The Exchanges

Q Yeah, that's, that's wild. So as app alpha fold applications expand, I mean, as Google comes up with more, you know, programs like this, does it change the nature of Google's business? I mean, alpha folds and Google search are very different. So talk a little bit about how that fits together.

A Yeah, it's, it's, uh, you know, it's a good point. These, these two things are quite different. Um, and alpha fold is a good example here of, you know, I described some of the ways That is having impact in the world. When I looked at AlphaFold, to your point, I was like, well, how does this work with Google search? It's not obvious, right? How does, how does, how do these two things knit together? What's the match there? Um, so it took a step back with my team and, um, thought about other ways we could employ and deploy this. And it seemed, we, we looked across a range of different areas and business opportunities, by the way, from agriculture to all sorts of areas. But in the end, we concluded that actually there was a great opportunity Here in drug discovery. You know, it takes 10 plus years to develop a drug, and then often when it goes into clinical trials, it fails. There's a very, very high failure rate, and so you've spent all that time and money and investment, and it doesn't actually make it through and solve the clinical need you're concerned about. So having understood this kind of scale and importance of the problem and the opportunity, that then gave us the impetus to form a new company. Uh, so we, we formed a new company, which is a sister company now to Google DeepMind. It's part of the overall alphabet group. It's called Isomorphic Labs. It's about two years old…

AI assessment note: “we formed a new company, which is a sister company now to Google DeepMind”

Q So that's even a little deeper than you just mentioned, resulting in significant advances for humanity. How, how will this technology, A, Provide a deeper understanding of our world, but be, you know, if, if we can apply, you know, if it can think like a human, then where are we gonna, how practically will it be able to achieve some of these, you know, goals that you just discussed?

A So I think that the application in science is a really interesting area to think about. Science is an incredibly complex area that we as humans over the years have made incredible advancements in, and those advancements in science have really enabled us to build the societies we have today, have the health that we have today, have the food production that we have today, but at some level it feels like we're meeting the limits of the knowledge that humans alone can create through this process, and a concrete example of this is Um, proteins, uh, they're the building blocks of life. They're what makes you and I work, Alex, right now. If we didn't have proteins, the little machines in our bodies, uh, we wouldn't function. And scientists for years have been trying to determine the structure of proteins, actually, because if they go wrong, if they're misformed, um, if the structure isn't quite right, then that can cause things like disease, um, and all sorts of, um, uh, maladies across in a whole range of different areas. The challenges for humans, we've been trying to do that using experimental means. So it takes years of painstaking research and millions of dollars of specialist equipment to determine the structure of just one of these proteins. Which is why, actually, experts for about 50 years now, actually, these are really, really, uh, advanced scientists have been trying to us…

AI assessment note: “a concrete example of this is Um, proteins, uh, they're the building blocks”

Q Right. And so it's so interesting because, you know, DeepMind's era areas have been the gaming, working on protein folding, which we're about to talk about, um, Google brain, you know, maybe more search related. So how much of your activities are now going to be focused on the core Google business versus some of this other type of research?

A I think it was interesting to know, uh, even at DeepMind, and actually this is very close to, uh, my role is that we, we have for a long time being Um, taking the technology that's been developed in our fundamental research programs and apply that to Google's products and services. So I thought that's actually been a cool part of, uh, both these groups and actually is now a fundamental part of what we do at Google DeepMind. So we're both advancing the state of the art in the technology, applying that to really big problems in science, and then using those breakthroughs to drive value and impact across these, you know, it was often billion user products At Google. And that's absolutely right, Alex. It's fundamental to the, the new setup at Google.

AI assessment note: “we're both advancing the state of the art... and then using those breakthroughs”

Q AI model a written list of principles, a constitution, and instructing it to follow those principles as closely as possible. A second AI model is then used to evaluate how well the first model follows its constitution and corrects it when necessary. You know, I'm curious what you think about this, um, this approach and whether that's something that, you know, Google would consider employing, and if not, why not?

A So this, this is kind of, uh, I would generally think about this approach and then other approaches like this as a way of ensuring these models are behaving in the way that we want them to behave. Um, and we think about, we do definitely think about that very deeply. It's very important to everything we do. And there are different ways to do that. Um, one way is actually by having, um, an AI system like the one you've described provide feedback to the model that you're training. About whether it's behaving in the way that the designers would like that system to behave. And that's certainly something, um, it's all part of the overall approach. Another important way actually is that you have humans providing feedback to the model. This is a process called ROHF that folks might be familiar with where human, human races interact with these models and, uh, observing the constitution and provide feedback to the model on whether and how well, The model is performing against that constitution. And actually at the moment, that's a really important part of, I think, the core research process because humans are actually very good at this. And there's a kind of secondary benefit of that is that we are, um, beginning to understand How we can begin to embed more and more human feedback into the model process. So I think in general terms, yeah, this is a really important part of how we approa…

AI assessment note: “And that's certainly something, um, it's all part of the overall approach.”

Q but you also compared the intelligence to that of humans, so it can't be both of those things, or maybe you can. Um, it seems like, in some ways, we're, like, both, um, you know, it seems like we're both in awe of, like, what these things can do, and still not fully in comprehension as a species of what we're working on. Do you think that's a fair assessment?

A I think, uh, we're at the early stages of a very long ladder, if you like. So we're on the first run and, uh, predicting exactly where research itself will go is always a precarious task. Um, I'm very hopeful if that's what you mean by an all of the potential for this technology to really help lift humanity to new levels, to help things like, uh, climate change, to help in things like health. I think that's really important to keep in mind. Um, We've got to continue to interrogate these systems to understand how they work, to make sure we're doing it in a responsible way, to make sure we get the right review in place each step of the way, so that we do understand, we do roll them out in a way that makes sense for society overall. Then we're going to do both those things. We're going to be bold in how we develop it, but also responsible and take care.

AI assessment note: “We've got to continue to interrogate these systems to understand how they work”

Q DeepMind have talked about, I think I'm going to just cite this, that the algorithm should be better at planning and problem solving. So that seems to be where we're going. So first of all, I'm going to get, you know, I have a few questions for you about Gemini, but we'll just talk about it on a broad level. How do you teach an AI to plan and predict?

A So, um, there's a whole range of different active research tasks here, and, and to be clear, there isn't an answer yet, which is why it's still active research. But one of the ways we motivate this research is by making sure we have tasks that require planning. So we spend a lot of time and investment in building a whole suite of different evaluations and tasks. Which then provide the target, if you like, for our research and our research programs to focus on. And that's a really interesting definition of intelligence. And one of the definitions of intelligence that we use at DeepMind and was actually created by one of the founders of, uh, Google DeepMind, Shane Legg, is intelligence is the ability to perform well across a range of different tasks. And I really liked that definition because it's, I think it's very descriptive and it's very easy to operationalize into a research program. And so this sense of building multiple different evaluations and tasks that provide then a way for us to measure our performance and progress against whether it's planning or adding memory, um, is, uh, is really central to actually the way we conduct research. And then behind that, it's a creative process. Uh, so what you're trying to do is bring together people from a whole range of different disciplines, from neuroscience, from different areas of AI research to, uh, Come together and have idea…

AI assessment note: “to be clear, there isn't an answer yet, which is why it's still active research”