00:01:07 · Sebastian Johnson, John Daley
[00:01:07] S0: Amazing. And where did this come from? You know, why is this a problem that you two set out to solve? [00:01:14] S1: Yeah. Sure. I can provide a little bit of context there, and then I'd love Jack to share more about the the technical angle that he has been leaning on. So I worked at at Palantir for eight years, worked across a wide variety of roles and functions there, but spent a lot time both on internal security compliance data protection type work, and was also embedded with a number of our customers in both government and financial services are working on these types of problems. And from 2023 onwards, as I began to use the Frontier AI models in my work, it just became obvious that, one, you're gonna have to totally rethink how this work was done as the offensive capabilities of of third actors became... Began to use the AI in novel ways, but also it was gonna totally change what was possible and what was economical to deliver on these workflows. And then I met Jack, who is an AI researcher and engineer by background, and I think he he made my vision even more ambitious because he said, you're not thinking deeply different how much the agents are gonna be able to do. Yeah, Jack.
00:02:23 · Jack Miller
[00:02:23] S2: Yeah. Yeah. Absolutely. I mean, I think similar similar story to John. You know, I was kind of in the AI world, actually, as an undergrad back when the original, like, scaling papers came out, and you were just seeing this kind of ever improving, scaling, know, originally from from OpenAI. Obviously, Google did some work as well, in establishing the scaling laws. And what you could see is that, like, pretraining was really working. You know, every model release, it was unclear what the new emergent behavior would be, but it was clear that there would be some improvement. And so you could see that scaling, you know, occurring, in the early in the early twenty twenties, and so it was pretty obvious that, you know, a model like kind of GPT-four, GPT-four point five, would eventually be reached, and that would be incredibly helpful for, loads of different tasks. But then what you saw after that, of course, was the introduction of reinforcement learning, And it really was at that point in time that things became foreseeable that we would reach the kind of situation we're in today where models are trained to... Trained on these cyber tasks to to go and, you know, capture the flag. It's a very easily reinforced reward to get really, really good at offensive cybersecurity, and we... We've seen that come in, you know, with the mythos class of models. And so, you know, the timelines were unclear, but the eventual state, I think, was foreseeable. And what I really wanted to do was start a company that is leveraging the frontier for defensive capabilities. And I think what you see is, especially, you know, around 2324, a lot of the advice was going... You know, if you really care about this sort of stuff, go and join the Frontier Labs. I think what we've seen is that the Frontier Labs have a significant incentive to to race, to get, you know, the best model out, that they can against both themselves and against China, of course. And that actually relatively few people are thinking about how to use the frontier for genuinely defensive capabilities. And I thought that that's where my time would best be spent. There was a moment where I thought I'd do a PhD, but, like, you know, if I did a PhD, I'd still have another couple of years. And I think by that time, maybe we would have AGI and, like, it could do my thesis in the last two months.
00:04:52 · Jack Miller, Sebastian Johnson, John Daley
So I was like, maybe... [00:04:54] S0: But you're like, is the time [00:04:55] S2: for sanity. You know? [00:04:57] S0: Yeah. Yeah. Yeah. Now is the time of building. [00:04:59] S1: Yeah. I mean, we're we're talking the day after g p six, Astra came out, and I think the AGI conversation is so funny. Right? Because if you drop someone from 2024 in front of g p... They're like, oh, this is, of course, AGI, but I think because it's like this incremental, but a very rapid improvement, we're sort of a nerd to the fact that, like, these capabilities are just amazing, and I think, you know, we really enjoy talking with prospects and customers and helping them understand how much you can actually do, let's say, and and and what that means to get deployed in your enterprise. [00:05:30] S0: And how do you react to a new model? You know, like like the new model that just came out, what happens for you in the business? Is this like, oh my god. Like, we have to test everything. We have to rebuild everything. How do you react and building the defensive capabilities? Well, yeah, what happens for you when a new model like this gets dropped?
00:05:30 · Sebastian Johnson, John Daley
[00:05:30] S0: And how do you react to a new model? You know, like like the new model that just came out, what happens for you in the business? Is this like, oh my god. Like, we have to test everything. We have to rebuild everything. How do you react and building the defensive capabilities? Well, yeah, what happens for you when a new model like this gets dropped? [00:05:47] S1: Yeah. I think there's a philosophical position and a technical angle on that. So maybe I'll speak to the first and then Jack speak to the second. I think we've built the company on the thesis that at the application layer, you can just ride the wave of improving model capability over time. And so we use a wide variety of models, the right task... Sorry. The right... A model for the right task, and we do a lot for our customers in terms of of optimizing cost and and performance. And we think that's just going to increase, and, ultimately, we're one of the beneficiaries of the hundreds of billions of dollars being spent at the frontier plus the amount... The the enormous amount of open source innovation coming. Now at the same time, you need to make sure that as new models come out, they're performing consistently with with expectation, and you're running evals. And so I'll check... Talk about how we think about that.
00:06:43 · Jack Miller
[00:06:43] S2: Yeah. Absolutely. And I'd I'd kind of put new model releases on kind of two dimensions. One is, like, improvements to frontier intelligence. So for example, you know, Astra, I haven't looked specifically at the benchmarks to see if it does in fact beat Fable 5.1. I guess they're probably pretty similar. But one is, like, improvements to the frontier, and there are some bits of our platform where that that helps, you know, particularly long running, tasks, but it it helps a lot obviously with internal stuff, you know, software engineering. We've got kind of a fleet of virtual machines that we use now for development where we just like basically parallelize everything, and have this this long verification loop and and new models are incredibly useful there. But then the second dimension is is cost to serve. And, you know, you see every year this kind of 10 x marginal decrease for the same level of intelligence. And so that helps us immensely in the cybersecurity domain because, you know, kind of the point of our platform is to move from a world where you sample. You just take a small sliver of an organization and work out in this part, are they doing what they're meant to be doing to a population sample. We wanna look at everything all the time to determine whether it's, secure and meeting kind of the requirements, either from the regulator or from from customers. And so I look at both, and it tends to be when, like, you get an improvement in the frontier, we tend to take something away from our scaffold. So, you know, as the models get more intelligent, they need less and less guidance. And then, if a... If you get a radically cheaper model, then we tend to expand the set of things that we look at, or we provide something for free that we otherwise wouldn't have provided for free before. So that's what that's what tends to change.
00:08:27 · Sebastian Johnson, John Daley
[00:08:27] S0: That's really interesting. I I feel like we're getting more into the meat of of, you know, what you're really building and how you're serving customers. Could you give back an example? Because I know that you're working with Callosum. Could you talk about a bit, like, you know, the type of customers and clients that you're serving and what it is that they're really looking for? [00:08:44] S1: Yeah. Sure. Jack, do you wanna talk about the Callosum work?