WEBVTT 00:00:00.160 --> 00:00:03.277 you play this forward in your brain. You you've been in the tech industry for a 00:00:03.360 --> 00:00:07.277 long time and from looking at your work, you it feels like you're describing this 00:00:07.360 --> 00:00:10.637 as the most sort of transformative, potentially harmful technology that 00:00:10.720 --> 00:00:14.237 humans have really ever seen. You know, maybe alongside the nuclear bomb, I 00:00:14.320 --> 00:00:16.637 guess, but some would say even 00:00:16.720 --> 00:00:21.437 potentially worse because of the nature of the intelligence and its autonomy. 00:00:21.520 --> 00:00:25.277 You must have moments where you you think forward into the future and your 00:00:25.360 --> 00:00:28.237 thoughts about that future aren't so rosy. 00:00:28.320 --> 00:00:30.077 >> Well, >> because I have those moments. Yes, but 00:00:30.160 --> 00:00:33.757 but let's let's think let's answer the question. I said think five years. In 00:00:33.840 --> 00:00:37.117 five years, you'll have two or three more turns of the crank of these large 00:00:37.200 --> 00:00:40.237 models. These large models are scaling 00:00:40.320 --> 00:00:43.037 with ability that is unprecedented. 00:00:43.120 --> 00:00:45.677 There's no evidence that the scaling has 00:00:45.760 --> 00:00:48.557 laws as they're called have begun to 00:00:48.640 --> 00:00:53.117 stop. They will eventually stop, but we're not there yet. Each one of these 00:00:53.200 --> 00:00:56.797 cranks looks like it's a factor of two, factor of three, factor of four of 00:00:56.880 --> 00:00:58.877 capability. So let's just say turning 00:00:58.960 --> 00:01:02.397 the crank all of these systems get 50 00:01:02.480 --> 00:01:04.957 times or 100 times more powerful in and 00:01:05.040 --> 00:01:09.437 of itself. That's a very big deal because those systems will be capable of 00:01:09.520 --> 00:01:11.437 physics and math. You see this with O.1 00:01:11.520 --> 00:01:14.637 and um and open AI all the other things 00:01:14.720 --> 00:01:17.757 that are occurring. Now what are the 00:01:17.840 --> 00:01:22.157 dangers? Well there's the most obvious one is cyber attacks. There's evidence 00:01:22.240 --> 00:01:26.397 that the raw models, these are the ones that have not been released, can do what 00:01:26.480 --> 00:01:30.157 are called day zero attacks as well or better than humans. A day zero attack is 00:01:30.240 --> 00:01:33.597 an attack that's unknown. They can discover something new. And how do they 00:01:33.680 --> 00:01:36.797 do it? They just keep trying because they're computers and they have nothing 00:01:36.880 --> 00:01:39.517 else to do. They don't sleep. They don't eat. They just turn them on and they 00:01:39.600 --> 00:01:44.397 just keep going. Um, so the so cyber is an example where everybody's concerned. 00:01:44.480 --> 00:01:48.637 Another one is biology. Viruses are relatively easy to make and you can 00:01:48.720 --> 00:01:52.077 imagine coming up with really bad viruses. There's a whole team. I'm part 00:01:52.160 --> 00:01:55.517 of a commission. We're looking at this to try to make sure that doesn't happen. 00:01:55.600 --> 00:01:57.757 I already mentioned misinformation. 00:01:57.840 --> 00:02:02.717 Another probably negative, but we'll see is the 00:02:02.800 --> 00:02:04.557 development of new forms of warfare. 00:02:04.640 --> 00:02:07.437 I've written extensively on how war is 00:02:07.520 --> 00:02:09.837 changing. And the way to understand 00:02:09.920 --> 00:02:12.157 historic war is that it's the 00:02:12.240 --> 00:02:16.877 stereotypically the the soldier with the gun, you know, on one side and so forth. 00:02:16.960 --> 00:02:20.797 World War trenches. You see this, by the way, in Ukra in the Ukraine fight today 00:02:20.880 --> 00:02:24.637 where the Ukrainians are holding on valiantly against the Russian onslaught. 00:02:24.720 --> 00:02:27.037 But it's sort of, you know, mono ammono, 00:02:27.120 --> 00:02:29.037 you know, man against man. Sort of all 00:02:29.120 --> 00:02:34.317 of the stereotypes of war. So in a drone world, which is the sort of the fastest 00:02:34.400 --> 00:02:37.917 way to build new robots is to build drones, you'll be sitting in a command 00:02:38.000 --> 00:02:42.077 center in some office building connected by a network and you'll be doing harm to 00:02:42.160 --> 00:02:46.077 the other side while you're drinking your coffee. Right? That's a change in 00:02:46.160 --> 00:02:49.517 the logic of war. Um, and it's applicable to both sides. I don't think 00:02:49.600 --> 00:02:51.677 anyone quite understands how war will 00:02:51.760 --> 00:02:56.717 change, but I will tell you that in UK in the Russia and Ukraine war, you're 00:02:56.800 --> 00:02:58.397 seeing a new form of warfare being 00:02:58.480 --> 00:03:01.917 invented right now, right? Um, both 00:03:02.000 --> 00:03:04.317 sides have lots of drones. Tanks are no 00:03:04.400 --> 00:03:06.957 longer very useful. A $5,000 drone can 00:03:07.040 --> 00:03:11.837 kill a $5 million tank. Um, so that's called the kill ratio. So basically, 00:03:11.920 --> 00:03:15.277 it's drone on drone. And so now people are trying to figure out how how to have 00:03:15.360 --> 00:03:17.517 one drone destroy the other drone. 00:03:17.600 --> 00:03:20.237 Right? This will ultimately take over 00:03:20.320 --> 00:03:22.797 war and conflict in our world in total. 00:03:22.880 --> 00:03:26.077 >> You mentioned raw models. This is a concept that I don't think people 00:03:26.160 --> 00:03:28.157 understand exists. The idea that there's 00:03:28.240 --> 00:03:33.197 some other model that's the raw model that is capable of much worse than the 00:03:33.280 --> 00:03:35.117 thing we play with on our computers every day. 00:03:35.200 --> 00:03:37.837 >> It's important to establish how these things work. So you the way these 00:03:37.920 --> 00:03:40.077 algorithms work is they have complicated 00:03:40.160 --> 00:03:41.677 uh training things where they suck all 00:03:41.760 --> 00:03:47.197 the information in and they uh one we currently believe we've sort of sucked 00:03:47.280 --> 00:03:50.077 all of the written word that's available. It doesn't mean there isn't 00:03:50.160 --> 00:03:53.677 more but we've we've literally done such a good job of sucking everything that 00:03:53.760 --> 00:03:57.437 humans have ever written. It's all in these big computers. And when I say 00:03:57.520 --> 00:04:01.117 computers I don't mean computers. I mean supercomputers with enormous memories 00:04:01.200 --> 00:04:05.517 and the scale is mindboggling. Uh, and of course there's this company called 00:04:05.600 --> 00:04:08.477 Nvidia which makes the chips which is now one of the most valuable companies 00:04:08.560 --> 00:04:12.397 in the world. Um, surprisingly so 00:04:12.480 --> 00:04:16.077 incredibly successful because they're so central to this revolution and good for 00:04:16.160 --> 00:04:18.157 Jensen and his team. So the important 00:04:18.240 --> 00:04:23.197 thing is when you do this training it comes out with a raw model, right? It 00:04:23.280 --> 00:04:27.197 takes six months and you know you wait 24 hours a day you can watch it. It gets 00:04:27.280 --> 00:04:31.037 close to there's a measurement that they use called the loss function. when it 00:04:31.120 --> 00:04:32.557 gets to a certain number, they say, 00:04:32.640 --> 00:04:35.117 "Good enough." So then they go, "What do 00:04:35.200 --> 00:04:38.077 we have?" Right? What do we do? Right? 00:04:38.160 --> 00:04:42.077 Um so the first thing is let's figure out what it knows. 00:04:42.160 --> 00:04:46.397 So they have a set of tests and of course it knows all sorts of bad things 00:04:46.480 --> 00:04:47.917 which they immediately then tell it not 00:04:48.000 --> 00:04:50.557 to answer. To me the most interesting 00:04:50.640 --> 00:04:55.357 question is in over a fiveyear period 00:04:55.440 --> 00:05:00.157 these systems will learn things that we don't know they learn. How will you test 00:05:00.240 --> 00:05:03.517 for things that you don't know they know? 00:05:03.600 --> 00:05:06.317 The answer in the industry is that they 00:05:06.400 --> 00:05:11.197 have incredibly clever people who sit there and they fiddle, literally fiddle 00:05:11.280 --> 00:05:13.277 with the networks and say, "I'm going to 00:05:13.360 --> 00:05:15.997 I'm going to see if it knows this. 00:05:16.080 --> 00:05:20.317 I'll see if it can do this." And then they make a list and they say, "That's 00:05:20.400 --> 00:05:24.637 good. That's not so good." Right? So all of these transformations. So for 00:05:24.720 --> 00:05:28.637 example, you can show it a picture of a website and it can generate the code to 00:05:28.720 --> 00:05:32.637 generate a website. All of those were not expected. They just happened. It's 00:05:32.720 --> 00:05:35.197 called emergent behavior. 00:05:35.280 --> 00:05:36.237 >> Scary. 00:05:36.320 --> 00:05:40.157 >> Scary but exciting. And so far um the 00:05:40.240 --> 00:05:44.957 systems have held. The governments have worked well. Um the these trust and 00:05:45.040 --> 00:05:46.797 safety groups are working here in the 00:05:46.880 --> 00:05:49.997 UK. Um one year ago was the first trust 00:05:50.080 --> 00:05:54.557 and safety conference. Um the government did a fantastic job. The team that was 00:05:54.640 --> 00:05:59.437 assembled was the best of all the country teams here in the UK. Um, now 00:05:59.520 --> 00:06:02.237 what's happening is these are happening around the world. The next one is in 00:06:02.320 --> 00:06:06.637 France in uh early February, and I expect a similar good result. 00:06:06.720 --> 00:06:09.517 >> Do you think we're going to have to guard I mean, you talk about this, but 00:06:09.600 --> 00:06:13.917 do you think we're going to have to guard these raw models with with guns 00:06:14.000 --> 00:06:16.317 and tanks and machinery and stuff? 00:06:16.400 --> 00:06:20.637 >> I worked for the Secretary of Defense for a while. Uh, in my in Google, you 00:06:20.720 --> 00:06:23.677 could spend 20% of your time on other things. So I worked for the secretary of 00:06:23.760 --> 00:06:25.837 defense to try to understand the US 00:06:25.920 --> 00:06:28.637 military and um one of the things that 00:06:28.720 --> 00:06:31.277 we did is we visited a plutonium um 00:06:31.360 --> 00:06:35.757 factory. Plutonium is incredibly dangerous and incredibly secret and so 00:06:35.840 --> 00:06:39.277 this particular base is inside of another base. So you go through the 00:06:39.360 --> 00:06:42.397 first set of machine guns and then you have normal thing and then you go into 00:06:42.480 --> 00:06:45.517 the special place with even more machines, guns and even because it's so 00:06:45.600 --> 00:06:49.357 secure. So the the metaphor is do you 00:06:49.440 --> 00:06:53.517 fundamentally believe that the computers that I'm talking about will be of such 00:06:53.600 --> 00:06:57.677 value and such danger that they'll have their own data center with their own 00:06:57.760 --> 00:07:01.357 guards which of course might be computer guards. But the important thing is that 00:07:01.440 --> 00:07:04.717 it's so special that it has to be protected in the same way that we 00:07:04.800 --> 00:07:06.877 protect nuclear bombs and prolifer uh 00:07:06.960 --> 00:07:09.917 and programming. An alternative model is 00:07:10.000 --> 00:07:15.197 to say that this technology will spread pretty broadly and there'll be many such 00:07:15.280 --> 00:07:19.917 places. If it's a small number of groups, the 00:07:20.000 --> 00:07:23.357 governments will figure out a way to do deterrence and they'll figure out a way 00:07:23.440 --> 00:07:26.877 to do non-prololiferation. So, I'll make something up. I'll say 00:07:26.960 --> 00:07:30.317 there's a couple in China, there's a few in the US, there's one in in Britain. Of 00:07:30.400 --> 00:07:33.517 course, we're all tied together between the US and Britain and maybe in a few 00:07:33.600 --> 00:07:37.197 other places. That's a manageable problem. On the other hand, let's 00:07:37.280 --> 00:07:40.157 imagine that that power is ultimately so 00:07:40.240 --> 00:07:43.357 easy to copy that it spreads globally 00:07:43.440 --> 00:07:47.677 and it's accessible to for example terrorists. Then you have a very serious 00:07:47.760 --> 00:07:49.757 proliferation problem which is not yet 00:07:49.840 --> 00:07:53.117 solved. This is again speculation 00:07:53.200 --> 00:07:55.037 >> because I think a lot about adversaries 00:07:55.120 --> 00:07:58.077 in China and Russia and Putin and and I 00:07:58.160 --> 00:08:02.397 think I know you talk about them being a few years behind, maybe one or two years 00:08:02.480 --> 00:08:05.517 behind, but they're eventually going to get there. They're eventually going to 00:08:05.600 --> 00:08:08.717 get to the point where they have these large language models or these AIs that 00:08:08.800 --> 00:08:11.437 can do these day zero attacks on our 00:08:11.520 --> 00:08:16.157 nation and they don't have the same in like sort of 00:08:16.240 --> 00:08:17.997 social incentive structure if they're a 00:08:18.080 --> 00:08:21.357 communist country to protect and to 00:08:21.440 --> 00:08:25.357 >> um guard against these things. Are you not worried about what China's going to 00:08:25.440 --> 00:08:29.117 do? >> Um I am worried and I'm worried because 00:08:29.200 --> 00:08:33.757 you're going into a space of great power without fully defined boundaries. What 00:08:33.840 --> 00:08:37.517 Kissinger and we talk about this in the book the the genesis book is 00:08:37.600 --> 00:08:40.957 fundamentally about what happens to society with the arrival of this new 00:08:41.040 --> 00:08:43.837 intelligence and the first book we did 00:08:43.920 --> 00:08:48.077 age of AI was right before chat GPT so now everybody kind of understands how 00:08:48.160 --> 00:08:51.757 powerful these things are we talked about it now you understand it so once 00:08:51.840 --> 00:08:53.917 these things show up who's going to run 00:08:54.000 --> 00:08:55.357 them who's going to be in charge how 00:08:55.440 --> 00:08:59.037 will they be used so from my perspective 00:08:59.120 --> 00:09:01.597 I believe at the moment anyway that 00:09:01.680 --> 00:09:04.557 China will behave relatively responsibly 00:09:04.640 --> 00:09:07.037 and the reason is that it's not in their 00:09:07.120 --> 00:09:09.597 interest to have free speech 00:09:09.680 --> 00:09:12.477 in every case in China when they have a 00:09:12.560 --> 00:09:17.437 choice of giving freedom to their citizens or not they choose non-freedom 00:09:17.520 --> 00:09:18.797 and I know this because I spent through 00:09:18.880 --> 00:09:20.877 all their I spent all that time dealing 00:09:20.960 --> 00:09:24.477 with it so it sure looks to to me like 00:09:24.560 --> 00:09:26.717 the Chinese AI solution will be 00:09:26.800 --> 00:09:31.917 different from the west because of that fundamental bias against freedom of 00:09:32.000 --> 00:09:36.077 speech. Because these things are noisy. They make a lot of noise. 00:09:36.160 --> 00:09:38.477 >> They'll probably still make AI weapons though. 00:09:38.560 --> 00:09:40.717 >> Well, on the weapon side, you have to 00:09:40.800 --> 00:09:44.157 assume that every new technology is 00:09:44.240 --> 00:09:47.037 ultimately strengthened in a war. Um the 00:09:47.120 --> 00:09:49.357 tank was invented in World War I. At the 00:09:49.440 --> 00:09:54.077 same time, you had the initial forms of uh airplanes. Much of the Second World 00:09:54.160 --> 00:09:56.317 War was an air campaign which 00:09:56.400 --> 00:10:01.357 essentially built many many things. And if you look at the the there's a a book 00:10:01.440 --> 00:10:03.117 called Freedom's Forge about the 00:10:03.200 --> 00:10:07.117 American U structure. According to the 00:10:07.200 --> 00:10:10.317 book, they ultimately got to the point where they could build two or three 00:10:10.400 --> 00:10:14.077 airplanes a day at scale. So in an 00:10:14.160 --> 00:10:17.517 emergency, nations have enormous power. 00:10:17.600 --> 00:10:21.837 >> I get asked all the time if everyone's if anyone's going to have a job left to 00:10:21.920 --> 00:10:25.517 do because this is the disruption of intelligence and whether it's people 00:10:25.600 --> 00:10:29.517 driving cars today. I mean, we saw the Tesla announcement of the robo taxis, 00:10:29.600 --> 00:10:32.957 whether it's accountants, lawyers, and everyone in between, that's job or 00:10:33.040 --> 00:10:36.157 podcasters. Are we going to have jobs left? 00:10:36.240 --> 00:10:41.827 >> Well, um, this question has been asked for 200 years. Um, there was 00:10:41.910 --> 00:10:45.277 [clears throat] there were the lites here in Britain way back when. And 00:10:45.360 --> 00:10:48.797 inevitably when these technologies come along, there's all these fears about 00:10:48.880 --> 00:10:51.837 them. Indeed, with the lites, there were riots and people, you know, destroying 00:10:51.920 --> 00:10:53.997 the looms and all of this kind of stuff, 00:10:54.080 --> 00:10:57.997 but somehow we got through it. So um my 00:10:58.080 --> 00:11:00.637 own view is that there will be a lot of 00:11:00.720 --> 00:11:02.637 job dislocation 00:11:02.720 --> 00:11:05.117 but there will be a lot more jobs not 00:11:05.200 --> 00:11:09.837 fewer jobs and here's why we have a demographic problem in the world 00:11:09.920 --> 00:11:12.637 especially in the developed developed world where we're not having enough 00:11:12.720 --> 00:11:16.957 children. Uh that's well understood. Uh furthermore we have a lot of older 00:11:17.040 --> 00:11:20.077 people and and the younger people have to take care of the older people and 00:11:20.160 --> 00:11:23.277 they have to be more productive. If you have young people who need to be more 00:11:23.360 --> 00:11:26.077 productive, the best way to make them more more productive is to give them 00:11:26.160 --> 00:11:29.117 more tools to make them more productive. 00:11:29.200 --> 00:11:34.157 Whether it's a machinist that goes from a manual machine into a CNC machine or 00:11:34.240 --> 00:11:37.917 in in the more modern case of a knowledge worker who can achieve more 00:11:38.000 --> 00:11:42.157 objectives. We need that productivity group. If you look at Asia, which is the 00:11:42.240 --> 00:11:46.317 centerpiece of manufacturing, they have all this cheap labor. Well, it's not so 00:11:46.400 --> 00:11:50.477 cheap anymore. So, do you know what they did? They added robotic assembly lines. 00:11:50.560 --> 00:11:53.837 So today when you go to China in particular, it's also true in Japan and 00:11:53.920 --> 00:11:58.237 Korea, the manufacturing is largely done by robots. Why? Because their 00:11:58.320 --> 00:11:59.837 demographics are terrible and their cost 00:11:59.920 --> 00:12:02.557 of labor is too high. So the future is 00:12:02.640 --> 00:12:05.437 not fewer jobs. It's actually a lot of 00:12:05.520 --> 00:12:09.597 jobs that are unfilled with people who may have a job skill mismatch, which is 00:12:09.680 --> 00:12:12.077 why education is so important. Now, what 00:12:12.160 --> 00:12:14.637 are examples of jobs that go away? 00:12:14.720 --> 00:12:18.077 Automation has always gotten rid of jobs 00:12:18.160 --> 00:12:22.797 that are dangerous. physically dangerous or ones which are 00:12:22.880 --> 00:12:26.557 essentially too repetitive and too boring for humans. I'll give you an 00:12:26.640 --> 00:12:29.437 example. Security guards. It makes sense 00:12:29.520 --> 00:12:34.077 that security guards would become robotic because it's hard to be a 00:12:34.160 --> 00:12:37.837 security guard. You fall asleep. You don't know quite what to and these 00:12:37.920 --> 00:12:42.317 systems can be smart enough to be very very good security. Now, these are these 00:12:42.400 --> 00:12:45.517 are important sources of income for these people. They're going to have to 00:12:45.600 --> 00:12:50.717 find another job. Another example in in the media in um Hollywood, everyone's 00:12:50.800 --> 00:12:54.157 concerned that AI is going to take over their jobs. All the evidence is the 00:12:54.240 --> 00:12:58.957 inverse. And here's why. Um the stars still get money, the producers still 00:12:59.040 --> 00:13:02.557 make money, they still distribute their movie, but their cost of making the 00:13:02.640 --> 00:13:05.917 movie is lower because they use more they use, for example, synthetic 00:13:06.000 --> 00:13:09.197 backdrops. So they don't have to build the set. Um they can do synthetic 00:13:09.280 --> 00:13:11.517 makeup. Now, there are job losses there. 00:13:11.600 --> 00:13:15.517 So the people who make the make make the set and do the makeup are going to have 00:13:15.600 --> 00:13:19.677 to go back into construction and personal care. By the way, in America, 00:13:19.760 --> 00:13:23.037 and I think it's true here, there's an enormous shortage of people who can do 00:13:23.120 --> 00:13:26.637 highquality craftsmanship, right? Those people will have jobs. They're just 00:13:26.720 --> 00:13:29.997 different >> and they may not be in Los Angeles. 00:13:30.080 --> 00:13:32.877 >> Am I going to have to interface with this technology? Am I going to have to 00:13:32.960 --> 00:13:36.797 get a neural link in my brain? because you we you you go over the subject of 00:13:36.880 --> 00:13:38.717 there being these sort of two species of 00:13:38.800 --> 00:13:43.997 humans potentially ones that do have a way to incorporate themselves more with 00:13:44.080 --> 00:13:48.237 artificial intelligence and those that don't and if and if that is the case 00:13:48.320 --> 00:13:51.677 what is the time horizon in your view of that happening 00:13:51.760 --> 00:13:54.557 >> I think neuralink is much more speculative because you're dealing with 00:13:54.640 --> 00:13:57.757 direct brain connection and nobody's going to drill on my brain until it 00:13:57.840 --> 00:14:01.277 needs it trust me >> I suspect you feel the same 00:14:01.360 --> 00:14:08.120 >> uh I I guess my o my overall view is that um 00:14:09.680 --> 00:14:12.237 you will not notice 00:14:12.320 --> 00:14:15.117 how much of your world has been co-opted 00:14:15.200 --> 00:14:17.117 by these technologies because they will 00:14:17.200 --> 00:14:19.517 produce greater delight. 00:14:19.600 --> 00:14:23.197 If you think about it, a lot of life is 00:14:23.280 --> 00:14:27.597 inconvenient. It's fix this, call this, make this happen. AI systems should make 00:14:27.680 --> 00:14:31.277 all that seamless. You should be able to wake up in the morning and have coffee 00:14:31.360 --> 00:14:35.037 and not have a care in the world and have the computer help you have a great 00:14:35.120 --> 00:14:37.997 day. This is true of everyone. Now, what 00:14:38.080 --> 00:14:40.717 happens to your to your profession? 00:14:40.800 --> 00:14:45.597 Well, as we said, no matter how good the computers are, people are going to want 00:14:45.680 --> 00:14:48.877 to care about other people. Another example, let's imagine you have Formula 00:14:48.960 --> 00:14:53.197 1 and you have Formula 1 with humans in it, and then you have an a a robot 00:14:53.280 --> 00:14:57.597 Formula 1, which where the cars are driven by the equivalent of a robot. Is 00:14:57.680 --> 00:14:59.357 anyone going to go to the robotic 00:14:59.440 --> 00:15:04.637 Formula 1? I don't think so. Because of the drama, the human achievement and so 00:15:04.720 --> 00:15:08.077 forth. Do you think that when they run the marathon here in London, they're 00:15:08.160 --> 00:15:11.837 going to have robots running with humans? Of course not. Right? Of course, 00:15:11.920 --> 00:15:13.837 the robots can run faster than humans. 00:15:13.920 --> 00:15:17.757 It's not interesting. What is interesting is to see human achievement. 00:15:17.840 --> 00:15:20.717 So, I think the commentators who say, "Oh, there won't be any jobs. We won't 00:15:20.800 --> 00:15:25.437 care." I think they miss the point that we care a great deal about each other as 00:15:25.520 --> 00:15:29.677 human beings. We have opinions. You have an a detailed opinion about me having 00:15:29.760 --> 00:15:34.157 just met me met me right now and I for you. We just sort of naturally set up 00:15:34.240 --> 00:15:35.917 your face, your mannerisms and so forth. 00:15:36.000 --> 00:15:40.077 We can describe it all right. The robot shows up is like, "Oh my god, what 00:15:40.160 --> 00:15:43.357 another robot. How boring." >> Why is Sam Alman working on the the 00:15:43.440 --> 00:15:46.637 founder of Open AI, one of the co-founders of OpenAI, working on 00:15:46.720 --> 00:15:49.677 universal basic income projects like Worldcoin then? 00:15:49.760 --> 00:15:51.197 >> Well, WorldCon is not the same thing as 00:15:51.280 --> 00:15:54.877 universal Bitcoin. uh um universal basic 00:15:54.960 --> 00:15:57.917 income. There is a belief in the tech 00:15:58.000 --> 00:16:00.477 industry that it goes something like 00:16:00.560 --> 00:16:03.677 this. The politics of abundance, what we 00:16:03.760 --> 00:16:06.397 do is going to create so much abundance 00:16:06.480 --> 00:16:11.117 that most people won't have to work and there'll be a small number of groups 00:16:11.200 --> 00:16:14.797 that work who typically these people themselves and there will be so much 00:16:14.880 --> 00:16:17.837 surplus everyone can live like a millionaire and everyone will be happy. 00:16:17.920 --> 00:16:20.957 I completely think this is false. I think none of what I just told you is 00:16:21.040 --> 00:16:24.237 false. But all of these UBI ideas come 00:16:24.320 --> 00:16:28.717 from this notion that humans don't behave the way we actually do. So I'm 00:16:28.800 --> 00:16:31.037 I'm a critic of this view. I believe 00:16:31.120 --> 00:16:33.837 that that we as humans, so I'll give an 00:16:33.920 --> 00:16:37.677 example is um we're going to make legal 00:16:37.760 --> 00:16:41.037 the legal profession much much easier because we can automate much of the 00:16:41.120 --> 00:16:44.317 technical work of lawyers. Does that mean we're going to have fewer lawyers? 00:16:44.400 --> 00:16:48.317 No. The current lawyers will just do more laws. They'll do more they'll add 00:16:48.400 --> 00:16:52.477 more complexity. The system doesn't get easier. the humans become more 00:16:52.560 --> 00:16:56.317 sophisticated in their application of the principles. We are naturally 00:16:56.400 --> 00:16:58.957 basically uh we have this thing called 00:16:59.040 --> 00:17:03.677 um basically reciprocal altruism. That's part of us but we also have our bad 00:17:03.760 --> 00:17:06.637 sides as well. Those are not going away because of AI. 00:17:06.720 --> 00:17:10.317 >> When I think about AI, this simple analogy you often think of is say my IQ 00:17:10.400 --> 00:17:14.717 is Steven Barla is 100 and there's this AI that's sat next to me whose IQ is 00:17:14.800 --> 00:17:19.037 1,000. What on earth would you want to give Steven to do? because because that 00:17:19.120 --> 00:17:21.197 10,00 IQ would have really bad judgment 00:17:21.280 --> 00:17:23.597 in a couple cases because remember that 00:17:23.680 --> 00:17:26.397 the AI systems do not have human values 00:17:26.480 --> 00:17:29.197 unless it's added right I would much 00:17:29.280 --> 00:17:33.837 rather talk to you about something involving a a moral or human judgment 00:17:33.920 --> 00:17:37.997 even with the thousand I wouldn't mind consulting it so tell me the the history 00:17:38.080 --> 00:17:42.157 how was this resolved in the past how were these but at the end of the day in 00:17:42.240 --> 00:17:46.637 my view the core aspects of humanity which have to do with morals and 00:17:46.720 --> 00:17:49.917 judgment and belief beliefs and charisma, they're not going away. 00:17:50.000 --> 00:17:52.157 >> Is there a chance that this is the end of humanity? 00:17:52.240 --> 00:17:56.877 >> No. Um, the way humanity does is much 00:17:56.960 --> 00:18:00.557 it's much harder to eliminate all of humanity than you think. All the people 00:18:00.640 --> 00:18:04.157 that I've looked with on these biological attacks say it's it takes 00:18:04.240 --> 00:18:07.197 more than one horrific pandemic and so 00:18:07.280 --> 00:18:11.197 forth to eliminate humanity. And and the the pain can be very very high in these 00:18:11.280 --> 00:18:14.077 moments. Look at the World War I, World 00:18:14.160 --> 00:18:17.117 War II, the holodore in uh Ukraine in 00:18:17.200 --> 00:18:19.517 the 1930s, the Nazis. You know, these 00:18:19.600 --> 00:18:21.997 are horrifically painful things, but we 00:18:22.080 --> 00:18:26.717 survived, right? We we as a as a humanity survived and we will. 00:18:26.800 --> 00:18:30.957 >> I wonder if this is the moment where humans couldn't see past around the 00:18:31.040 --> 00:18:34.957 corner because, you know, I've heard you talk about how the AIs will turn in, 00:18:35.040 --> 00:18:37.517 they'll be agents and they'll be able to speak to each other and we won't be able 00:18:37.600 --> 00:18:40.717 to understand the language. >> I I have a specific proposal on that. 00:18:40.800 --> 00:18:43.357 Um, there are points where humans should 00:18:43.440 --> 00:18:47.837 assert control and I've been trying to think about where are they. I'll give 00:18:47.920 --> 00:18:51.277 you an example. There's something called recursive self-improvement where the 00:18:51.360 --> 00:18:54.477 system just keeps getting smarter and smarter and learning more and more 00:18:54.560 --> 00:18:59.357 things. At some point, if you don't know what it's learning, you should unplug 00:18:59.440 --> 00:19:02.077 it. >> But we can't unplug them, can we? 00:19:02.160 --> 00:19:05.597 >> Sure you can. There's a power plug and there's a circuit breaker. Go and turn 00:19:05.680 --> 00:19:07.277 the circuit breaker off. Another 00:19:07.360 --> 00:19:10.877 example, um there's a there's a scenario 00:19:10.960 --> 00:19:15.757 theoretical where the system is so powerful it can produce a new model 00:19:15.840 --> 00:19:18.797 faster than the previous model was checked. 00:19:18.880 --> 00:19:22.157 >> Okay, that's another intervention point. 00:19:22.240 --> 00:19:25.677 So in each of these cases um if the if 00:19:25.760 --> 00:19:29.277 agents and the technical term is called agents, what they really are is large 00:19:29.360 --> 00:19:33.837 language models with memory and you can begin to concatenate them. You can say 00:19:33.920 --> 00:19:36.877 this model does this and then it feeds into this and so forth. You can build 00:19:36.960 --> 00:19:41.197 very powerful decision systems. We believe this is the the the thing that's 00:19:41.280 --> 00:19:42.797 occurring this year and next year. 00:19:42.880 --> 00:19:45.837 Everyone's doing them. They will arrive. 00:19:45.920 --> 00:19:50.877 The agents today speak in English. You can see what they're saying to each other. 00:19:50.960 --> 00:19:54.797 They're not human, but they are communicating what they're doing. 00:19:54.880 --> 00:19:58.717 English to English to English. As long as, and it doesn't have to be English, 00:19:58.800 --> 00:20:01.997 but as long as they're human understandable. But let's So the thought 00:20:02.080 --> 00:20:05.437 experiment is one of the agents says, "I have a better idea. I'm gonna 00:20:05.520 --> 00:20:08.797 communicate in my own language that I'm going to invent that only other agents 00:20:08.880 --> 00:20:12.637 understand. That's a good time to pull the plug. 00:20:12.720 --> 00:20:15.197 >> What is your biggest fear about AI? 00:20:15.280 --> 00:20:18.557 >> My actual fear is different from what you might imagine. My my actual fear is 00:20:18.640 --> 00:20:21.517 that we're not going to adopt it fast enough to solve the problems that affect 00:20:21.600 --> 00:20:24.637 everybody, right? And the reason is that 00:20:24.720 --> 00:20:27.437 the that if you look at every everyone's 00:20:27.520 --> 00:20:32.317 everyday lives, what do they want? They want safety. They want health care. They 00:20:32.400 --> 00:20:35.917 want great schools for their kids. Why don't we just work on that for a while? 00:20:36.000 --> 00:20:39.837 Why don't we make people's lives just better because of AI? We have all these 00:20:39.920 --> 00:20:41.597 other interesting things. Why don't we 00:20:41.680 --> 00:20:45.357 have a um a teacher that is an AI 00:20:45.440 --> 00:20:50.397 teacher that works with existing teachers in this language of the kid in 00:20:50.480 --> 00:20:54.317 the culture of the kid to get the kid as smart as they possibly can. Why don't we 00:20:54.400 --> 00:20:56.077 have a doctor or doctor's assistant 00:20:56.160 --> 00:21:01.357 really that enables a a human doctor to always know every possible best 00:21:01.440 --> 00:21:04.477 treatment and then based on their current situation, what the inventory 00:21:04.560 --> 00:21:07.757 is, which country is, how their insurance works, what is the best way to 00:21:07.840 --> 00:21:11.517 treat that patient. Those are relatively achievable solutions. Why don't we have 00:21:11.600 --> 00:21:13.997 them? If you just did education and 00:21:14.080 --> 00:21:17.037 healthcare globally, the impact in terms 00:21:17.120 --> 00:21:21.117 of lifting human potential up would be so great, 00:21:21.200 --> 00:21:23.597 >> right? that it would change everything. 00:21:23.680 --> 00:21:26.557 It wouldn't solve the various other things that we complain about about, you 00:21:26.640 --> 00:21:31.197 know, this celebrity or this misbehavior or this conflict or even this war, but 00:21:31.280 --> 00:21:35.517 it would establish a level playing field of knowledge and opportunity at a global 00:21:35.600 --> 00:21:39.917 level that has been the dream for decades and decades and decades. If you 00:21:40.000 --> 00:21:41.757 love the D CEO brand and you watch this 00:21:41.840 --> 00:21:43.997 channel, please do me a huge favor. 00:21:44.080 --> 00:21:48.637 Become part of the 15% of the viewers on this channel that have hit the subscribe 00:21:48.720 --> 00:21:51.837 button. It helps us tremendously and the bigger the channel gets, the bigger the 00:21:51.920 --> 00:21:54.080 guests.