00:28:57 · Aditya Agarwal
[00:28:57] S1: 100%. Yeah. I know I was joking with someone that at times, teams in these labs, when they're complaining about the lack of compute, that seems to be more than what, as a country, that we have in India, right? So interesting times. Let's talk about the future a bit here, right? Of course, language is a scaffolding for human thought. But intelligence needs various other inputs. Have a bunch of talk around world models and models for specific purposes, models for science, and a bunch of other things. How do you view that part of, you know, emergent thoughts around some of the new ways of how possibly people should be interacting with, yeah, with both processing and availability of intelligence on those access.
00:29:46 · Srinivas Narayanan
[00:29:46] S0: No. I mean, the the opportunity to take intelligence and actually make useful things in the real world is is really different now compared to six months ago. Yeah. It's very exciting that, you know, the forefront of what you can imagine has moved from just purely digital software to how does intelligence interact with the real world. Right? So you're seeing enormous amount of innovation happening in robotics, in AI interacting with the physical world, in being able to do science, whether it's material science, whether it's life science. You know, math is already happening to some extent. Right? So I think you're gonna see that quite a bit. The cost of writing software has, you know, gone down dramatically as a result of all these models. So it is sort of like push the frontier of what people are thinking about from purely being in a digital world to actually being, okay. Now how can we use this in the physical world? Physical world starts... Has constraints. You know? How fast can we get verifiable data back and verifiable rewards back because, you know, that's not happening at the speed of light anymore? So that creates different challenges, but that is the opportunity. You know? You need new challenges, and I think there's a lot of innovation happening in how do you accelerate the feedback loop and how do we do that. So I think it's quite exciting.
00:30:59 · Aditya Agarwal
[00:30:59] S1: And one of the things that I've been personally curious about is why are these models, right, all of them so fundamentally inefficient when you compare them to carbon based life forms, right, for relatively low use of memory, low power, and basically small fluctuation in voltage, we are able to do amazing stuff. Right? And we... At least we also seem to be continual learning machines of short sight. One of our founders, Ravi Shu, is building in that problem space. So I'm just curious on what are the technical challenges? What are possibly the research based challenges that you're struggling with.
00:31:30 · Srinivas Narayanan
[00:31:30] S0: Yeah. Biology is awesome, and sometimes you have to pay respect to that. We've had six billion years of evolution, so give a give a little bit more time to AI. We've only had a little bit of time. But, also, I see this trend in every technology. Right? Generally, what happens in technology is that you prove that something is possible. And when something is possible and it becomes valuable, an entire industry collapses around how do we then make it scalable, efficient, fast, cheap, etcetera. And you're seeing that happen as well. Look at the amount of innovation and investment that is going in chips, in energy, and so... And in algorithmic innovation as well. And, you know, the cost per unit of intelligence has continued to come down pretty dramatically over the last few years. Right? I would say it probably comes down 10 x, like, every year or every two years. Right? And so that continues to happen through algorithmic innovation, hardware innovation, and innovation in the infrastructure layer. I think that'll continue to happen. You know, I still think there are there are differences between biology and stuff, but, you know, whether we do continue on learning. But you should continue to expect the cost curve to come down quite a bit.
00:32:37 · Aditya Agarwal
[00:32:37] S1: You know, today, one of the things that most builders struggle with is both memory degradation and context compaction, right, that happens once in a while. And suddenly, it seems that a bunch of things that you thought that can be outsourced, you have to take care of them yourselves, right? We internally have built our own harness for working with agents. So I'm just curious. How are the teams inside the labs thinking about some of these things?
00:33:03 · Srinivas Narayanan
[00:33:03] S0: Yeah. There's quite a bit of innovation happening around the model itself in terms of how do you manage the context, how do you compress context if it gets too big, How do you learn whether it's memory or other things? How do you get them really good at using tools? Mhmm. How do you express them in skills, and what are the primitives? So, yeah, I think it's not just the model, but the entire environment that it happens in. One of the things that all the mall... The the labs also do is that they try to incorporate a lot of that into the training process itself so that, you know, what you encounter in different situations matches the distribution. And so that, you know, the RL training kinda matches what you see, especially in coding products and things like that. But at the same time, I think there is also... You know, if I am starting a company and I want independence from a model, you want your hardness to be able to work with different models. I think that gives you more flexibility. So I think that is also exciting because you can then kind of build your own sort of customization and modes around these things. So I I think there is tremendous innovation left to be happen, not just at the intelligence layer, but things around it as well.