00:07:20 · Ksenia Se
[00:07:20] S1: Do robotics have the same bottlenecks as, autonomous systems like self driving or it's different?
00:07:25 · Ulrik Hansen
[00:07:25] S0: Yeah. No. I think, a lot of the... They have a lot of the same constraints. So I think, ultimately, the difficulty in deploying, like, these, like, robotic systems is, like, you have the additional constraints of you have to deploy the model on, like, specific, like, hardware devices on edge. Right? So you have GPU constrained environments that you don't have your... If you're leveraging a foundation model like a chat interface. You have to, first and foremost, like, make sure that, like, your systems can actually work on these, like, embodied AI devices, which is, I think, no small feat of of engineering. And then, like, creating the iterative feedback loop between the actual product usage and the model itself is also much, much harder because you might have, like, a robot folding laundry in your house, and it might fail. So now, look at, okay. How do you relay that back into, like, the system and, like, how it's trained? It becomes, like... There's, of course, like, latency for, like, uploading the data. You might, like, be in, like, an area without, like, Wi Fi, so you don't necessarily have Internet connectivity. Whereas, like, where we have seen the fastest progress in AI has been, like, in the pure digital workflows. Right? So for language models, like, ChatGPT and and otherwise, for coding, where, like, the feedback loop between, like, the actual user and the model itself is instant. That's another luxury that you have in physical and and embodied AI. So the companies that are building those types of domains, like, they have to really think carefully about, like, how they collect that feedback. And we see it, like, being solved in a number of different ways. Like, some companies are adding 10 operations to, like, help steer the robot if they kind of go off the rails. But, ultimately, like, if you want them to work on a fully autonomous basis, right, you don't necessarily, like, will be able to rely on that. So being clever about, like, how you build those, like, feedback loops is gonna be what determines a lot of the future compounding advantage.
00:08:59 · Ksenia Se
[00:08:59] S1: Feedback loops, we will see more and more of that in AI agentic systems. Right? How does infrastructure evolve to address that?
00:09:08 · Ulrik Hansen
[00:09:08] S0: Yeah. That's a great question. It's something we've thought a lot about because we we are, like, building a a key part of, like, that infrastructure that makes, like, the human feedback possible. I think what we've seen, like, generally over the last couple years is, like, the companies have evolved from building, like, out the kind of core foundations of, like, the models that they've been building and, like, as they go live in production, then, like, building the human feedback infrastructure, like, becomes much more important. And per use case, right, like, you have to have a lot of, like, custom tooling. So if you're building, like, a model that can work, like, LIDAR or three d data or, like, video data, you need, like, good tools and, like, the whole, like, AI tool chain, like, it's just very much, like, in development. Right? Like, a lot of this stuff, like, hasn't really been built yet. So, like, we will see that happening over the next, like, I think, five to ten years. If you think about, like, how the Internet economy has been built over the last thirty years, the development life cycle has, like, gotten to to where we are now. I think AI is likely to be much more... It will happen on a much more compressed timeline. And, yeah, of course, like, we're kind of, like, one of the companies that are building that space, us alongside, like, other ones. What that tool chain would look like exactly, I think it will probably look like a lot of the functionality that we have in in software development. So, like, version control systems, like, we have, like, good networking protocols and all these things. We will have to build out, like, similar things for for AI specifically. But the interesting quirk about AI systems is that becomes a little bit more difficult to do because they are, like, dynamic. Right? And they're sort of, like, working in a real environment. So, like, not constrained to to the same environments, like, software are working in, which are, like, inevitably, like, much more programmatic.
00:10:36 · Ksenia Se, Ulrik Hansen
[00:10:36] S1: Yeah. A lot of uncertainty to deal with. [00:10:38] S0: 100%. [00:10:39] S1: You were talking about involving more PhD students for annotation, working with data. So if expert feedback is the new goal, what is the role of humans in this human? Is it, like, coexisting task with agents? Are humans just global functions in this in this system? Who are humans here?
00:11:00 · Ulrik Hansen
[00:11:00] S0: Yeah. No. I think, ultimately, like, the humans will be providing human preference and feedback to the AI systems, and I think that will happen in, like, many different ways. So let's, like, maybe take, like, two examples. One example would be you with your personal AI digital assistant. You will provide feedback to that digital system or, like, how you like to... I don't know. Like, when do you like to wake up in the morning? Like, what do you like to have for breakfast? And, like, that will continue to, like, learn based on, like, the human preference data that you give it. So that's, like, on the more consumer side, and I think that's already happening in ChatGPT, right, where, like, the model is kind of getting personalized based on how much context that you've fed us. And I think that's some of the things that the Frontier Labs are working on is really, like, giving, like, the models as much context as you can from, like, your personal human preference data. It will, like, be able to offer, like, more and more customized advice and, like, be more of, like, a... An assistant to you. I think in the enterprise and, like, more of the kind of commercial side, all the human experts now are, like, really people that are, like, doing their day to day jobs. So if you're, like, a... Let's say, an investment banking analyst at, like, a big investment bank, you are, like, doing the work. AI will be observing, like, what you're doing. And so you are, like, quote, unquote, like, human experts that are providing feedback to the model, like, how, like, certain workflow processes and all these things should, like, operate. And I think as, like, AI is kinda getting pushed into the the enterprise, that's something that is likely to, of course, like, become more and more, like, prevalent as, like, you will be at work constantly providing feedback and, like, work with AI to, like, do your job, essentially. So it's happening both in, like, the rollout of AI, but also happening, of course, like, as discussed before, at the kind of pure training layer for pushing the frontiers of, like, the models where, like, we've started, like, seeing the asymptote.