00:22:04 · Molly O'Shea, Dylan Field
[00:22:04] S0: Really? [00:22:05] S1: I think so. And That's crazy. Yeah. And, like, we're still prompt in a way. There are other ways to explore latent space. Mhmm. I just think that the interfaces for how you interact with models will evolve so much. And to the degree that models are part of software, we will see a ton of exploration there. [00:22:26] S0: When we're talking about brand and and personality and forming trust, AI has really stopped a lot of trust in technology and just kind of like mainstream media. It's like I think it's like literally worse than ICE. So what builds a trusted brand?
00:22:47 · Dylan Field
[00:22:47] S1: Well, AI, first of all, I would say is really bimodal case, right? Because you've got folks that are literally in the case of a 4.0 or Quad, in love with the model. And I don't just mean, like, they really like its outputs. Like, they have affection for the model. I would say that's trust. And then you've got people that are like, this take it could take my job. I'm listening to things people are saying about how it's gonna destroy the world or the economy or whatever and doom. And so that is couldn't be further apart, like the super fans versus the super detractors. But in terms of just generally brand, I think it's consistency over time. And how do you actually show up in a way that people can expect, can model in their heads, and keep living your values, keep showing up the right way, and really doing things with the customer's interest at heart. If you're not doing that, what are you doing? Yeah. So I also go back to a line that Tim O'Reilly would always say. My first internship professionally was, like, at O'Reilly Media, and I loved it there. They were amazing to me, and I'm so grateful to them. Tim O'Reilly would always talk about create more value than you capture. And I think that is extremely important. And when you're perceiving, you know, singletons for AI, folks that, you know, you perceive might want to become monoliths and they're not saying that they're not going to want to do that. There's going to kind of be like a black hole and suck it all up, which I'm sure that that's not going to be what ends up happening. I think that they're all gonna get a lot smarter than that. And I think most don't even wanna do that in the first place. But yeah, the definition of a platform is that you're creating even more value outside of the platform than you are from the platform itself. This is also the Microsoft definition. It's like, what's the GDP of everything created on Windows back when, you know, Windows was the top priority of Microsoft? They wanted that sort of, GDP of stuff built on Windows to be greater than the revenue of their Windows line. And I think that's a really special way to think about ecosystem. And yeah, overall, the more that I think we lean into that kind of mindset as an industry and AI companies are vocal about that and show that, then I think people have more trust.
00:25:34 · Dylan Field, Molly O'Shea
But the doomerism has to stop as well, and it has to stop with good rationale. We have to come up with the best arguments we can for not what can go wrong, but how we're going to go address it and paint an optimistic future vision for humanity. [00:25:52] S0: Right. I've been seeing a shift in sentiment towards productivity, not job loss. And so while, like, I guess it's weird, but kind of the new North Star is to have more agents than employees. Because they'll help equip your employees and you'll run faster, you'll be able to do more as an organization, you can be more on the offensive side, which I think you and Figma have been doing really well. And then, I mean, you could also see this with, like, Figure AI. Like, they now have more humanoid robots at their office than employees. Obviously, they're gonna sell them, but still. [00:26:25] S1: Yeah. It's a little different. [00:26:26] S0: It's a little bit different. But I'm curious your standpoint as the proliferation of AI agents increase, like how does that change the internal dynamics of software and how you're selling and the products you're creating and all that?
00:26:41 · Dylan Field
[00:26:41] S1: I think that there's a bunch of ways this could play out and so we'll have to see. But overall, first, I think it's not even clear how do you define the unit of an agent? Yeah. So that's kind of a difficult one. But yeah, in general it should be the case that there's great services that you offer if folks are trying to do inference and use AI with your software and your platform that should result in like value to the customer and it should be the case that that's driven I think by humans. Overall, I think that humans being the loop will be the way that you steer towards the best outcomes. And I think that all my experience with AI so far have shown that on formal domains, things that are verifiable, you can get to amazing outcomes with enough IQ points. And yet judgment is somehow not purely correlated with just adding IQ points to the model. Design certainly does not seem to be correlated with adding IQ points to the model. These seem to be other capabilities that you have to figure out how to train and judgment I'm really curious about in particular because yeah just intuitively you know the folks that are really good at strategy like I have not even found with humans the common characteristic of how to find them. In design, think, you know, how do you teach a model true empathy for a user? How do you make it so that they can not just, you know, read a transcript of a user interview, but understand the relationships that user has, the inflections in their voice when they're talking about the problems they have and how it matters, and connect that with, like, their world and actually have that empathy for their worldview and where they're coming from to build for them, I I think it's really hard.
00:28:56 · Molly O'Shea, Dylan Field
[00:28:56] S0: What is the biggest question right now that people should be asking? [00:29:00] S1: Actually, I think that, you know or Corey, this on a Tuesday. Maybe I I haven't looked at the news this morning. Maybe it's already a different question. But as of right now, we've got live situation with US government in Thropic where they're discussing a jailbreak. I'm not even sure if jailbreak's the right word, but the use case they're concerned about seems important. But I would say that jailbreaking is something that is understudied, and I actually think that there needs to be a kind of different posture towards trying to get people to break jailbreak models in order to battle test them the same way that we do with like HackerOne. There are programs that are being run by the labs, of course, but I think that there needs to be way more incentive out there to do that. And we need to define what it means. Like, jailbreaks are very wide in terms of the spectrum of what they do, you know? And I I don't think we can have a conversation like this without even, like, defining what a jailbreak is. So that's one question I would say is what's a jailbreak? [00:30:13] S0: You seem to be doing some jailbreaking yourself. So what are what the process I don't know.
00:30:13 · Molly O'Shea, Dylan Field
[00:30:13] S0: You seem to be doing some jailbreaking yourself. So what are what the process I don't know. [00:30:20] S1: Is a jailbreak? So no. I mean, I think that there's a period of time where I enjoyed just, like, trying to figure out how models worked, reasoned from the outside end. I'm more enjoying now just, like, with open weight models looking at, you know, interpretability research and whatnot when you can actually get access to compute [00:30:42] S0: Mhmm. [00:30:43] S1: For that. It's pretty fun. But, overall, I think, you know, it's mean, just just talking with models long enough leads to states that you could call a jailbreak for basic ID model. Yeah. But, again, what is a jailbreak? It's not like you have to do some fancy JSON thing. Like, that's what people like to popularize it online with. But, I mean, you can just talk to them and ask them questions. [00:31:13] S0: Are you token maxing? [00:31:14] S1: I think that yes and no. Like, I try to have for stuff that I'm working on. Oftentimes I'm trying to get the agent running longer as long as it's got a defined goal. But yeah, mean, I could be running like an agent swarm or something and I'm not doing that. We have plenty of people at Figma that are and are getting great results though. And it's amazing how much we can do there when you have folks that really understand the infrastructure that they're trying to build and they go and, you know, they're already 10x and they can 10x themselves again by having the agents really break up tasks properly.