{"url": "https://sphinx.acast.com/ft-tech-tonic/howdeepmindvanquishedgo/media.mp3", "page_url": "https://play.acast.com/s/ft-tech-tonic/howdeepmindvanquishedgo", "title": "How DeepMind vanquished Go", "show_or_channel": "FT Tech Tonic", "date": "2017-04-12", "duration_s": 1924, "type": "podcast", "speakers": ["Mustafa Suleyman"], "confidence": "high", "notes": "FT Tech Tonic: Suleyman on lessons from AlphaGo and applied AI.", "people": ["mustafa-suleyman"], "id": "2017-04-12_ft-tech-tonic-how-deepmind-vanquished-go", "audio_duration_s": 2027.46775, "worker": "jc-w15", "transcribe_s": 22.8, "model": "large-v3", "segments": [{"start": 4.78, "end": 35.01, "text": "Hello and welcome back to Tectonic, a podcast that looks at the way technology is changing our lives. I'm Madhumita Merjia, European Technology Correspondent at the Financial Times in London. Last week we heard from Microsoft's Lily Cheng, who talked about the evolution and challenges of chatbot technology. This week we hear from DeepMind, the company behind AlphaGo, the computer program that conquered the ancient board game of Go."}, {"start": 35.01, "end": 65.09, "text": "I think the really exciting prospect for us over the next couple of decades as machine learning systems and AI get better and better is to figure out how we can control these systems, tame them and use them to deploy at our most challenging problems. The voice of Mustafa Suleiman, co-founder of DeepMind. He came into the FT studio to talk to my colleague John Thornhill about what he learned from the AlphaGo experience."}, {"start": 65.09, "end": 95.26, "text": "and about the complex problems DeepMind has been trying to resolve since the company was acquired by Google in 2014. So thank you very much for coming into the studio, Mustafa. I wondered if we could start by talking about you. You have a very intriguing background for someone who is involved in AI. Could you tell us about what you did when you left university and how you then ended up at DeepMind? Well, I started off studying philosophy and theology at Oxford and I got to about halfway through my second year."}, {"start": 95.26, "end": 121.36, "text": "I guess I was kind of frustrated with the theoretical nature of the work that I was doing and the issues I was thinking about and kind of anxious to get on and be very practical and try to apply my ethics in the real world. And so I got together with a friend of mine who I met at university and we started a small charity, a mental health support service called Muslim Youth Helpline. It was essentially a telephone counselling service for young British Muslims."}, {"start": 121.36, "end": 150.61, "text": "And the basic idea was that we would take conventional, well-established, known techniques in counselling, so non-judgmental listening and support services, and try to make them a little bit more culturally and faith sensitive. Try to lace them in the language that other young British Muslims would find more accessible and easier to engage with. And so Muslim Youth Helpline was launched and it's still going today. It's now had its 14th year. And then you branched out into conflict resolution, is that right?"}, {"start": 150.61, "end": 178.59, "text": "Yes, essentially after three or four years or so, I realised the, in some sense, fundamental limitations of charities. It was really difficult to scale the organisation, to raise funds in a sustainable way. And from there I went on to work for the Mayor of London, Ken Livingstone. Briefly, I spent a few years there. It was pretty challenging and despite all of the high-minded principles, it was actually really difficult to get practical things done on a day-to-day basis."}, {"start": 178.59, "end": 205.92, "text": "But I learned an enormous amount about the challenges of navigating through large bureaucracies and trying to reconcile your principles in everyday practice of local government. And I guess all of those experiences culminated in me realizing that what was actually of importance in the world is to try to bring together people who speak different social languages and try to be a translator, an interlocutor at scale. And so I ended up working for a whole bunch of different organizations, including the UN, the US government."}, {"start": 205.92, "end": 225.47, "text": "Dutch government, WWF, Shell. And we worked all over the world, ended up growing the organization Rios Partners, which is still going today to about five or six offices around the world, specializing in large-scale conflict resolution and negotiation. And that culminated in 2009 when I was at the climate negotiations in Copenhagen."}, {"start": 225.47, "end": 254.83, "text": "After a series of preparations over 18, 24 months, I finally sort of cracked and realised that actually engaging with these real-world complex human organisations wasn't going to be something we could deliver on the kind of timescale that we need to given the sort of complex social challenges that we face, given the problems that we face in food and disease and climate, rapidly growing inequality, and somehow our humanly designed organisations are struggling to keep up."}, {"start": 254.83, "end": 276.62, "text": "And so how did you end up at DeepMind from there?"}, {"start": 276.62, "end": 303.9, "text": "We have to be practical, we have to engage with real institutions, and his view was always that one day we would build grand simulations of the world that would allow us to play with and model the complexities of our economy or our environment well enough to avoid that experimentation in the real world. And so clearly we came from very different perspectives, but we were always fundamentally aligned with the same purpose, to try to figure out how we could dedicate our lives to making the world a better place."}, {"start": 303.9, "end": 333.04, "text": "Corny as that might sound, that was genuinely what we've always been motivated by. And so there was a very natural alignment back in late 2009, early 2010, when I had just sort of finished the climate negotiations, which of course were at the time a massive disaster and everybody was really broken hearted and massive high promises from the Obama administration, but never quite managed to materialize. And we finally decided that now is the time to start a company building general purpose learning systems. Now we'll come back to a lot of those issues, but I'd like to touch on..."}, {"start": 333.04, "end": 348.14, "text": "The success of AlphaGo, because I think that catapulted DeepMind's global attention. And could you explain to us why AlphaGo beating the world champion of Go is so much bigger a deal than DeepBlue beating Garry Kasparov at chess in 1997?"}, {"start": 348.14, "end": 374.29, "text": "Yes, there are a bunch of reasons. In the first instance, the system that beat Kasparov in 97 was largely a rule-based, heuristics-driven system. So a whole bunch of really smart grandmasters programmed a set of handcrafted rules into the Deep Blue system. And what it basically did is look up from a table, if you like, positions that were approximately similar to the position that it faced at any given moment against Kasparov in a live game."}, {"start": 374.29, "end": 397.82, "text": "And tried to play a similar type of position based on the rule that it had been programmed with. The way that we trained AlphaGo was really to try to design the system so that it learnt its own representation of what types of moves would be rewarding or effective in any given moment. And in doing so, we taught it to learn. And that's really important, rather than programming it with structured rules."}, {"start": 397.82, "end": 405.71, "text": "We taught it to learn its own representations and it eventually learned some really interesting strategies, which you might like to call..."}, {"start": 405.71, "end": 434.43, "text": "Creative intuitions. It certainly has discovered new knowledge, which has changed thousands of years of Go wisdom and been very, very exciting to the community. The other reason I think that it's really, really different is the size of the search space. So if you roll out all the possible combinations of board positions in the game of chess, there are something like 10 to the power of 40 or so different positions. In the game of Go, there are 10 to the power of 170 different combinations of the board."}, {"start": 434.43, "end": 453.01, "text": "And so this is exponentially more positions than in chess. And in fact, many, many, many more positions in a 19 by 19 go board than there are atoms in the known universe, which is pretty incredible if you think about it. Every liquid, solid and gas in you and me in this room, you know, in the entire known universe."}, {"start": 453.01, "end": 473.14, "text": "There are exponentially more board positions that are possible in the game of Go. So this is a vastly large search space for which the traditional methods of heuristic-based rule search that were so common in old school AI simply don't work. Now Lee Sidol, the champion Go player, I think was not expecting the outcome and he appeared initially quite shattered by the result."}, {"start": 473.14, "end": 494.26, "text": "But since then, he said that he has learned a huge amount from the AlphaGo experience, and there's almost like an augmented intelligence that you can get from understanding how different systems of intelligence work. Can you talk a bit about that? That's exactly right. I mean, obviously, in the moment, he was pretty devastated. He's a professional player, one of the best in the world, been playing for his entire adult life."}, {"start": 494.26, "end": 503.68, "text": "But afterwards, we gave him the opportunity to go back through the games and look at the range of moves that AlphaGo was considering before it picked any particular move."}, {"start": 503.68, "end": 525.65, "text": "And that gave him the opportunity to really expand his own horizon and challenge his own intuitions and assumptions that had been passed down through the many generations of Go wisdom. And if you think about it, given that the size of the search space is just so large, we've only really ever explored a tiny subset of all possible moves that..."}, {"start": 525.65, "end": 534.83, "text": "in the game of Go. And so as we build systems that can help us to reframe the complex world around us and provide us with..."}, {"start": 534.83, "end": 564.11, "text": "Parallel intuitions, alternative ways of interpreting complex systems, then that again informs our own judgment of what's possible and adjusts our own biases, helps us to realize in some sense the kind of subset of the world that we're in, the filter bubble in some ways that we're actually in. And that I think has made him, as he said, much more creative. In fact, subsequently he went on to win a whole string of games back to back and he said he felt more energized than at any time before. And that might give us an intuition for how a man working"}, {"start": 564.11, "end": 579.58, "text": "Synchronously with machine could potentially do some really quite transformative things. Right. Now could you define for us what your understanding of artificial intelligence is at DeepMind? Because the term gets bandied about a lot and used in lots of very different contexts. Sure."}, {"start": 579.58, "end": 606.26, "text": "So there's probably a bunch of different definitions. The way that we think of artificial intelligence systems is that they're systems that we train in an environment to learn their own representations of rewarding behavior. So we define a reward signal. We give them an objective, whether it's to play Atari games really well, whether it's to reduce the energy required to cool the data centers, as we did in our collaboration with the Google Data Centers team."}, {"start": 606.26, "end": 610.58, "text": "Or whether it's to play the game of AlphaGo to superhuman performance."}, {"start": 610.58, "end": 640.02, "text": "But we allow the system to explore a whole bunch of different strategies that it essentially discovers largely through interacting with a data environment on its own, independent of our excessive control. So, of course, we set boundary constraints to ensure that they operate safely. But fundamentally, we're looking for the system to teach us new knowledge. And that's what we want from artificial intelligence. Traditional methods of AI are largely about scripting handcrafted rules"}, {"start": 640.02, "end": 641.22, "text": "Thank you."}, {"start": 641.22, "end": 670.61, "text": "prompting the system to reproduce those rules when they encounter a novel example of the particular task that the algorithm is trying to accomplish. So if it were classifying cups, for example, traditional methods for vision recognition, vision classification, require human intelligent programmers to say these are the broad sub-features of a cup. These are edges and shapes and lines and this is the colour and the texture and so on and so forth. And when you encounter that set of sub-features,"}, {"start": 670.61, "end": 698.64, "text": "And the big breakthrough that we've had in the last five or six years is in deep learning. These are large neural networks which learn their own representations of what are effective and valuable features in some training environment. Now, a lot of very extravagant claims are made for the impact that AI could bring about. I mean, Andrew Ng of Baidu has compared it to the new electricity in the way that it can magnify the force of different systems."}, {"start": 698.64, "end": 717.95, "text": "How much impact do you think AI is going to have? It always surprises me the extent to which people are comfortable making grand speculative statements about how the future is going to turn out. I personally feel much more cautious. Maybe it's my desire to practically make things work in the next few months or few years."}, {"start": 717.95, "end": 746.5, "text": "You know, when I engage with these systems, I see the pain of processing data and the challenge of getting our algorithms to work at scale and the regulatory environment that we need to operate in and so on and so forth. There's all sorts of practical considerations. And so, you know, I think these methods and approaches herald great potential. And that's exactly why I'm working on them. I really, truly believe that there's enormous value here and we really could do incredibly useful and socially impactful things for the world."}, {"start": 746.5, "end": 759.73, "text": "But I think at the moment our gains and our progress looks far more incremental than I think much of the hype would lead you to believe. And in what areas do you think it will have a very big social impact?"}, {"start": 759.73, "end": 782.72, "text": "Struggling to make progress with some of our most complex social problems. Over the last few centuries, capitalism has done an incredible job of inventing the idea of the limited liability corporation and defining the objective function of profit and framing problems in that regard. And that's been enormously valuable and we've driven a tremendous amount of progress."}, {"start": 782.72, "end": 804.58, "text": "Thank you. Thank you. Thank you."}, {"start": 804.58, "end": 829.38, "text": "And so I think there's a whole class of challenges which aren't easily understood in those terms. And I think the really exciting prospect for us over the next couple of decades as machine learning systems and AI get better and better is to figure out how we can control these systems, tame them, and use them to deploy at our most challenging problems. One of those that we've been working a great deal on is in healthcare. I think we are..."}, {"start": 829.38, "end": 838.38, "text": "Currently seeing the costs of healthcare spiral and the complexity and the bureaucracy of our healthcare systems completely swamp us."}, {"start": 838.38, "end": 863.63, "text": "So only 50% of care that we receive every day in the NHS or in the US systems is consistent with known best practice. So the massive variation in the quality of care is actually causing 1 in 10 avoidable experiences of harm in the UK when you go to visit any hospital. Or in the US, it's now the third biggest killer. So preventable, avoidable medical death is the third biggest killer."}, {"start": 863.63, "end": 888.13, "text": "Killing 250,000 people a year. So there's a huge opportunity. Is that a particular use of AI or is that really just smart data analytics? I think AI systems help us to model complexity. And that's what is potentially exciting about them. And one of the biggest challenges that we face in healthcare, I think, is one of diagnostics and also of coordinating timely and efficient and safe care."}, {"start": 888.13, "end": 917.42, "text": "And that, of course, means that we need to digitise the patient experience in hospital, whether it is diagnosing particular conditions in radiology scans or in pathology, or in fact tracking the course of interventions that have happened to you over the 24 hours that you've been in admission. And it's that kind of information that help us to learn the kind of structure so that we can better predict the likely trajectory that you're on. The big challenge in acute hospitals today is that the vast majority of care..."}, {"start": 917.42, "end": 945.1, "text": "97% of care is actually reactive. So nurses and doctors are largely doing things to you, responding to you, only after you have gone into some kind of decline or something has gone wrong. And what we all know, what all the literature tells us, is that the more we can be proactive, predict when somebody is likely to suffer a particular reaction or to deteriorate in some sense, the more likely we are to be able to deliver cheaper, safer and more effective treatments. And that's exactly what machine learning systems do well."}, {"start": 945.1, "end": 973.15, "text": "Okay, now there's been some controversy about the use of health data by private companies generally, but particularly in the National Health Service. How can you ensure as a private company that everyone has equal, fair access to that data? Investing with Schwab is like spending a Saturday at a great farmer's market. You can fill your reusable tote with a bit of everything. Maybe you go for some free-range, self-directed investing."}, {"start": 973.15, "end": 999.78, "text": "Or perhaps you pick up a few farm fresh trades while you peruse. You can even get help from a dedicated advisor. That's full service wealth management. Mix, match, and change your mind whenever you want. Because at Schwab, you can invest your way. No matter your goals or appetite for investing, Schwab has everything you need, all in one place. Visit Schwab.com to learn more. There is a not abusive privacy."}, {"start": 999.78, "end": 1017.55, "text": "Yeah, I think that's a great question. So we've worked very hard to think about this question from the outset. So we launched DeepMind Health in February of last year. On the day that we launched, we announced that we'd be setting up a panel of independent reviewers, so nine independent policy and clinical experts."}, {"start": 1017.55, "end": 1046.83, "text": "to scrutinize us in the public interest. And so this is the first time, as far as we know, that a UK-based technology company has invited this kind of scrutiny and oversight proactively. We know that we need to be as innovative on the governance and the oversight and the accountability that we invite upon ourselves as we are in the development of the core technology. And so what we've asked this panel to do is to come and spend time with us quarterly. They can interview anybody in my team. They haven't signed any contract. They're not paid."}, {"start": 1046.83, "end": 1064.3, "text": "We haven't even signed an NDA, so they can say whatever they like about us publicly. They have access to all of our agreements. They've audited our technical infrastructure. We've in fact given them a budget to hire a secretariat, to commission consultants, to do a technical audit, to do a legal and governance audit."}, {"start": 1064.3, "end": 1085.97, "text": "And we've been very open with them. I think success is a function of our ability to prove utility. But as importantly, we need to be able to build trust. And building trust, the first step, in my opinion, is for us to be trusting of our panel of independent reviewers to scrutinize and oversee us. And just to be clear, the use of that data is not exclusive to you."}, {"start": 1085.97, "end": 1104.3, "text": "But other researchers or people who want access to that data would have to go via your platform, is that right? No, they don't have to come via our platform at all. So there are many, many, many research relationships with various hospitals and academic universities across the UK between companies and researchers and NHS organisations."}, {"start": 1104.3, "end": 1133.34, "text": "And so we have a relationship with the Moorfields Eye Hospital, just like any other organisation can have a research relationship. We don't own any of that data. We simply process the data and we publish the results of our work openly in peer-reviewed journals. So what the Moorfields Hospital are really excited about is the potential for us to diagnose instantly and hopefully near or at human performance really difficult conditions like glaucoma, age-related macular degeneration, diabetic retinopathy."}, {"start": 1133.34, "end": 1162.32, "text": "So the interesting thing is that 99% of radiologists across the UK in a recent study said that they were overwhelmed by their workload. And so this is exactly the kind of opportunity, just like we did with AlphaGo, where potentially we could provide an intelligent assistant to the human radiographers to help them to speed up their diagnoses, review scans that have been made by other colleagues, and be given a standardized best practice tool that helped them to..."}, {"start": 1162.32, "end": 1177.9, "text": "consistently deliver the highest quality possible diagnosis that they can. Okay, now you've been describing this extraordinary range of possibilities about how AI can be applied. How do you go about determining what your priority should be? Because you are also part of a very large..."}, {"start": 1177.9, "end": 1196.46, "text": "We were in a very privileged position when we were acquired. We had an enormous amount of cash in the bank for a very long time, and so we were under no pressure to join Google at the time before Alphabet was announced."}, {"start": 1196.46, "end": 1213.2, "text": "And so that put us in a great position to be able to essentially guarantee our operating autonomy and our independence. And so our research and our product roadmap is entirely defined by us. And so within that, as this independent separate entity within Alphabet."}, {"start": 1213.2, "end": 1230.88, "text": "We've obviously also chosen to collaborate with Google. There are some really interesting machine learning problems across Google, and I think there are some great opportunities to have a really significant impact, both commercially and also in terms of testing and developing our algorithms on some really interesting data sets."}, {"start": 1230.88, "end": 1246.5, "text": "Two broad divisions that I run in our applied group. I run all of our commercial applications, one of which is DeepMind for Google. And so we have a pretty large team collaborating with various different groups and products across Google, one of which was our data centers and energy work."}, {"start": 1246.5, "end": 1271.74, "text": "Precisely picked this because it was a problem that kind of sits at the intersection of our values and our desire to generate revenue. So data centers consume an enormous amount of energy and growing vastly over the next five to ten years. And the real challenge in data centers is to be able to efficiently extract heat from these enormous physical utilities. And it turns out that we were able to build a..."}, {"start": 1271.74, "end": 1301.07, "text": "you know really intelligent control system that has a model of the incoming demand for compute so when you write a query into youtube and you stream a video essentially what that's doing is spinning up cycles in a data center somewhere very far away and that's generating heat and when the physical apparatus of the data center infrastructure responds to that heat extraction in coolers and chillers and set points of temperature valves and so on and so forth we actually are able to control"}, {"start": 1301.07, "end": 1318.77, "text": "And you're talking with National Grid. Is that right about how you can apply that on a far bigger scale?"}, {"start": 1318.77, "end": 1332.02, "text": "Develop at DeepMind are inherently general. You know, one of the defining characteristics of the machine learning systems that we build are that, as I said, they should be able to learn their own representations of rewarding behavior, and also they should be general."}, {"start": 1332.02, "end": 1354.42, "text": "The same suite of methods should be able to work in multiple different environments and hopefully get better as they encounter new environments. And so we're very excited about the prospect of being able to use some of these methods to better manage some of the challenges that the grid has. The grid is obviously very keen to onboard more renewable energy, but the challenge with renewables is that they're very variable."}, {"start": 1354.42, "end": 1383.7, "text": "So some days it can be sunny or windy and other days it can be not at all. And so the difficulty is given that demand for energy remains very consistent, how do we supply the grid with more and more renewables? And actually there's a significant component of that which is an intelligent prediction problem. If we know with very high certainty how much energy is likely to be consumed and how much energy we're likely to produce, then clearly there's an opportunity for us to make that matching process a little bit more..."}, {"start": 1383.7, "end": 1407.22, "text": "Now, as we know from the Superman comics, with great power comes great responsibility."}, {"start": 1407.22, "end": 1433.14, "text": "And at the time of the acquisition by Google, you talked about setting up an ethics committee and oversight of a lot of your activities. How have you progressed on that? Where are we with that? We do have an ethics board that we've set up. We've not yet announced it publicly. And we've been working really hard internally to operationalize the principles of our ethics board. So as I mentioned, we were in a really strong position when we were acquired."}, {"start": 1433.14, "end": 1446.61, "text": "gave us the opportunity to put our values front and center as part of the acquisition. And so one of the things that we were able to agree is that we would have an ethics and safety board staffed by the DeepMind founders as well as three representatives."}, {"start": 1446.61, "end": 1475.89, "text": "from Alphabet along with some independent representation to help us to navigate some of the complexities of developing technology like this. And one of the things that we've always been interested in is figuring out how we can be innovative and proactive on holding ourselves accountable. And most of these things will be ongoing experiments. That's the way I like to think about them. It's really important to avoid some of the kind of high-minded speculation and hype that you get around AI as well as the grand."}, {"start": 1475.89, "end": 1493.7, "text": "Fear-mongering that we've seen in the last few years around superintelligence and so on and the singularity. We're much more practical and engaged in the day-to-day battle of making governance and transparency and accountability real. And there's lots of different ways that we do that. Some are new mechanisms of human governance and oversight."}, {"start": 1493.7, "end": 1523.02, "text": "Others are technical mechanisms. So one of the things that we have been working really hard on is a set of technical methods that we call verifiable data audit. So given some particular data set that we have access to, say for example in our collaborations with the NHS, we process data on behalf of some of the NHS hospitals. And obviously there are very strict governance, regulatory and legal processes around that. But most people across the system,"}, {"start": 1523.02, "end": 1539.95, "text": "struggle to interpret those very long contracts and those very complex pieces of regulation and apply them to the practical everyday relationship that we have with the hospital. And so what we've designed is a method for verifying that everything that we have done with the data..."}, {"start": 1539.95, "end": 1569.23, "text": "In our capacity as data processors is only what we have been allowed to do with that data. So when we move data from one database to another, or when a member of staff accesses the register of data, or when there is an edit or a modification to that data, all of those interventions into the data structure on our servers generates an untamperable log that describes that activity. So what we're starting to do is to shine a light of transparency on"}, {"start": 1569.23, "end": 1581.23, "text": "At the moment, I think many people feel distrustful and frustrated because, you know, we're creating all of this data through our interactions with these digital systems."}, {"start": 1581.23, "end": 1609.23, "text": "And it just sort of sits off in the cloud somewhere in some black box and you or I have no oversight or no ability to control or influence what happens to that data. And so I think one first step that will help in that direction is to at least create cryptographic certainty around who has done what with your data at what time and to verify that that is consistent with the legal processing requirements. And how can you ensure cryptographic certainty? It sounds like blockchain."}, {"start": 1609.23, "end": 1638.67, "text": "Yes, it is very much inspired by blockchain, but with some modifications on the blockchain system that ensure that we don't face the sort of proof of work challenge that you get when it's a distributed ledger. So ours is a sort of private ledger that will be run by us, but that can be scrutinized by independent technical third parties, by our independent reviewers, by other governance processes, namely regulators or data controllers or hospitals or anyone we collaborate with. And the aim is to make this..."}, {"start": 1638.67, "end": 1667.2, "text": "As you say, it's quite striking how open Google DeepMind has been in terms of publishing research, the transparency of what you're doing. That in itself, though, has led some people to worry about the extent to which some of these potentially incredibly powerful technologies are open. What do you think of that debate? You know, I think we try to remain as practical as possible. I think there is, at the moment, no risk."}, {"start": 1667.2, "end": 1680.1, "text": "that we are being too open. If anything, I think companies in general need to be much more open around how they process and manage data. We need to publish much more of our algorithms and our work."}, {"start": 1680.1, "end": 1709.49, "text": "Thank you very much."}, {"start": 1709.49, "end": 1738.43, "text": "We are to try to sort of rebuild some of that trust that is obviously driving so much of the populism that we see today. Right. And one of the ways of rebuilding that trust clearly is to have more of open political debate about artificial intelligence. I was looking at a report the other day by Future Advocacy, which noted that the term artificial intelligence had only been mentioned in the House of Commons 32 times since electronic records had begun compared to the hundreds of thousands of mentions of Brexit."}, {"start": 1738.43, "end": 1747.18, "text": "That seems a bit of a strange priority to me. But how can you better inform politicians about what this technology is and what kind of impact it can bring about?"}, {"start": 1747.18, "end": 1776.46, "text": "Yeah, I mean, that's a great question. I mean, I think we think very deeply about that. So we spend a lot of time trying to publish as much information as possible on our websites. We have a whole bunch of Q&As that explain the data that we have, the algorithms that we're trying to develop, the products that we're trying to build. We've given many, many public presentations about the long-term product roadmap that we see and the potential for this to be really transformative. We also spend time with think tanks, with academics, with some politicians too."}, {"start": 1776.46, "end": 1799.31, "text": "Trying to explain more on the kind of benefits of our technologies, but also some of the risks. I think we're very sensitive and honest about the potential for misuse. I mean, it's naive to believe that these technologies are only and inevitably going to be beneficial. I think the best framing for these systems are they should be technology systems that we control and that we direct."}, {"start": 1799.31, "end": 1828.59, "text": "As a society, as a population, collectively, not just a particular company or a particular government, we have to be able to control these systems so that they do what we want when we want them and that they don't sort of run away ahead of us. And I think that's where some of the frustration with some of the technology companies comes from because I think people are concerned that the rate of progress is outstripping our ability to really understand these systems and be able to control them. And I think that's probably well reflected in the fact that there aren't very many parliamentary mentions."}, {"start": 1828.59, "end": 1853.71, "text": "Final question, but it's quite a big one. Do you think you will see human-level machine intelligence in your lifetime?"}, {"start": 1853.71, "end": 1883.02, "text": "Practical, real-world, applied challenges. I think that we have enormous growing challenges ahead of us from modelling our climate to managing antimicrobial resistance to improving the efficiency of our healthcare systems. You know, 800 million people in the world today are malnourished and yet a third of the food that we produce every year goes wasted. One of the biggest problems we have is that of obesity and the consequential diabetes that arises from that. So we have a fundamental problem with production."}, {"start": 1883.02, "end": 1912.82, "text": "We must end it there. But thank you very much, Mr. Stafford, for a fascinating conversation. Thank you very much, Sean. Cheers."}, {"start": 1915.5, "end": 1942.7, "text": "We'll be back next week with another episode of Tectonic when we hear from Jeremy Conrad of Lemnos Labs, who is making a bet that self-driving cars will be a boom for the robotic startups backed by his incubator-turned-venture capital fund. If you'd like to comment on today's show or suggest any topics you'd like us to cover in future episodes, please email us at tectonic at ft.com."}, {"start": 1943.31, "end": 1985.06, "text": "This episode of Tectonic was produced by Fiona Simon. For years, credit scores were generated with just a snapshot. A snapshot that left millions of people out of the picture. So VantageScore is modernizing the credit score. VantageScore uses 400% more credit report and alternative data than legacy models, bringing more people into the credit economy. Like 5 million more mortgage eligible people. Banks win and francs win. Lenders win and so do lenders."}, {"start": 1985.06, "end": 2013.2, "text": "When you build credit scores the right way, everybody wins. Get started at VantageScore.com. The new episode of the next five podcast is all about sovereignty, data, and control. Charlie Giancarlo at Everpure joins me. You always get a very embarrassed response if you ask them if they know where all their data is. As does Professor Sana Karaghani. There isn't any modern AI system that's entirely sovereign. And Claire Schramm-Fergus at Lloyds Banking Group."}, {"start": 2013.2, "end": 2020.98, "text": "Listen to The Next Five wherever you get your podcasts. Enjoy."}]}