00:37:35 · Thariq Shihipar
[00:37:35] S0: Yeah. Is harness design... Is tricky, I I think. Like, I I think you could probably imagine... Like like, before I came to Anthropic, I I built a spreadsheet agent that was a cogen harness myself, and I do think that that would outperform Claude Code at the specific spreadsheet, like, work. But would it outperform it in a way that's worth it in the sense that the Claude Code harness is getting more on the exponential and I would need to continue to update this harness myself? I I I don't know. You know? So... Yeah. And so your second question, oh, like like crazy things people have done with Claude Code. Yeah. I mean, I... There was a interesting company I saw recently that does it like material discovery. And so basically, what they do is like, you know, you can write code to simulate the properties of a material object and, kind of like, you know, a couple years ago, there was this like LK 99 superconductor thing, that turned out to be fake. But like, what if you could simulate a bunch of these and then figure out like from the simulations, what should I test in the lab and, you know, go test them? So, I thought that was like one of my favorite like, you know, in in new ways to use Claude Code. Yeah.
00:38:51 · S7
[00:38:51] S7: I wanted to get your opinion on the seven x usage gap. So with research saying that power users are getting up to seven times more advanced reasoning capabilities out of frontier models than the average users who are kinda merely using it as a search engine. As a builder, how are you thinking about designing the Claude Code interface in a way to kind of stop that? And can we have a UX design that stops that capability overhang by teaching the users to do more multistep agentic tasks?
00:39:25 · Thariq Shihipar
[00:39:25] S0: Yeah. I I think there's both there. Like like like, you know, there's definitely like an interest I have in particular. I think it's both we have to improve the u... Agent UX, and I think that teaching people is very valuable, you know. And so there's a lot of work to do in both ways. I I think that like I am... Like, I think there is some fatigue from people in like, oh, like, I need to learn this new thing all the time, you know, and I think that kind of sucks. And I think ideally, should be Claude's job to do things for you and proactively and know exactly the right amount of things. It's a hard problem. Right? Because like let's say that I have a request and Boris has a request for example. Like Boris knows the code base like in in and out. If I'm like, hey, could you make this change? Maybe Claude should be like, okay, well, do you know that this change would have this follow on effect? You know what I mean? And I'd be like, no, I didn't actually. But like Boris might be like, actually, yeah, did. Why are you asking me? You know what I mean? And so I think there's this like... The model is very complicated. It's not just model and harness. It's model harness user and like the world as well. Right? Like what is possible? What code base are you operating in? What does the user know? And how does the user like model like both the world and the harness? And so it it... Yeah. I think this is, one of the biggest problems in AI right now.
00:40:42 · Aditya Agarwal
[00:40:42] S1: Yeah. So we have ten minutes left, and I'm gonna sneak this in because we were chatting about it on text. Because I think a lot of people are familiar with, the the iOS Sora notion. But because we have somebody who is at Sora, it's worth asking. And obviously, that was... I mean, admittedly, last time we were up here talking about it, it was kind of a very clickbaity type phrase. There's a lot that you can build that's still very interesting, and most can be built. But, like, just a couple comments on from the inside, and you don't have to like share specifics. How would you advise a friend, your friends, who are here at SBC, of navigating what is coming with models, how to actually build in a way that takes advantage of models versus getting steamrolled, specific areas where like, you you can be like, I don't know, buddy. Like, not a good one.
00:41:29 · Thariq Shihipar
[00:41:29] S0: To start, I I think AI is one of these things that is like... It... It's unpredictable for all of us. You know what mean? And and at Anthropic, like Dario's talked about how, for example, we have to pre allocate how much compute we need to buy years ahead of time. You know what I mean? And it's very risky. Right? Like, because we need to buy billions and billions of dollars of compute and we need to hope that we will make that revenue. The way I think of it is that this is like kind of an age of exploration. You know, like, think everyone including the frontier models labs are like figuring out what is possible now. And it is like really exciting. There are so many unknown unknowns about, yeah, where is the direction going? How do you hire? How do you build? How do you organize a team? And it's like, I think some people look at you when you're making a startup and be like, this is like, why are you doing this? You know? Like, it's like extremely like, everything's changing all the time. Like, how do you know it's going to work? Like, what do you know? Like, Well, isn't Anthropic conducive? Is an OpenAI gonna do this? So I I think at a high level, I think you just have to be excited about the exploration and uncertainty and like be like, no, I can do it. You know, like, really believe that I can do it and have that sort of like internal confidence I think is the most important thing. Because it's not, I think, the safe thing to do anymore and it was for a long time at Silicon Valley or the high status thing to do. So I think that's one. I think specific areas of like how to think about it. Again, like, dude, I have a I'm
00:42:57 · Aditya Agarwal
[00:42:57] S1: just gonna get steamrolled. Scroll the beans.