00:17:34 · Ksenia Se, Ulrik Hansen
[00:17:34] S1: So context is coming from people. And what about synthetic data? It seems to be everywhere now for edge cases, rare classes, bootstrapping new domains. Do you think it helps to bridge the gap, or, like, where does it work, and where does it fall short? [00:17:50] S0: Yeah. Good question. So I think the areas, like, where we generally sense... See synthetic data working really well is, like, where you could do physics based simulation. So, like, environments, like, where you can, like, build some sort of, like, model of the world, and you can you can, like, build some, like, stochasticity into the model, and, like, you kinda, like, simulate these, like, real world environments. I think that's been worked really well in self driving where you can, like, simulate some of these, like, long tail edge cases that would be very, very hard to capture in the real world. I think, like, where it's falling short is generally, like, when you are using a model to generate synthetic data and then maybe, like, generating synthetic data and then, like, having another model trained on that data that is already seen. So that can sometimes lead to, like, model collapse because it's kind of like, the snake eating its own tail. But you can also use, like, a bigger model to train a smaller model. So there are, like, different ways of, like, doing it. I think the most successful companies are the ones that use the human or real world data in conjunction with synthetic data. And I, yeah, see that, like, being the the winning formula going forward.
00:17:50 · Ulrik Hansen, Ksenia Se
[00:17:50] S0: Yeah. Good question. So I think the areas, like, where we generally sense... See synthetic data working really well is, like, where you could do physics based simulation. So, like, environments, like, where you can, like, build some sort of, like, model of the world, and you can you can, like, build some, like, stochasticity into the model, and, like, you kinda, like, simulate these, like, real world environments. I think that's been worked really well in self driving where you can, like, simulate some of these, like, long tail edge cases that would be very, very hard to capture in the real world. I think, like, where it's falling short is generally, like, when you are using a model to generate synthetic data and then maybe, like, generating synthetic data and then, like, having another model trained on that data that is already seen. So that can sometimes lead to, like, model collapse because it's kind of like, the snake eating its own tail. But you can also use, like, a bigger model to train a smaller model. So there are, like, different ways of, like, doing it. I think the most successful companies are the ones that use the human or real world data in conjunction with synthetic data. And I, yeah, see that, like, being the the winning formula going forward. [00:18:46] S1: The hybrid hybrid method. [00:18:47] S0: That's right.
00:18:47 · Ulrik Hansen, Ksenia Se
[00:18:47] S0: That's right. [00:18:48] S1: From what I understand, you work with high stakes domains like medical AI, robotics, autonomous systems. Right? What are the biggest mistakes that these companies do from prototype to production in terms of dealing with data? [00:19:02] S0: I think, like, the most most of, like, the very savvy companies that we work with, like, they probably understand, like, what their shortcomings are and, like, what their, like, shortfalls on, like, the infrastructure side would be. And, like, of course, like, they are working very hard to eliminate those. I think, like, at the end of the day, the challenge that, like, companies always have is, like, they don't necessarily have access to the same resources as, like, the frontier labs or the hyperscalers. Of course, like, we're plugging that gap. Right? We're, like, building the the tools and the info that is really, like, only been available to the labs and, like, making that available to the enterprise so that they can move as fast as, like, the frontier labs on, like, building performant AI systems. [00:19:37] S1: So you are their, like, guide, basically, in this scenario?
00:19:37 · Ksenia Se, Ulrik Hansen
[00:19:37] S1: So you are their, like, guide, basically, in this scenario? [00:19:41] S0: Yeah. We're the guide, and, like, we, of course, like, build up a product that that helps them, like, bring their models from from prototype into production, like, doing that faster than they would have if they had, like, tried to, like, build a lot of the stuff in house. And a lot of times, like, it's also just not feasible because the complexity of AI systems being developed, I think, is only really going one way. Right? Like, you wanted to solve increasingly complex types of tasks and the problems. And so there is, like, a never ending and evolving need to continue to push the frontiers of your own AI models and systems. So, like, one thing we've seen, for example, is, like, all the models have gone from being unite models to be multimodal. That, like, opens up, like, a whole new set of things that you need to, like, be able to support. This is not, like, attainable for, like, a Fortune 500 enterprise similar to, like, how they move to the cloud because of all the companies that had, like, previously built in house data centers and infrastructure. Now they can, like, rely on, like, a cloud platform, which means that they can, like, move a lot faster on building in house apps and all these other things. I think the same thing will happen in in AI development. [00:20:37] S1: That's very interesting. Everything that we were talking about is on a very practical level. Are there any conversations about AGI with the companies you work with, or what's your personal take on it?
00:20:48 · Ulrik Hansen
[00:20:48] S0: There are, like, of course, like, companies that are working on on AGI. I think my general sense is, like... So we we principally work on, like, more applied AI use cases where they try to actually bring that intelligence into production. At the end of the day, like, treating intelligence as a single construct is difficult because there's, like, different levels of intelligence. Right? There's, like, the cheap types of intelligence, which is just, like, raw knowledge and facts where, like, the models can, like, spit out infinite versions of something at, like, near zero cost. There's, like, pattern recognition on, like, common things, like writing boilerplate code, summarizing documents, and, like, generating standard designs. All that stuff is, like, quite uneasy to do, but it's been a lot of lot of time development. But I think there will be a kind of, like, bifurcation of, like, expensive and cheap intelligence. So, like, expensive types of intelligence are things like human agency, creativity, original insight, taste, and vision, and the stuff that, like, feels like it's, like, genius or original because it can't be reverse engineers from from the data. And so I think the... Kind of going back to the analogy around having AI systems that evolve and, like, kind of, like, doing both of those things, like the ones where, like, we could do it for cheap, there'll be less need for, like, a human loop and fact checking for, like, the ones that... The things that are, like, will be more difficult or require, like, more expensive types of intelligence. Like, that's really where, like, human agency tastes and all these things that you can't really explain necessarily, like, in data will become, like, more important.
00:22:10 · Ksenia Se, Ulrik Hansen
[00:22:10] S1: Yeah. For me, applied AI seems much more exciting than the conversations about something. [00:22:16] S0: At the end of the day, like, I think, like, a lot of the models themselves are, like, good enough. Right? So, like, the main constraint now really is to, like, embed those models and, like, kind of wrapping them into products that can, like, solve real problems. So I think, like, that's where the main constraint is because, like, the form factor, like, matters a lot. Right? Chatability can help you write emails and all these other things, but you can't, like, necessarily have it do your whole, like, job for you. Like, of course, like, these things are getting better, but there'll be, like, a long tail of, like, different types of, like, human tasks and work that, at least, like, for a substantial amount of time, will be, like, outside, like, the rebid of, like, what the the existing form factors that AI is, like, delivered in can support. [00:22:57] S1: In this world where you work with data in all the stages, you know how it works for training, you know how it works in post training, what it needs for the context, working with high stakes companies when it's super important to really manage your data well and be safe about it. What concerns you the most in this world that you help building, and what excites you the most?