{"version":"1.0.0","segments":[{"startTime":48.225,"endTime":50.965,"body":"So why are some people concerned, right?"},{"startTime":62.985,"endTime":67.645,"body":"So imagine I had like a general purpose robot, and I'm sitting in faculty's offices."},{"startTime":68.295,"endTime":71.51,"body":"We've just moved in and we still haven't managed to get our coffee machine plumbed"},{"startTime":71.51,"endTime":71.725,"body":"in."},{"startTime":88.555,"endTime":90.925,"body":"Alright, okay, bloody hell, robot."},{"startTime":97.515,"endTime":101.831,"body":"Just open the door, go through it as the robot steps on the cat, okay,"},{"startTime":101.831,"endTime":105.285,"body":"next day, robot, don't smash the door, don't step on the cat."},{"startTime":105.785,"endTime":107.245,"body":"Get me a coffee as fast as you can."},{"startTime":107.705,"endTime":112.895,"body":"So it runs downstairs and just steals a coffee from the hand of the person"},{"startTime":112.895,"endTime":113.241,"body":"who's just get got served their coffee, okay, robot, don't smash the door, don't step"},{"startTime":113.241,"endTime":119.125,"body":"on the coffee."},{"startTime":127.545,"endTime":132.673,"body":"And so the robot races in, starts making the coffee, comes back totally burned, like,"},{"startTime":132.673,"endTime":134.725,"body":"you know, half, half the robot."},{"startTime":134.725,"endTime":139.245,"body":"It was, okay, robot, don't, don't smash the door."},{"startTime":140.225,"endTime":140.965,"body":"Why don't do this?"},{"startTime":150.385,"endTime":158.24,"body":"It sounds small, but hopefully it illustrates the actual difficulty of specifying a precise objective"},{"startTime":158.24,"endTime":164.525,"body":"function for something that doesn't have your common sense like humans are."},{"startTime":164.835,"endTime":169.692,"body":"When I ask somebody to get me a coffee, there's an enormous amount of sort"},{"startTime":169.692,"endTime":170.016,"body":"of common sense that goes alongside that, that we simply don't know how to put"},{"startTime":170.016,"endTime":175.845,"body":"into computers right now."},{"startTime":176.545,"endTime":182.024,"body":"And so in the robot coffee example, maybe that's a little bit silly, but what"},{"startTime":182.024,"endTime":182.389,"body":"happens if we start making very powerful ais and we start putting them to work"},{"startTime":182.389,"endTime":188.965,"body":"on really important problems?"},{"startTime":193.985,"endTime":201.35,"body":"And we know from mythology, humans have known for thousands of years that actually this"},{"startTime":201.35,"endTime":203.805,"body":"is a hard problem, right?"},{"startTime":203.995,"endTime":206.445,"body":"King Midas wanted to be wealthy."},{"startTime":215.245,"endTime":221.02,"body":"He found out the difficulty of specifying an objective function to be what you actually"},{"startTime":221.02,"endTime":222.945,"body":"want, not what you say."},{"startTime":243.685,"endTime":245.385,"body":"So this is Sam Altman from Open ai."},{"startTime":246.385,"endTime":251.273,"body":"I mean, I think the best case is like so unbelievably good that it's like"},{"startTime":251.273,"endTime":255.185,"body":"hard to, I think it's like hard for me to even imagine."},{"startTime":255.255,"endTime":260.059,"body":"Like I can sort of, I can sort of think about what it's like when"},{"startTime":260.059,"endTime":260.379,"body":"we make more progress of discovering new, new knowledge with these systems than humanity has"},{"startTime":260.379,"endTime":265.825,"body":"done so far."},{"startTime":292.965,"endTime":297.835,"body":"But I can't quite like, I think the, the, the good case is just so"},{"startTime":297.835,"endTime":302.705,"body":"unbelievably good that you sound like a really crazy person to start talking about it."},{"startTime":309.725,"endTime":314.842,"body":"Um, I'm more worried about like an accidental misuse case in the short term where,"},{"startTime":314.842,"endTime":315.183,"body":"you know, someone gets a super powerful, like, it's not like the AI wakes up"},{"startTime":315.183,"endTime":321.665,"body":"and decides to be evil."},{"startTime":332.165,"endTime":337.625,"body":"But, but I can see the accidental misuse case clearly, and that's, that's super bad."},{"startTime":337.725,"endTime":337.945,"body":"Um,"},{"startTime":338.055,"endTime":342.637,"body":"Okay, so this is, that's Sam Altman from OpenAI talking about, uh, this, and then"},{"startTime":342.637,"endTime":342.943,"body":"this is Dario Moe from a, he was actually the person, the scientist probably most"},{"startTime":342.943,"endTime":343.248,"body":"prominently associated with building gp, PT three and open ai and then moved to start"},{"startTime":343.248,"endTime":356.385,"body":"his own startup anthropic, which is maybe the third biggest, um, of the kind of"},{"startTime":375.475,"endTime":379.001,"body":"Yeah, I, I I think it's popular to give these percentage numbers and, and you"},{"startTime":379.001,"endTime":379.237,"body":"know, I mean the truth is that I'm, I'm not, I'm not sure it's easy"},{"startTime":379.237,"endTime":384.645,"body":"to put to, to put a number to it."},{"startTime":387.625,"endTime":392.863,"body":"Um, you know, I I think I've, I think I've often said that, you know,"},{"startTime":392.863,"endTime":393.212,"body":"my, my chance that something goes, you know, really quite catastrophically wrong on the scale"},{"startTime":393.212,"endTime":393.561,"body":"of, of, you know, human civilization, you know, it might be somewhere between 10 and"},{"startTime":393.561,"endTime":393.91,"body":"25% when you put together the risk of something going wrong with the model itself,"},{"startTime":393.91,"endTime":394.26,"body":"with, you know, something going wrong with human, you know, people or organizations or nation"},{"startTime":394.26,"endTime":394.609,"body":"states misusing the model or, or it kind of inducing conflict among them or, or"},{"startTime":394.609,"endTime":423.245,"body":"just some way in which kind of society can't, can't handle it."},{"startTime":437.665,"endTime":442.772,"body":"Um, again, this stuff about curing cancer, I think if, if we can avoid the"},{"startTime":442.772,"endTime":443.113,"body":"downsides, then this stuff about, you know, about curing cancer, extending the human lifespan, um,"},{"startTime":443.113,"endTime":450.605,"body":"you know, solving problems like, like mental illness."},{"startTime":462.545,"endTime":466.814,"body":"And I think one of the big motivators for reducing that 10 to 25% chance"},{"startTime":466.814,"endTime":467.098,"body":"is, you know, how how great it'll \u003claugh\u003e, you know, is trying to increase, is"},{"startTime":467.098,"endTime":473.645,"body":"trying to increase the good part of the pie."},{"startTime":493.115,"endTime":494.245,"body":"It's great to be part of it."},{"startTime":494.275,"endTime":497.725,"body":"It's, you know, it's, it's great to be one of the ones building it and"},{"startTime":497.725,"endTime":500.485,"body":"causing it to happen, but there, there's a certain robustness to it."},{"startTime":500.905,"endTime":505.558,"body":"And, you know, I find I find more meaning, I find more, you know, when,"},{"startTime":505.558,"endTime":505.869,"body":"when this is all over, I think, you know, I personally will feel I've done"},{"startTime":505.869,"endTime":513.005,"body":"more to contribute to, you know, whatever utopia results."},{"startTime":527.825,"endTime":532.831,"body":"The lesson that I wanted to take from this is one powerful AI is probably"},{"startTime":532.831,"endTime":533.165,"body":"coming."},{"startTime":534.025,"endTime":541.172,"body":"And two, once you put AI to important decisions, those decisions intrinsically contain a notion"},{"startTime":541.172,"endTime":542.125,"body":"of ethics."},{"startTime":547.505,"endTime":552.085,"body":"And so this discussion about how we use this technology has to involve everyone."},{"startTime":552.865,"endTime":557.085,"body":"And so that's what I'm hoping my talk is gonna lay the foundations for today."},{"startTime":593.745,"endTime":594.965,"body":"And then where should we be?"},{"startTime":597.425,"endTime":602.405,"body":"AI over its history has really had these two branches, one the kind of good"},{"startTime":602.405,"endTime":605.725,"body":"old fashioned, what we now call good old fashioned ai."},{"startTime":646.585,"endTime":649.685,"body":"Now, learning things from data may sound extremely mysterious."},{"startTime":663.465,"endTime":668.325,"body":"We want to separate cats from dogs and we have a couple of measurements about"},{"startTime":668.325,"endTime":671.565,"body":"them, maybe their weight and their tooth length or whatever."},{"startTime":685.825,"endTime":687.205,"body":"But a computer doesn't know that."},{"startTime":710.625,"endTime":713.254,"body":"So all the dogs are on one side and all the cats are on the"},{"startTime":713.254,"endTime":713.43,"body":"other."},{"startTime":717.145,"endTime":721.195,"body":"And it can see that like, okay, I've got sort of 100% dogs on one"},{"startTime":721.195,"endTime":724.165,"body":"side, but basically 50 50 dogs and cats on the other."},{"startTime":725.025,"endTime":726.565,"body":"And then it makes a new guess."},{"startTime":727.105,"endTime":731.973,"body":"Now it's got a hundred percent cats on one side, two thirds dogs and one"},{"startTime":731.973,"endTime":734.245,"body":"third cat on the other, slightly better."},{"startTime":742.425,"endTime":746.768,"body":"And so if a new data point comes in like this, a triangle like this,"},{"startTime":746.768,"endTime":749.085,"body":"the computer can just look at the boundary."},{"startTime":752.705,"endTime":753.445,"body":"And make a guess."},{"startTime":777.305,"endTime":781.388,"body":"So there's a limited amount that I can fit to, and this is where neural"},{"startTime":781.388,"endTime":782.205,"body":"networks come in."},{"startTime":785.595,"endTime":792.401,"body":"They have non-linear, they have the ability to do non-linear fitting, uh, and they have"},{"startTime":792.401,"endTime":796.485,"body":"millions, billions, or even trillions of parameters these days."},{"startTime":830.705,"endTime":836.811,"body":"And so neural networks let us take in thousands of inputs, give out outputs into"},{"startTime":836.811,"endTime":837.218,"body":"thousands of categories and do this kind of fitting, but in not just two dimensional"},{"startTime":837.218,"endTime":843.325,"body":"space."},{"startTime":848.605,"endTime":853.261,"body":"So you won't be able to visualize it, but sort of keep in your mind"},{"startTime":853.261,"endTime":856.365,"body":"the picture of three-dimensional space and volumes in three-dimensional space."},{"startTime":860.745,"endTime":864.325,"body":"And then if the new point falls inside that making that your prediction."},{"startTime":871.265,"endTime":874.67,"body":"So I'm sure basically everyone can people just sort of nod if they've, if they"},{"startTime":874.67,"endTime":875.805,"body":"know what, yeah, okay, good."},{"startTime":901.065,"endTime":905.85,"body":"So a dog has four and we show this to chat GPT and we say,"},{"startTime":905.85,"endTime":907.445,"body":"please predict the next word."},{"startTime":910.805,"endTime":915.173,"body":"And then we compare the two things and we say, uh, the 'cause we know"},{"startTime":915.173,"endTime":915.465,"body":"the real sentence and we know its prediction and we say, cha, GPT, you got"},{"startTime":915.465,"endTime":920.125,"body":"that wrong."},{"startTime":920.985,"endTime":926.375,"body":"Now adjust those boundaries, the kind of the equivalent of that line, but in this"},{"startTime":926.375,"endTime":926.734,"body":"high dimensional space of all words, adjust those boundaries such that you predict legs next"},{"startTime":926.734,"endTime":932.125,"body":"time."},{"startTime":946.435,"endTime":952.943,"body":"Okay, so you're now basically at the cutting edge of AI in terms of how"},{"startTime":952.943,"endTime":954.245,"body":"these things work."},{"startTime":1013.105,"endTime":1015.925,"body":"How do we like conceptualize these on a single axis?"},{"startTime":1049.425,"endTime":1052.045,"body":"And then perhaps chat GPT sits somewhere in the middle."},{"startTime":1072.995,"endTime":1075.085,"body":"Like kind of like adding two numbers together."},{"startTime":1086.465,"endTime":1091.343,"body":"And so this here is the compute used on some of the famous algorithms for"},{"startTime":1091.343,"endTime":1092.645,"body":"the last 50 years."},{"startTime":1097.785,"endTime":1103.847,"body":"And uh, this point outlier at the top there is, um, is GPT-4 that's gone"},{"startTime":1103.847,"endTime":1107.485,"body":"up 10 million times in the last 10 years,"},{"startTime":1111.105,"endTime":1115.86,"body":"But to show the next year or two, I have to shrink all of those"},{"startTime":1115.86,"endTime":1117.445,"body":"data points down to this."},{"startTime":1180.065,"endTime":1186.365,"body":"And so if we estimate in the 18 years, the first 18 years of life,"},{"startTime":1186.365,"endTime":1192.245,"body":"roughly how many sort of flop equivalents do we think the human brain uses?"},{"startTime":1192.265,"endTime":1196.765,"body":"So effectively the training time up to 18, where do we think it falls on"},{"startTime":1196.765,"endTime":1197.365,"body":"this plot?"},{"startTime":1242.565,"endTime":1247.185,"body":"Of course, big caveats on whether we can actually do these calculations effectively."},{"startTime":1249.025,"endTime":1253.765,"body":"Now, something that will run through many people's heads is aren't human beings somehow special?"},{"startTime":1253.865,"endTime":1258.83,"body":"Is this not too me mechanistic reductionist, uh, to try and compare the compute like"},{"startTime":1258.83,"endTime":1260.485,"body":"in this really obvious fashion?"},{"startTime":1263.185,"endTime":1268.96,"body":"And in fact I claim that that kind of magical thinking has actually misled science"},{"startTime":1268.96,"endTime":1270.885,"body":"in a bunch of circumstances."},{"startTime":1285.785,"endTime":1290.669,"body":"We now know that in our galaxy, well, we're not even the center of our"},{"startTime":1290.669,"endTime":1293.925,"body":"solar system, nevermind the center of our galaxy in biology."},{"startTime":1296.505,"endTime":1298.725,"body":"We were given a very, very special place."},{"startTime":1299.625,"endTime":1306.405,"body":"And nowadays we think we evolved like all the other animals in quantum mechanics."},{"startTime":1307.105,"endTime":1313.442,"body":"Uh, for a decent period of time, probably a couple of decades, people genuinely wondered"},{"startTime":1313.442,"endTime":1317.245,"body":"whether it required consciousness to decohere a wave function."},{"startTime":1378.905,"endTime":1382.223,"body":"Uh, I'm happy to ask answer questions about it at the end if people are"},{"startTime":1382.223,"endTime":1382.445,"body":"interested."},{"startTime":1383.585,"endTime":1390.169,"body":"Um, but that does then like genuinely provoke the question, what is in the top"},{"startTime":1390.169,"endTime":1391.925,"body":"right of this plot?"},{"startTime":1392.825,"endTime":1397.125,"body":"And so this is where people will start talking about super intelligences."},{"startTime":1406.225,"endTime":1412.725,"body":"Now, people's default notion immediately when thinking about this is to sort of anthropomorphize it"},{"startTime":1412.725,"endTime":1417.925,"body":"as like either a god or a devil and nothing in between."},{"startTime":1437.045,"endTime":1442.085,"body":"I think it's better to think of a super intelligence as kind of like, uh,"},{"startTime":1442.085,"endTime":1442.421,"body":"this, uh, sort of disembodied idea of it has some goals, whatever, it's whatever we've"},{"startTime":1442.421,"endTime":1442.757,"body":"either programmed it with or it's learned and it can make moves in the, uh,"},{"startTime":1442.757,"endTime":1443.093,"body":"in the real world with whatever actions it has available to it and it has"},{"startTime":1443.093,"endTime":1458.885,"body":"an understanding of the world."},{"startTime":1469.635,"endTime":1475.766,"body":"Whereas Sam said, and Dario would've said, but didn't quite, uh, uh, it's actually these"},{"startTime":1475.766,"endTime":1479.445,"body":"accidental use ca misuse cases that are more concerning."},{"startTime":1484.195,"endTime":1486.925,"body":"Like a totally crucial question here."},{"startTime":1490.545,"endTime":1496.942,"body":"So, uh, nobody has built one as far as we know and, uh, and so"},{"startTime":1496.942,"endTime":1497.368,"body":"we can't say for certain it's possible, but we can say that there is no"},{"startTime":1497.368,"endTime":1506.325,"body":"known theory of science that prevents it."},{"startTime":1517.445,"endTime":1522.466,"body":"They can send signals faster, they can store signals for longer times, they can retrieve"},{"startTime":1522.466,"endTime":1523.805,"body":"things with higher veracity."},{"startTime":1524.465,"endTime":1529.747,"body":"And so I think you have to, like the extraordinary claim that demands extraordinary evidence"},{"startTime":1529.747,"endTime":1534.325,"body":"is positing that a super intelligence can't be built rather that it can."},{"startTime":1536.585,"endTime":1538.365,"body":"So what could we do about ai?"},{"startTime":1557.305,"endTime":1559.965,"body":"So let's take a slightly hypothetical scenario."},{"startTime":1568.345,"endTime":1574.436,"body":"And so we set it going and it looks around at the world and it"},{"startTime":1574.436,"endTime":1577.685,"body":"says, okay, uh, I'm slightly anthropomorphizing it here."},{"startTime":1585.725,"endTime":1590.41,"body":"And it recognizes that there are actors out there whose livelihoods to depend on the"},{"startTime":1590.41,"endTime":1592.285,"body":"production of carbon in the atmosphere."},{"startTime":1626.505,"endTime":1630.855,"body":"Now, that's not to say there aren't ideas about how you create an AI with"},{"startTime":1630.855,"endTime":1631.725,"body":"an off switch."},{"startTime":1631.865,"endTime":1638.431,"body":"So a professor called Stuart Russell at Berkeley, um, has a set of ideas where"},{"startTime":1638.431,"endTime":1643.685,"body":"you insist that the intelligence, uh, maintains great uncertainty about human preferences."},{"startTime":1664.795,"endTime":1669.283,"body":"It's actually not obvious that if you just like let things happen and train the"},{"startTime":1669.283,"endTime":1669.583,"body":"most powerful agents you could, there would be, uh, there would be an obvious off"},{"startTime":1669.583,"endTime":1676.765,"body":"switch in the way you might imagine to begin with."},{"startTime":1691.745,"endTime":1697.285,"body":"Uh, however, however, in this green area, there's a lot of safe region."},{"startTime":1697.985,"endTime":1704.355,"body":"So in the top right there is, you know, a whole set of, uh, uh,"},{"startTime":1704.355,"endTime":1710.725,"body":"potential technologies that we can't currently understand and can't currently control in the green region."},{"startTime":1711.745,"endTime":1717.158,"body":"We do understand them and we can control them, like artificial intelligence has been used"},{"startTime":1717.158,"endTime":1720.045,"body":"totally safely for decades in the real world."},{"startTime":1745.255,"endTime":1750.325,"body":"It'll build great technology companies that will actually meaningfully change the way we live."},{"startTime":1787.345,"endTime":1790.285,"body":"And powerful AI can be a path to solving these."},{"startTime":1798.745,"endTime":1803.445,"body":"But critics will say, you know, it is unknown and it's not obviously safe."},{"startTime":1804.225,"endTime":1807.685,"body":"And the dynamics of building it are genuinely complicated."},{"startTime":1808.595,"endTime":1816.205,"body":"Individuals, nations, companies, they're all have reasons to race a bit more than they'd like."},{"startTime":1830.255,"endTime":1833.311,"body":"But because we want to encourage a love of learning, we think it's well worth"},{"startTime":1833.311,"endTime":1833.515,"body":"it."},{"startTime":1834.415,"endTime":1837.795,"body":"We never make you pay for lectures, although donations are needed."},{"startTime":1834.415,"endTime":1837.795,"body":"We never make you pay for lectures, although donations are needed."},{"startTime":1842.815,"endTime":1846.731,"body":"And if you haven't already, click the follow or subscribe button from wherever you are"},{"startTime":1846.731,"endTime":1847.515,"body":"listening right now."},{"startTime":1842.815,"endTime":1846.731,"body":"And if you haven't already, click the follow or subscribe button from wherever you are"},{"startTime":1846.731,"endTime":1847.515,"body":"listening right now."},{"startTime":1866.555,"endTime":1874.55,"body":"Is it really realistic that we'll regulate a hypothetical unknown technology effectively in a way"},{"startTime":1874.55,"endTime":1878.815,"body":"that actually ends up, uh, making things safer?"},{"startTime":1879.815,"endTime":1886.265,"body":"A poor regulation could easily make things worse if it privileges bad actors over good"},{"startTime":1886.265,"endTime":1886.695,"body":"actors."},{"startTime":1906.475,"endTime":1911.914,"body":"Um, because I think there are two fundamental crucial questions that underlie this, that enable"},{"startTime":1911.914,"endTime":1914.815,"body":"us to have a bit more thoughtful discussion."},{"startTime":1922.155,"endTime":1923.215,"body":"Do we have a short time?"},{"startTime":1924.035,"endTime":1929.039,"body":"Now, five, 10 years ago, the long time you would've found people who said, we"},{"startTime":1929.039,"endTime":1931.375,"body":"probably have like a century or two."},{"startTime":1939.995,"endTime":1943.522,"body":"And now the people who say the long time are like, you know, a few"},{"startTime":1943.522,"endTime":1946.815,"body":"decades and the people who say a short time say two to five years."},{"startTime":1949.315,"endTime":1953.911,"body":"Uh, and then how easy would it be to what we call align super intelligence,"},{"startTime":1953.911,"endTime":1956.975,"body":"but basically get a super intelligence to care about us?"},{"startTime":1979.525,"endTime":1982.945,"body":"I think we have a long time and it's gonna be really easy to get"},{"startTime":1982.945,"endTime":1984.085,"body":"it to care about us."},{"startTime":1985.645,"endTime":1988.972,"body":"I think it's gonna be an extremely hard problem to get it to care about"},{"startTime":1988.972,"endTime":1990.525,"body":"us, to align it to our values."},{"startTime":2020.335,"endTime":2021.765,"body":"We've got ages to figure it out."},{"startTime":2021.765,"endTime":2025.79,"body":"The technologies that you work on to try and make things safe right now aren't"},{"startTime":2025.79,"endTime":2028.205,"body":"actually gonna be the technologies that it's built with."},{"startTime":2092.745,"endTime":2096.245,"body":"Nobody knows where the actual world really sits."},{"startTime":2102.025,"endTime":2102.845,"body":"Or is it gonna be easy?"},{"startTime":2108.065,"endTime":2113.716,"body":"Um, actually I think probably the right attitude is to kind of fill in the,"},{"startTime":2113.716,"endTime":2114.092,"body":"all the regulations on this two by two and then say, how do we, uh,"},{"startTime":2114.092,"endTime":2114.469,"body":"uh, like mitigate the worst downsides and um, uh, like get, take advantage of the"},{"startTime":2114.469,"endTime":2126.525,"body":"most of the upsides."},{"startTime":2143.065,"endTime":2149.27,"body":"And so what would a better, a better version be was perhaps two paths that"},{"startTime":2149.27,"endTime":2150.925,"body":"strike me as safer."},{"startTime":2167.625,"endTime":2172.509,"body":"And the other is to start in the red zone, uh, but use kind of"},{"startTime":2172.509,"endTime":2175.765,"body":"formal mathematical guarantees that what we're doing is actually safe."},{"startTime":2186.425,"endTime":2191.648,"body":"And so, you know, one of the things that seems important is in the early"},{"startTime":2191.648,"endTime":2191.997,"body":"stages of training these, uh, models, um, you know, we are, we protect intellectual property"},{"startTime":2191.997,"endTime":2192.345,"body":"carefully with copyright and people demand or people rightly get the value of what they're"},{"startTime":2192.345,"endTime":2202.445,"body":"doing."},{"startTime":2231.725,"endTime":2237.352,"body":"Before you can build a nuclear power station, you have to give all kinds of"},{"startTime":2237.352,"endTime":2237.727,"body":"quantitative numbers to demonstrate the safety of your, um, nuclear power reactor, which means that"},{"startTime":2237.727,"endTime":2245.605,"body":"we do have relatively few nuclear incidents."},{"startTime":2246.465,"endTime":2251.857,"body":"Of course, the kind of contrary side to that is that actually they have sufficiently"},{"startTime":2251.857,"endTime":2252.217,"body":"limited the application of nuclear power, that now more people die of the kind of"},{"startTime":2252.217,"endTime":2261.205,"body":"effects of air pollution from coal than die from nuclear incidents."},{"startTime":2277.035,"endTime":2282.902,"body":"This seems like an important concept that we can lift across for AI from the"},{"startTime":2282.902,"endTime":2283.685,"body":"pharmaceutical industry."},{"startTime":2286.305,"endTime":2291.145,"body":"So before we, uh, are, before we put a drug out into the world, we"},{"startTime":2291.145,"endTime":2291.468,"body":"are very careful to make sure it's tested in every, uh, regard possible so that"},{"startTime":2291.468,"endTime":2298.245,"body":"it can be as safe as possible."},{"startTime":2298.885,"endTime":2304.138,"body":"Although of course the downside to this one is that people say this is, this"},{"startTime":2304.138,"endTime":2304.488,"body":"process is now taking too long and too expensive and actually harming people by preventing"},{"startTime":2304.488,"endTime":2314.645,"body":"like good medicines from getting out rather than purely, uh, uh, a safety, um, factor."},{"startTime":2338.505,"endTime":2344.223,"body":"And finally, in aerospace, um, if there's ever an accident, there's a very truth-seeking investigation"},{"startTime":2344.223,"endTime":2344.605,"body":"on trying to figure out what happened, like a sort of very carefully blame-free culture"},{"startTime":2344.605,"endTime":2344.986,"body":"so that we really understand exactly what failed so that we can make these things"},{"startTime":2344.986,"endTime":2356.805,"body":"safer over time."},{"startTime":2421.625,"endTime":2424.965,"body":"So I'll just take two minutes to explain a little bit about this approach."},{"startTime":2426.185,"endTime":2431.401,"body":"So we've designed it to be both powerful and trusted, and that means that there"},{"startTime":2431.401,"endTime":2431.748,"body":"are sort of three basic tenets, um, for every system that's built according to these"},{"startTime":2431.748,"endTime":2436.965,"body":"principles."},{"startTime":2447.905,"endTime":2453.996,"body":"Uh, and every algorithm has a governance mechanism controlling its actions, it's modular so that"},{"startTime":2453.996,"endTime":2457.245,"body":"each component can be separately tested and understood."},{"startTime":2457.705,"endTime":2463.427,"body":"So it means that you really can build in this incremental approach, um, making sure"},{"startTime":2463.427,"endTime":2468.005,"body":"things are safe as you go, and it is fundamentally human first."},{"startTime":2468.905,"endTime":2474.245,"body":"So it is explainable both structurally and algorithmically."},{"startTime":2474.545,"endTime":2479.368,"body":"And so basically what this means is humans can actually look at the systems and"},{"startTime":2479.368,"endTime":2479.69,"body":"understand what they're doing and why in such a way that they can make good"},{"startTime":2479.69,"endTime":2486.765,"body":"choices about whether to implement it or not."},{"startTime":2525.385,"endTime":2528.685,"body":"So what should we do about ai?"},{"startTime":2556.875,"endTime":2560.725,"body":"But that is unfortunately the sort of position that we are in, in the debate"},{"startTime":2560.725,"endTime":2561.495,"body":"in AI today."},{"startTime":2562.555,"endTime":2566.175,"body":"And so at the very least, we need to start carving out these two categories."},{"startTime":2578.675,"endTime":2582.135,"body":"The red zone, I think it is sensible to be cautious."},{"startTime":2583.115,"endTime":2588.924,"body":"The problem is, it's not totally clear what the right way to be cautious is,"},{"startTime":2588.924,"endTime":2589.311,"body":"is the sensible path to regulate with all the downsides and complexities of regulation, or"},{"startTime":2589.311,"endTime":2589.698,"body":"should we build in a more open fashion in the kind of open source ideals"},{"startTime":2589.698,"endTime":2604.415,"body":"of many eyes, make light bugs or make shallow bugs?"},{"startTime":2606.795,"endTime":2612.975,"body":"And crucially, if we are to build these things, what values should AI represent?"},{"startTime":2621.555,"endTime":2628.305,"body":"And while this may seem slightly unsatisfying as an ending, building transformative AI is going"},{"startTime":2628.305,"endTime":2633.255,"body":"to force us to decide what it means to be human."},{"startTime":2633.875,"endTime":2638.975,"body":"We are going to have to decide what we put into these algorithms, what we"},{"startTime":2638.975,"endTime":2642.375,"body":"care about, and what we are willing to fight for."},{"startTime":2643.355,"endTime":2646.855,"body":"And at the moment, we are just in the middle of the story."},{"startTime":2646.905,"endTime":2649.615,"body":"There is no ending right now."},{"startTime":2649.795,"endTime":2654.295,"body":"The the end would depend on the decisions made by people in this room."},{"startTime":2655.305,"endTime":2656.015,"body":"Thank you very much."},{"startTime":2665.965,"endTime":2669.936,"body":"Gosh, mark, that was a brilliant sort of sweep across everything I remember in awe"},{"startTime":2669.936,"endTime":2671.525,"body":"of what you were talking about."},{"startTime":2671.585,"endTime":2672.965,"body":"Now, uh, my name's Richard Harvey."},{"startTime":2693.145,"endTime":2696.925,"body":"Um, let's start about, start with bias."},{"startTime":2710.585,"endTime":2715.397,"body":"And I suppose \u003claugh\u003e a biased AI perhaps isn't the sort of killer AI that"},{"startTime":2715.397,"endTime":2718.605,"body":"one might have imagined, but it's certainly a bad ai"},{"startTime":2718.755,"endTime":2723.805,"body":"Certainly, and so, so yes, uh, so AI can be biased."},{"startTime":2724.345,"endTime":2731.095,"body":"Um, it certainly depends on the training data, although there are, it is interesting in"},{"startTime":2731.095,"endTime":2731.545,"body":"that once you put something into maths, it's actually often easier to fix faster because"},{"startTime":2731.545,"endTime":2731.995,"body":"human brains are, well, we've tried for centuries, maybe millennia to remove bias from human"},{"startTime":2731.995,"endTime":2745.045,"body":"brains."},{"startTime":2774.745,"endTime":2780.085,"body":"But I just, one like slight, uh, kind of tangent to that question."},{"startTime":2793.425,"endTime":2798.834,"body":"So people talk about like AI ethics as different to kind of AI safety and"},{"startTime":2798.834,"endTime":2799.195,"body":"you know, there are all sorts of inside baseball fights between whose bit is more"},{"startTime":2799.195,"endTime":2804.605,"body":"important."},{"startTime":2808.705,"endTime":2813.227,"body":"Um, I like to use the analogy of a car, like I want seat belts"},{"startTime":2813.227,"endTime":2813.529,"body":"and I want a catalytic converter because I care about the short-term risks of crashes,"},{"startTime":2813.529,"endTime":2820.765,"body":"but I care about the long-term risks of global warming."},{"startTime":2821.225,"endTime":2824.942,"body":"And I think insisting that you have to do one or you have to do"},{"startTime":2824.942,"endTime":2826.925,"body":"other of those is just fundamentally a mistake."},{"startTime":2835.575,"endTime":2836.185,"body":"Yeah, yeah."},{"startTime":2855.025,"endTime":2861.657,"body":"Do, do you know of any progress on, on that, you know, um, so for"},{"startTime":2861.657,"endTime":2867.405,"body":"characterizing intelligence for example, or, um, even characterizing malevolence in any numerical way?"},{"startTime":2878.985,"endTime":2885.426,"body":"So I think people feel like as we step forward in our knowledge of, um,"},{"startTime":2885.426,"endTime":2885.856,"body":"ai, we can sort of make analogies back to the human brain, but I don't"},{"startTime":2885.856,"endTime":2894.445,"body":"know that anyone would want to."},{"startTime":2923.625,"endTime":2928.795,"body":"Now, I'm not sure if that assertion is completely \u003claugh\u003e true, but, but perhaps behind"},{"startTime":2928.795,"endTime":2933.965,"body":"it is the how, how general, how, how do you measure generality of an ai?"},{"startTime":2936.305,"endTime":2938.605,"body":"Uh, well the, I mean, so these are huge, huge questions."},{"startTime":2938.685,"endTime":2939.685,"body":"I think I, sorry, I've"},{"startTime":2942.285,"endTime":2943.365,"body":"There's metro question."},{"startTime":2967.305,"endTime":2974.312,"body":"And so I think, you know, it, it seems fairly clear that we are making"},{"startTime":2974.312,"endTime":2974.779,"body":"progress in that like it's, um, GPT-3 four have substantially increased our, um, ability to"},{"startTime":2974.779,"endTime":2986.925,"body":"have a single algorithm, um, uh, solve tests in very different domains."},{"startTime":2987.785,"endTime":2992.142,"body":"Uh, so, so I, I don't think I, I basically don't think I agree with"},{"startTime":2992.142,"endTime":2993.885,"body":"the underlying assumption of the question."},{"startTime":2993.885,"endTime":3000.302,"body":"Mm-Hmm, \u003caffirmative\u003e, yeah, I think we are making progress, um, how fast is really gonna"},{"startTime":3000.302,"endTime":3003.725,"body":"depend on what extra scientific breakthroughs we need."},{"startTime":3003.825,"endTime":3009.594,"body":"So I don't think that LLMs as they're these large language models, the, the name"},{"startTime":3009.594,"endTime":3009.979,"body":"for the algorithm that sits behind, uh, chat GPT, I don't think those are enough"},{"startTime":3009.979,"endTime":3020.365,"body":"for general intelligence, um, uh, in the way that most people mean it."},{"startTime":3020.825,"endTime":3026.845,"body":"Uh, but we get into some quite detailed specific questions around, um, definitions here."},{"startTime":3040.705,"endTime":3041.125,"body":"Mm-Hmm,"},{"startTime":3041.165,"endTime":3041.885,"body":"\u003caffirmative\u003e, yeah."},{"startTime":3048.805,"endTime":3050.565,"body":"Need like PhDs these days."},{"startTime":3094.265,"endTime":3097.125,"body":"And that can't be human led in this regard."},{"startTime":3110.185,"endTime":3114.645,"body":"So let's talk about a circumstance that, that faculty would, uh, work in."},{"startTime":3118.585,"endTime":3123.558,"body":"So, you know, digital marketing is, you know, there are these exchanges where transactions happen"},{"startTime":3123.558,"endTime":3123.889,"body":"on the millisecond and obviously a human being is not gonna execute every one of"},{"startTime":3123.889,"endTime":3124.221,"body":"those transactions, but the human being can set the framework for, uh, the strategy, how"},{"startTime":3124.221,"endTime":3124.552,"body":"much you want to spend in what given time on what types of campaign, and"},{"startTime":3124.552,"endTime":3142.125,"body":"then can set the governance protocols that sit around the algorithm."},{"startTime":3144.985,"endTime":3149.011,"body":"If, um, we, if you spend this much in this unit time, then stop trading"},{"startTime":3149.011,"endTime":3150.085,"body":"these kinds of things."},{"startTime":3164.635,"endTime":3165.485,"body":"Yeah, great answer."},{"startTime":3165.945,"endTime":3169.705,"body":"Um, yeah, here's a good one."},{"startTime":3176.215,"endTime":3184.261,"body":"Yeah, I think, um, I don't know that we have to be super precise though"},{"startTime":3184.261,"endTime":3189.625,"body":"for, uh, for the purposes of, um, oops, ensuring safety."},{"startTime":3192.875,"endTime":3194.305,"body":"Maybe the boundary doesn't really matter."},{"startTime":3194.455,"endTime":3196.745,"body":"What matters is exactly you, you know, the end points."},{"startTime":3197.055,"endTime":3197.545,"body":"Exactly."},{"startTime":3202.925,"endTime":3207.279,"body":"And then in some sense there's a fuzzy boundary to the red zone and then"},{"startTime":3207.279,"endTime":3207.57,"body":"it's gonna be our decision on how much risk we choose to take in that"},{"startTime":3207.57,"endTime":3211.925,"body":"boundary."},{"startTime":3212.905,"endTime":3218.022,"body":"And there's an interesting question here, which I've temporarily lost, but let me paraphrase, which"},{"startTime":3218.022,"endTime":3218.363,"body":"is, um, so AI is, is very dependent on its training data and the questionnaire"},{"startTime":3218.363,"endTime":3224.845,"body":"is asking about data poisoning."},{"startTime":3254.505,"endTime":3259.573,"body":"Um, in the detail there are many, like the kind of coffee robot failure mode,"},{"startTime":3259.573,"endTime":3259.911,"body":"there are many, many examples of sort of equivalent failure modes that can happen including"},{"startTime":3259.911,"endTime":3260.249,"body":"sort of a data poisoning such that, uh, the coffee robot thinks it's executing on"},{"startTime":3260.249,"endTime":3273.765,"body":"what I asked, but, but in fact is doing something completely different."},{"startTime":3275.665,"endTime":3282.853,"body":"And, um, people are working on, uh, on a variety of paths, including just making"},{"startTime":3282.853,"endTime":3288.605,"body":"sure the data's reasonably good, um, uh, to kind of solve that."},{"startTime":3290.535,"endTime":3296.024,"body":"So when you are talking about ai, you are implicitly talking about, um, not only"},{"startTime":3296.024,"endTime":3296.39,"body":"the algorithm, but also the, the data that that feeds it, this sort of classic"},{"startTime":3296.39,"endTime":3304.075,"body":"chat GPT problem, isn't it that it,"},{"startTime":3305.055,"endTime":3305.475,"body":"For sure."},{"startTime":3305.745,"endTime":3306.035,"body":"It's,"},{"startTime":3306.385,"endTime":3310.195,"body":"It's ingesting stuff that might not have been safe in the first place."},{"startTime":3311.415,"endTime":3318.315,"body":"Uh, so I mean, of course the way our algorithms work at the moment mostly"},{"startTime":3318.315,"endTime":3324.755,"body":"is, um, that they get trained and then that trained model goes into production."},{"startTime":3366.715,"endTime":3367.475,"body":"\u003claugh\u003e, ask someone here."},{"startTime":3416.955,"endTime":3421.525,"body":"Like that is a magnificent opportunity to transform education."},{"startTime":3433.875,"endTime":3434.685,"body":"Yeah, great answer."},{"startTime":3444.905,"endTime":3447.245,"body":"So I, so do you guess can't remember."},{"startTime":3447.245,"endTime":3453.145,"body":"Anyway, what I would say is, um, I think there will turn out to be"},{"startTime":3453.145,"endTime":3459.045,"body":"some notion like, so in computing we have this notion of a universal classical computer."},{"startTime":3467.945,"endTime":3471.955,"body":"And then we, so that was for a while, for a long time thought to"},{"startTime":3471.955,"endTime":3475.965,"body":"be it if you just had those logic gates, you could do perform any computation."},{"startTime":3545.525,"endTime":3546.765,"body":"There's, there must be a hundred here."},{"startTime":3547.415,"endTime":3548.125,"body":"Sorry about that."},{"startTime":3548.265,"endTime":3552.322,"body":"For those of you in person, I guess you're around for a few minutes so"},{"startTime":3552.322,"endTime":3553.405,"body":"people can badger you."},{"startTime":3553.415,"endTime":3556.405,"body":"Other than that, mark, thank you very much indeed."}]}