[0:00:00] And without any further ado, I'm going to introduce our first guest, who also happens to be my guest. I'm going to be doing this panel. So please welcome to the stage Google DeepMind CEO, Mustafa Suleiman. Please welcome him. Here we go. How are you doing? Pretty good. Thank you. I'm good. We were just talking about how neither of us are morning people. Right. So when we started the company,
[0:00:28] My co-founders and I, neither of us were morning people, so we actually made the start time 10 a.m., and we still stick to that. So getting up at 7 to get here was a little tricky. Are you hiring? Are we hiring? Just kidding. We're hiring. We're always hiring. I want to work at 10 a.m. All right. I think that DeepMind is interesting because you guys are, just to be clear, DeepMind is the AI research branch of Google. Right. Fair way to put it.
[0:00:57] And you guys have your hands in a lot of different honeypots, right? You've got AlphaGo, which is a machine learning or deep learning tool that can play games. And then you have DeepMind Health, which is such a far cry from AlphaGo. And I guess for the layman, I want you to explain what DeepMind actually is. What is the goal? Right.
[0:01:24] So our mission is to solve intelligence and use it to make the world a better place. And so when we started out five or six years ago, the idea that we would commit to trying to reduce or express intelligence in an algorithmic form and try to take advantage of all the benefits of computing.
[0:01:44] Infinite computation, a lot of storage, infinite data, permanent memory, the ability to run very large simulations and predictions. It seemed kind of crazy that we could treat intelligence like an engineering problem. But we've managed to demonstrate that we can make a lot of progress and develop systems that are increasingly capable of doing the sorts of things that make you and I very smart. And it turns out that that's very useful and pretty valuable, especially when it comes to
[0:02:12] Making progress with some of our most complex social problems. And that's really what motivates us at the company. I think we founded the company on the premise that many of our most complex social problems are becoming increasingly stuck. So we're struggling to make progress towards solving climate change at the rate that we need to. We have enough food on the planet to feed everybody equally, but we're just not distributing it in the right way.
[0:02:38] And on many fronts, we really could be making much more progress, especially when you consider the fact that over the next...
[0:02:46] 30 years or so, we have a massively aging population, a massively growing population, and everybody wants to consume at an incredible rate, just like we here in the UK, Europe, and the US do. And so trying to feed and support that population is going to be a huge global challenge over the next two or three decades, which is why we believe that we need to try and build computers that are smarter and smarter to help us with those sorts of problems. And I think it was 2013 is when you sold to Google.
[0:03:15] Right? Right. 2014? 2014, yeah. 2014, so two years ago, 500 million. Right. There we go. And I'm curious, why sell it all? So, yeah, I mean, that's a fair question. We were in a really, really strong position. We had plenty of runway. We had about 120 or so staff, and we had runway for about four or five years without generating any income. And we had a lot of amazing people backing the company.
[0:03:44] But there were a couple of really big attractions to working with Google. First, they have an incredible compute infrastructure, which was very, very useful to us. And secondly, I think having the opportunity to work on Google products was a massive attraction. I mean, there are more than five products that have a billion users a month, and that kind of scale gives us an opportunity to develop really interesting and increasingly general algorithms. Yeah, data, right?
[0:04:14] Data, in a lot of ways, is the lifeblood of teaching a machine to learn. Right, yeah. I mean, another way of putting that is the training environment. So it's not so much the underlying data, it's the way that you set up the training environment. So when we were a startup independent of Google, we made a lot of progress with training algorithms in the old school Atari 2600 environment. And the benefit of this is that we could generate an enormous amount of training cycles. So we could give our agents an experimental...
[0:04:43] You know, place to learn to build general representations of what is rewarding behavior without really depending on Google-type data. And so there are many different ways that you can look to train those sorts of algorithms, and there's lots of other companies now that don't have Google-sized data but are getting around that in very successful ways. Google has a history kind of of, I remember when we realized that all of our accounts would merge from YouTube and Gmail and...
[0:05:11] All of those things ended up merging kind of into the same pot. I'm curious if that will end up being... I mean, what's the relationship? What's the relationship with you and Google? To what extent do you have access to that data? And are you getting directives? Are you asking for okays? You're here in the UK. They're all in San Francisco. What's going on? How does that work? So I think...
[0:05:37] One of the things that made us successful as a startup was, I think, being in London and being somewhat removed from the sort of valley way of doing things. We were very committed to building a new kind of organization with a kind of hybrid culture, one that really focuses on the benefits of working on very hard, long-term research problems, just like you get in academia.
[0:06:00] But then also combining that with the kind of commercial rigor that you get in a company where you basically have very well structured, parallelizable engineering teams to work on big projects that help us to train machine learning algorithms, good testing environments, good training environments.
[0:06:19] And all the while underpin that kind of culture with the values of the public sector, you know, a real social mission, really wanting to work on the most interesting and exciting problems that have real-world social impact. And I think we found that, you know, being slightly outside of that ecosystem, specifically focused in London, allowed us to, you know, really attract the very best talent. Like, I think it turns out that, you know, the very, very top tier of people can work anywhere they want in the world. They can work in...
[0:06:48] They can work in finance and banking. They can work in the Valley. They can go to New York. And actually, I think the very top tier of people are clearly not just motivated by comp or working on really hard problems. And we can obviously offer both of those. They also want to work with the very best peers.
[0:07:06] They're colleagues that are also awesome. But they want to work on problems that really matter. I think that's something that's changing in the last five to 10 years or so, particularly recently. People really want to make a difference in the world. And I think presenting the opportunity to work on those sorts of problems is something that is very core to our mission and motivation. Yeah. But again, for example, you worked on Google's data centers energy-wise and cooled them down by 40%. Is that accurate?
[0:07:36] So is that Google's idea? Was that your idea? Were you saying, oh, we want to get into energy, and Google was like, sure, come work on this? Or was Google like...
[0:07:44] We need cooler data centers. Can you get on this? Well, so, I mean, when we sold the company, you know, it was very important to us that we were able to continue to run independently and be able to control the kind of research that we do independently of Google and work on the products that we think are important and interesting. And that's been an amazing relationship over the last couple years. And, of course, energy is something that's important to our social mission.
[0:08:10] And so getting the opportunity to work on one of the largest computer infrastructures in the world was something that we really couldn't turn down. And so we did actually some really interesting work. We managed to train a model that looked at the incoming load or demand from typing in a query on YouTube or in search, for example. And of course, that generates heat in the data centers, and we need to efficiently extract that heat.
[0:08:39] Basically, the process by which we do that, we made much, much more efficient. And so we ended up reducing the cooling load required by 40%, which was a net 15% reduction in power usage efficiency, which was obviously a very significant saving to Google. And another part of that social mission, you were talking to me earlier, is this partnership with the NHS, the National Health Service, working to detect AKI, which is a kidney issue.
[0:09:12] If only we had some AI to solve that problem. Can we get the noise down on the monitor? Back to my thought. So the partnership with the NHS, why don't you explain what that's all about and what you're doing there exactly?
[0:09:31] Sure. So we're collaborating with a hospital in North London called The Royal Free. And we've just agreed a five-year partnership to work with them on a bunch of conditions that patients get when they get admitted to the hospital. So we're really trying to detect when somebody is at risk of deterioration. So you mentioned acute kidney injury. This is when your kidneys essentially become dehydrated.
[0:09:56] And when they do get dehydrated, they do a bad job of extracting the toxins and ureas and the bad things from your blood. And that can often lead to infections, which can back up and create further complications. And in some cases, as you say, dialysis. And that's extremely costly and really, really painful for the patient. And so what we were able to identify is that actually if we look at the blood results as they come in in real time,
[0:10:22] We can identify which patients are at risk of contracting some form of deterioration, be it sepsis or acute kidney injury or a bunch of other things. And we can generate alerts and reminders and notifications on mobile to the right nurse or doctor so that they can intervene ahead of time, preventatively getting to that patient before anybody else realized that they were at risk of a deterioration. But that was with an algorithm that the NHS developed.
[0:10:48] Right, exactly. So there is no deep mind machine learning baked into this app, this streams app that you're building. That's right. So then why would you build a front-facing app, right? Well, it's still very early days for us, right? So we've really only been collaborating with the role-free for just over 12 months now. And we think in the future there's a lot of long-term potential to deploy machine learning algorithms to help to do a better job of these problems.
[0:11:17] One area with the NHS that we're really successfully deploying machine learning systems is with the Moorfields Eye Hospital, which is a pretty awesome research collaboration, quite different to the collaboration that we have with the Royal Free Hospital. And essentially, it's a similar kind of problem. We would like to look at some of the data in real time in order to identify which patients are at risk of really blinding conditions like diabetic retinopathy and age-related macular degeneration.
[0:11:46] And so what we're working on there with the Moorfields Eye Hospital is to train machine learning systems that can identify where in an image a particular pathology exists and label that scan, identify the pixels that correspond to the diagnosis that we care about, and then prompt a clinician, an ophthalmologist, to let them know that this patient is at risk.
[0:12:11] Get a preventative treatment in as quickly as possible. So in some cases, this is really as simple as some injections that if you get them right within a 48-hour period, then you really can reduce some of the most extreme forms of sight loss and essentially save people's sight. And so there are lots and lots of examples across the healthcare system. I mean, these are just two small examples, to be honest with you, where getting in there as early as possible.
[0:12:35] can have a massive difference both on cost and improving patient outcomes. And that's why we're really motivated to work in this space. Yeah, and obviously it's more cost effective to prevent something than it is to deal with it after the fact, right? And in a lot of ways, machine learning is about predicting what will happen based on the data, which I guess, again, brings us back to the point of data. I mean, Google doesn't have troves of
[0:13:00] Patient data the same way that it does how you click through on search or what you choose after you watch a YouTube video. And I think that from a layman perspective or from a user perspective, you think, OK, well, Google owns DeepMind. DeepMind is looking at all of this patient data. And not for a bad reason, right? To solve problems, to prevent health crises. But I think psychologically, there's a trust issue.
[0:13:29] That could arise from that. And I'm curious how you go about building trust. And you say social impact, and I believe you, but there's a good chance that some people don't, right? And so how do you go about telling someone, hey, I have patient identifiable data on you. I know your medical history. Trust me, I'm doing it for the right reasons. You're not going to be paying more because I can do it better than a doctor that can do it.
[0:13:54] How do you have that conversation? Yeah, I mean, that's a great question. So to me, trust is a function of control and transparency. So to address both of those two points, control is the first one. So clearly, there's a whole series of legal agreements which ensure that the hospital retains complete control over the data. If they want to boot us out, then they can do that instantly.
[0:14:18] They tell us exactly what they want us to do with the data and what we can and can't do with it. And so clearly that's an important part of the process. The second thing is transparency. So not just relying on our assertion that the data is stored in a particular way or processed in a particular way, but actually having independent third parties to scrutinize that. So when we launched DeepMind Health earlier this year, knowing that it was going to be a sensitive and important question for many people who were worried about exactly the question that you're raising,
[0:14:47] We put in place a panel of independent reviewers. So these are nine public experts, ranging from clinicians to policymakers to ex-member of parliaments. And we invited them to scrutinize us in the public interest. We have proactively given them a budget, which they can completely control. They've hired a secretariat. They're hiring auditors, legal experts, governance experts, technical reviewers. And we meet with them periodically. They can come into our office and meet any of our team.
[0:15:16] Interview them. They're planning to publish an annual report. And so this is just a small first step towards us introducing a level of oversight and transparency, which I think will go some way towards helping to create that trust that we're both talking about. But there are other technical approaches that we're taking as well. We are pioneering a general transparency architecture.
[0:15:40] So this is a distributed verifiable logging architecture that describes where data has moved when. Yeah, who looked at what, when, for how long, for what purpose. And that's so important because otherwise you're sort of relying on the kind of assertions of a third-party expert only. So it's necessary but not sufficient. And it would also be awesome if you could look up on your mobile phone, you know,
[0:16:05] Who has your data? Which clinician looked at it? When did they look at it? Under what policy and why? And we believe that we can build a system like that in a very distributed and ultimately in an untamperable way. We believe that we can generate a log of who has interacted with that data and where it has moved in a way that is mathematically proven to be.
[0:16:27] And so that's another step that we're taking to try to increase the level of transparency around your most sensitive and personal data in conjunction with the kind of human oversight that I mentioned. Well, and I'm glad you brought up that human oversight piece because that's for DeepMind Health. But also as part of the acquisition with Google, you did insist upon kind of an ethics board, right, for DeepMind as a whole. And Google has not been okay with sharing the names of those people.
[0:16:57] So look, I've said many times, we want to be as innovative and progressive and open with our governance as we are with our technology. It's no good for us to...
[0:17:10] Just be technologists and in a vacuum independently of the social and political consequences, build technologies that we think may or may not be useful in the world and throw them over the wall. And so we've been experimental in that respect. The ethics board is ongoing and it's something that we have internally to oversee some of our efforts. But we've also tried other approaches, not least the independent reviewers on health, but also...
[0:17:36] Most recently, we got together with Amazon, Facebook, IBM, Microsoft, and Google to start the partnership on artificial intelligence. And this is specifically looking at how we can develop and share best practices around how we build and deploy machine learning systems in the world. And we've invited an equal number of non-corporate, completely independent members to join us on the governing board of this new not-for-profit organization, which we'll be announcing in early January.
[0:18:06] But that original kind of anonymous ethics board, can you tell us anything about the composition of it? Are there people from Google on that board? Is it racially and gender diverse? I mean, can you tell us anything about who those people are? Look, I think at the moment that's something that we're working on internally. And we've always said that that's going to be very much focused on full general purpose learning systems. And I think that's very, very far away. We're decades and decades away from.
[0:18:33] So you're talking about kind of removing the black box?
[0:19:00] In a way, because that is a big concern when it comes to AI. You don't know how you get from A to Z. With an algorithm, you can figure that out. With a machine learning or a deep learning model, you don't know how it made its decisions. Well, I mean, to put it in context, I think we have this issue across the board. I mean, many of our most complicated software systems are incredibly difficult to debug, and when they go wrong, they cause massive impacts, whether it's a...
[0:19:26] Airports or hospitals or in transport systems. So in general we have this broader question of how we verify what our technical systems are doing and how we scrutinize them, ensure that they're transparent and ensure that we have control of them. As humans we clearly must remain completely in the loop with these systems and I think they're some of the research questions that we're hoping to coordinate across all of the companies on and make all of that research public and open and share it with the community.
[0:19:55] Yeah, and I think there is in a lot of ways like kind of fear of reliance on these types of systems, especially when it comes to things like world hunger, right, or our health. We can see just in something as simple as the Facebook news feed, you know, how does Facebook's, you know, algorithm know the difference between an accidental click or a real click? And I think, I'm wondering, are we going to pass on our,
[0:20:26] Our flaws to our machine learning algorithms and systems. Because you can see there's that ProPublica story about Compass where our historical data of putting more black people in prison than white people, the machine learning can see the correlation there and not necessarily understand the causation and assume that black people are more guilty than white people. And how do we recognize that and how do we prevent it?
[0:20:55] Yeah, I mean, that's a great question. So the ProPublica piece is awesome. And I think Julia Angwin is doing really, really exciting work. I think this is one of the most important questions of our day. I think the way I think about these things is that, you know,
[0:21:09] We are destined to project our biases and our judgments into our technical systems. And if we don't think consciously as designers and technologists about how we're building those systems, then we will, without really realizing it, unwittingly introduce all of those same biases into those systems. And so, in some ways, the exciting thing about technology systems is that it presents an opportunity for us to...
[0:21:34] Critically reflect on how we're designing systems that interact with the real world. And so we should constantly attempt to do that in an open and transparent way and try to rebuild our world in some sense with fewer of those biases and those judgments and those things that we want to shed as we move forward as a species and evolve. Yeah. And you did mention that you just went into a partnership with IBM and a number of other organizations, which makes me feel like you might not answer this honestly, but is Watson legit?
[0:22:04] I'm not sure. You have to use it. I haven't used it. Ah, liar. You are lying to me right now. You haven't used Watson? I have not used Watson. Oh, you just not read up on it. No, I don't pay attention to those sorts of things. Oh, you're kidding me. Okay, final question because we are over time. I think that a lot of people learn more about AI from a science fiction movie than they do about having, they're not as lucky as I am to have a conversation with you.
[0:22:31] What would you point to if someone was saying, what's the closest representation we have in Hollywood or otherwise of what is actually happening in the world of AI? Well, I think in general I would say don't look at Hollywood. I mean, Hollywood does a good job of making fun and light of, you know, I don't know, politics or love or, you know, war. I mean, these are things that we have experienced in the past and we know well.
[0:22:58] When it comes to imagining what the future is going to be like, a lot of that is fun and entertaining, but it doesn't really bear a great deal of resemblance to the kinds of systems that we're actually building. I can't really think of a film that makes me think, yeah, AI is going to look like that. If there's any filmmakers out there, I think this is a great opportunity to produce much better quality sci-fi.
[0:23:25] And I lied, that's not my last question. We consistently hear that like general AI, AGI, is 20 years out. We've heard that for the last like 80 years. How far out are we from a kind of general, broad, and artificial intelligence?
[0:23:43] I think when we say that it's 20 years out or it's decades away, what we're actually saying is it's so far out that we can't really measure it. So anything beyond that sort of time horizon, it's very difficult for me to say the difference between 20 years and 50 years. And so I think it's definitely a long way off, and it's just very far from the kind of practical things that we're able to make work today. So it's a long way away. Cool. Well, I really appreciate you taking the time, Mustafa. Thanks very much. Appreciate it. Thank you. You can take that with you.
[0:24:13] That mic, yeah.