starting, whatever, December, January. And right now in some ways we are in peak love with Claude Code because these models, they feel like you want to be raw, you want to be close to them, and it gives you a sense of like, wow, between Claude Code, Claude Code work and so on, there's a tremendous amount of capability. But pretty soon you're like, well, I kind of don't want five terminal windows running. I want something that's a bunch more managed. I want different UI. I want more capabilities, right? I guess the first question is, what do you think changed in December? I'm curious to get your analysis of what changed in the model capability. And number two, how quickly are you seeing users get, in some ways, frustrated with the minimal UI of these raw models and wanting more?
00:24:47 · Amjad Masad, Aditya Agarwal
[00:24:47] S0: I think there are different phases of model improvements. Say 2021 or like 2020 to 2023, it was a pre training era where you just put in a bunch of crap, really any kind of data, it didn't matter. And you just have as much parameters as possible, and you grow this model and it just keeps getting better. We have the scaling laws than we thought we [00:25:17] S2: could In a simple word, more data, more parameters, everything was better, yeah.
00:25:20 · Amjad Masad
[00:25:20] S0: Yeah, but then we a plateau, and the plateau for a bunch of reasons. One is there wasn't a lot of data left in the world. Another is the training infrastructure got a little too complicated for growing these models a lot more. And so we entered the sort of RL and thinking era. So like, say, 2024, 'twenty five, starting with one and deep seq and all that, you had this moment of time where the training regime changed and we entered this moment of you're doing reinforcement learning and these models are getting more autonomous because not only the raw intelligence is improving, but they're staying coherent longer over time. I think the big unlock that happened just in late December, November, December, I think the coherence just became almost indefinite. Interesting. Like, Agent three was kind of the most autonomous agent in September 2025, and we struggled to make it run for an hour. Right now, if you have enough evaluation in the middle, if you have computer use or unit tests or whatever, really the models can run indefinitely. The other thing is models also gained some self consciousness almost. I think that's too much of a loaded word, but they started training on their own context window. So they actually also got really good at compression and really good at taking context and storing it into files and being able to search across these files and recollect the contacts based on the next tasks that you're going to do. And all these things culminating together created these different models. The other big thing, at least this is the rumor, is that it looks like the big labs have figured out how to scale pre training again. So I think we're gonna enter another phase of scaling where we're gonna see bigger and bigger models. So Opus is like a really big model, and I think they're gonna get even bigger. And I think the reason token prices are not going down is because the models are getting bigger. But it's sort of ebbs and flows and easy to understand what stage we're sort of in. So I expect the raw capabilities to continue improving in the near future. That being said, I still feel that there's a lot of white space between the AGI providers and what the users want. And the kind of thing that is actually fun and exciting to use, that doesn't change. 100%. That's human, and humans don't change all that much. We're visual creatures.
00:28:06 · Amjad Masad, Aditya Agarwal, Ruchi Sanghvi
We always want to have visual feedback as opposed to just pure terminal. [00:28:12] S2: Yeah, no, I strongly agree. And I think that if you take the other side of that, if you're the kind of founder who believes that in an AGI world there will not be that much for humans to do, well then my real question is why are you starting a company, right? [00:28:25] S0: Yeah, go shopping. [00:28:29] S2: Go learn how to certify, I guess. [00:28:33] S3: That's what I'm doing. So you could believe that. Okay. [00:28:36] S2: Perhaps Ruchi should we move on to company building beliefs? [00:28:39] S3: Yeah. I mean, Replit has a pretty fascinating internal company culture. Like, how many of you have heard about it? Come on, guys. Well, this is your opportunity to learn. So one of Replit's core values is seek pain. I cannot imagine putting up on my walls, seek pain. But where does that come from? And do you scare people away?
00:29:06 · Haya Odeh, Amjad Masad
[00:29:06] S1: Yeah. So I'm gonna answer because this is mine, actually. I created it. Seek pain comes from curiosity when something is working or not working, and going deep, deep, and trying to understand what the hell is going on. It's really difficult to de attach yourself from the work and not feel that it's personal. It's not personal. It's not about you. It's just work. And this is very difficult for people to go to users or go to actually people and ask them, Hey, can you use this thing? Or like, Can you let me know what, like, can you tell me how does this feel to you? And they tell you it suck. And like, you don't feel, you know, like you're now demotivated. So it's such a skill to master that skill. So it's just like learning that, right? So, and then like there's the growth into it, and there's like a lot in seek pain, seek pain in hard conversations, seek pain in understanding like what's going on in the product, seek pain is how can I get better in X, Y, Z? So all of that is a principle that I take to heart, and I wanted to bring it to Replit. And every time someone is shipping something and they're celebrating, and it feels like a bubble that they're celebrating something and they're saying, oh, this is working, yay. And I'm like, can you tell me what's not working? [00:30:37] S0: So anytime everyone's celebrating in the company, Haya will be like, why are you happy? What sucks about this? [00:30:43] S1: What sucks? Can someone tell me what suck about it? Because if I know, I would be like more fulfilled, to