The Exchanges

Q So what will that do? Because that's another big meme that people talk about.

A So what do you, well, it would help with the energy crisis and Climate crisis, because, um, if you had sort of cheap, uh, superconductors, you know, then you can transport energy from one place to another without any loss of that energy, right? So you could potentially put solar panels in the Sahara desert and then just have a, the, the, the, the superconductor, you know, uh, funneling that into Europe where it's needed. At the moment you would just lose a ton of the power to heat and other things on the way. So then you need other technologies like batteries and other things to store that. Cause you can't, you can't just pipe it to the place that you want without, without, without being incredibly inefficient. So, uh, but also materials could help with things like batteries too, like, but come up with the optimal battery. I don't think we have the optimal battery designs, um, that maybe we can do things like a combination of materials and, and, and proteins. We can do things like carbon capture, you know, modify, uh, algae or other things to, to do carbon capture, uh, better than, um, uh, our artificial systems. Um, I mean, even the one of the most famous and most important chemical, chemical processes, the harbor process to make fertilizer and ammonia, you know, to take nitrogen out of the air, um, was, was, was something that allows modern civilization. Uh, but there might b…

AI assessment note: “it would help with the energy crisis and Climate crisis, because”

Q That there have been a lot of tricks that have been put on top of LLMs. Um, I hear often about scaffolding and orchestration and AI that can use a tool to search the web, but it won't remember what it learns. As soon as you close that session, it forgets. Is that just a limitation of the large language model paradigm?

A Well, look, I think there is, and I'm definitely a subscriber to the idea that maybe we need one or two more big breakthroughs before we'll get to AGI, and I think they're along the lines of things like continual learning, better memory, longer context windows, or Or perhaps more efficient context windows would be the right way to say it. So don't store everything, just store the important things. That would be a lot more efficient. That's what the brain does. Um, and better long-term reasoning and planning. Now, it remains to be seen whether just sort of scaling up existing ideas and technologies will be enough to do that. Uh, or we need one or two more, uh, uh, really big insightful innovations. I'm probably, if you were to push me, I would, I would be in the latter camp. Um, but I think, um, no matter what camp you're in, we're gonna need Large foundation models as the key component of the final AGI systems. Of that, I'm sure. So, I don't, I'm not subscriber to someone like Jan LeCun who thinks, you know, that there's sort of some kind of dead end. I think the only debate in my mind is, are they a key component or the only component? So, I think it's between those two, two options. And, and for me, we, this is one advantage we have of having such a deep and rich research bench. We can go after both of those things At maximum, with maximum, uh, force, both, you know, scaling …

AI assessment note: “we need one or two more big breakthroughs... along the lines of things like continual learning”

Q And Debus, do you think the majority of your improvement is coming from building bigger data centers and using more chips? Like, there's talk about how the world will be just wallpapered with data centers. Is that your vision?

A Well, no, look, I mean, we're definitely gonna need a lot more data centers. Um, it's amazing that, you know, it still amazes me from a scientific point of view, we turn sand into thinking machines. It's pretty incredible. But actually, it's not just for the training. Um, it's, it's now we've got these models that everyone wants to use, you know, and actually we're seeing incredible demand for 2.5 pro, and I think Flash, we're really excited about how performant that is for, uh, the incredible sort of low cost. Um, I think the whole world's gonna want to use these things. And so, we're going to need a lot of data centers for serving, and also for inference time compute. Giving, you know, you saw, you saw DeepThink today, T.P.P. Pro DeepThink. The more time you give it, the better it will be, and certain tasks, very high value, very difficult tasks, you'll want to, it will be worth letting it think for a very long time, and we're thinking about how to push that even further, and again, that's going to require a lot of chips at runtime.

AI assessment note: “we're going to need a lot of data centers for serving, and also for inference”

Q Yeah. And then on the reasoning front, you said that's another thing that's missing. I mean, that's, everybody's talking about reasoning right now. So how does that end up getting us closer to artificial general intelligence?

A Right, so reasoning and, and mathematics and other things, and there's a lot of progress on maths and coding and so on, but let's take maths, for example. You have systems, uh, uh, some systems that we work on, like alpha proof, alpha geometry, that are getting, you know, silver medals in maths olympiads, which is fantastic, but on the other hand, Some of our systems, those same systems are still making some fairly basic mathematical errors, right? And for various reasons, um, like the classic, you know, counting the number of Rs in strawberries and so on, and, and, and the word strawberry and so on. And, and, um, is 9.11 bigger than 9.9. Uh, and so, and things like that. And, and, and of course you can fix those things and we are, and everyone's improving on those systems, but we shouldn't really be seeing those kinds of flaws in a system that is that capable In other domains, in more narrow domains, of doing, you know, Olympiad-level mathematics. So there's something still a little bit missing, in my opinion, about the robustness of, ah, these systems, and then that's, I think that speaks to the generality of these systems. A truly general system would not have those sorts of weaknesses. It would be very, very strong, maybe even better than the best humans in some things, like playing Go or doing mathematics, but it, it would be overall consistently good.

AI assessment note: “A truly general system would not have those sorts of weaknesses. It would be very, very strong”

Q You had an AI that knew it was going to lose a game and decided to hack its way around?

A I think we're anthropomorphizing these things quite a lot at the moment because I feel like these systems are still pretty basic. I wouldn't get too alarmed about them right now, but I think it, it, it, it, it shows the type of issue we're going to have to deal with. Maybe in two, three years time when these agent systems become quite powerful and quite general. So, and that's exactly what AI safety, uh, experts are worrying about, right? Where systems where, you know, there's unintentional effects of the system. You don't want the system to be deceptive. You don't, you want it to do exactly what you're telling it to report, report that back reliably. But for whatever reason, it's, uh, interpreted the goal has been given in a way where it causes it to do these undesirable behaviors.

AI assessment note: “I think we're anthropomorphizing these things quite a lot at the moment”

Q yourself and some colleagues, uh, kept pointing your phone at things and asking an assistant, Alpha, what was going on. And I was yelling at the computer, as I usually do, uh, and said, this guy needs glasses. Like, he needs smart glasses to be able to do it. The phone is the wrong form factor. Um, what is your vision for AI glasses and when is the rollout happening?

A Yeah, I think you're exactly right. And that, that was our conclusion. It's very obvious when you sort of dog food these things and internally that, as you saw from the film, we were holding up, you know, you're holding up your phone to, to, uh, get it to tell you about the real world. And it's, it's, it's, it's, it's, it's amazing. It works, but it's not the, it's clearly not the right form factor for a lot of things you want to do, you know, cooking or you want roaming around the city and Asking for directions or recommendations, um, or even helping the, you know, partially sighted. There's a huge, I think, use case there to help, uh, uh, with those types of, uh, situations. And, um, and for that, I think you need something that's hands-free, and the obvious thing is, for those of us, anyway, that wear glasses, like me, is there, is to put it on glasses, but there may well be other devices, too. I'm not sure that glasses is the final form factor, but it's definitely, it's obviously a clear Next form factor. And of course at Google and an alphabet, we have a long history with glasses and, um, maybe we're a bit too early in the past, but I think the, my analysis of it and talking to the people working on that project was a couple of things that the form factor was a bit too chunky and clunky and the battery life and these kinds of things, which are now more or less solved. Um, …

AI assessment note: “I think the thing it was missing was a killer app”