WEBVTT 00:00:00.160 --> 00:00:03.348 Eric Schmidt, welcome to Inside Defense. I'm glad to be here. 00:00:03.432 --> 00:00:04.797 I am delighted to be having 00:00:04.880 --> 00:00:08.525 this conversation with you, particularly here in Kiev, Ukraine. Um, 00:00:08.609 --> 00:00:09.917 you will be known to most 00:00:10.000 --> 00:00:13.830 of our viewers as one of the fathers of the modern Silicon Valley. 00:00:13.913 --> 00:00:14.957 You're a former CEO 00:00:15.040 --> 00:00:19.306 of Google. You're now a mentor and investor in some of the leading lights 00:00:19.389 --> 00:00:20.557 of the AI revolution, 00:00:20.640 --> 00:00:25.501 but you've also long been a leader in innovation to national security, 00:00:25.585 --> 00:00:27.197 and you've played a very 00:00:27.280 --> 00:00:30.983 outsized role here in Ukraine. You've built and scaled a drone company. 00:00:31.067 --> 00:00:32.317 You've gone frequently to 00:00:32.400 --> 00:00:37.270 the front lines. On this trip, I went uh to visit the command headquarters 00:00:37.353 --> 00:00:38.877 of Nemesis, which is one 00:00:38.960 --> 00:00:44.434 of the drone brigades uh in Ukraine. It's the subject of another insider show we've 00:00:44.517 --> 00:00:45.037 done. And 00:00:45.120 --> 00:00:50.895 I have to say, you see, you get a glimpse of the future of war and it's pretty 00:00:50.978 --> 00:00:52.397 dystopian. Um war is 00:00:52.480 --> 00:00:57.207 really bad. And now I've seen it, it's much worse than you think it is. 00:00:57.291 --> 00:00:59.037 If you've been watching war 00:00:59.120 --> 00:01:03.837 movies, to watch people really be killed, really struggling, 00:01:03.920 --> 00:01:06.717 the death and destruction of a city, 00:01:06.800 --> 00:01:10.232 little old ladies being killed by the Russians, it's just horrible. 00:01:10.316 --> 00:01:11.597 The goal should be to stop 00:01:11.680 --> 00:01:17.013 it. The only way I know of to stop an aggressor is to fight back. 00:01:17.097 --> 00:01:19.597 And these new technologies have 00:01:19.680 --> 00:01:23.906 both offensive and defensive versions of each others. Um, 00:01:23.989 --> 00:01:29.410 Russia and Ukraine are together in this unfortunate pairing inventing that future. 00:01:29.494 --> 00:01:32.157 And I had always thought that until recently 00:01:32.240 --> 00:01:35.999 Ukraine had the advantage in innovation, but you wrote recently in the Washington 00:01:36.083 --> 00:01:40.556 Post, it was very sobering, that you were more worried than at any time since 00:01:40.639 --> 00:01:42.077 the earliest days of this 00:01:42.160 --> 00:01:46.355 invasion. You said Russia was going towards a strategy of quote total war 00:01:46.438 --> 00:01:47.117 and that that 00:01:47.200 --> 00:01:51.958 was blunting Ukraine's innovative edge. Can you explain what you meant by that? 00:01:52.041 --> 00:01:52.877 Well, I've been 00:01:52.960 --> 00:01:58.331 doing this for more than three years. Um, and I've watched the Ukrainians um, 00:01:58.414 --> 00:01:59.677 work really, really 00:01:59.760 --> 00:02:06.397 hard to innovate and they've held the line largely well at great cost to themselves. 00:02:06.480 --> 00:02:07.597 Um, it's always 00:02:07.680 --> 00:02:14.643 been presumed that Ukraine was the innovator and Russia was big and slow 00:02:14.727 --> 00:02:20.397 and something changed. Russia is now on an innovation path at a slope, if you will, 00:02:20.480 --> 00:02:22.797 of improvement that's similar to that of 00:02:22.880 --> 00:02:28.321 Ukraine, but with much more resources. Here are my data points. Um, 00:02:28.404 --> 00:02:30.877 we have no effective you uh way 00:02:30.960 --> 00:02:34.809 of handling the ballistic missiles that they're using. And indeed, 00:02:34.892 --> 00:02:36.477 the anti-ballistic missiles, 00:02:36.560 --> 00:02:40.843 al generally known as Patriots, uh, in the West are in such short supply that there's 00:02:40.926 --> 00:02:45.843 not enough to go around. And I understand that. If you look at jet shaheds, 00:02:45.927 --> 00:02:48.157 Ukraine sorry Russia invented two 00:02:48.240 --> 00:02:51.949 things and put them together which I can describe separately which has made jet 00:02:52.032 --> 00:02:52.717 shahads that are 00:02:52.800 --> 00:02:57.415 roving around the country the major killers of things that are not on 00:02:57.498 --> 00:02:59.117 the front line um there's 00:02:59.200 --> 00:03:03.932 a talk of a large mobilization which will be more uh human fighters on the Russian 00:03:04.015 --> 00:03:04.637 side against 00:03:04.720 --> 00:03:09.480 Ukraine very hard to fight against there's going to be a new satellite network in 00:03:09.563 --> 00:03:10.317 the air that's 00:03:10.400 --> 00:03:14.533 similar to Starlink it's called Rusevet sometime next year. And finally, 00:03:14.616 --> 00:03:18.815 the Russians are building these very powerful so-called glide bombs, uh, 00:03:18.899 --> 00:03:21.757 which are independently targeted and are keeping 00:03:21.840 --> 00:03:26.749 the front line from moving and really destroying a lot of people. 00:03:26.832 --> 00:03:29.437 It's any one of those Ukraine could 00:03:29.520 --> 00:03:32.681 overcome. The five is going to be, in my view, 00:03:32.765 --> 00:03:35.997 almost impossible without Western help. And the 00:03:36.080 --> 00:03:39.051 Western help, to be very precise, is they need money. 00:03:39.135 --> 00:03:41.357 Russians have now shut down all of their 00:03:41.440 --> 00:03:44.979 commerce, all of their ports. They need money. Literally money. 00:03:45.062 --> 00:03:46.877 You're watching an interview from 00:03:46.960 --> 00:03:51.803 The Economist. To get early access to our upcoming interviews and watch others 00:03:51.886 --> 00:03:56.226 from our archive, subscribe to The Economist by clicking the link above. Now, 00:03:56.310 --> 00:03:58.157 back to the interview. And second, 00:03:58.240 --> 00:04:03.418 they need help on scaling. They have the courage to win that is or hold the fight, 00:04:03.502 --> 00:04:04.317 but they don't 00:04:04.400 --> 00:04:08.034 have the capability of scaling. Yeah, I I I hear that loud and clear, 00:04:08.118 --> 00:04:09.597 but let's just dive down into 00:04:09.680 --> 00:04:12.873 some of those specifically. So jet shahads, you know, 00:04:12.956 --> 00:04:15.037 the Russians took the shahed drone, 00:04:15.120 --> 00:04:18.782 Iranian drone that they now make in huge numbers and added jet propulsion, 00:04:18.865 --> 00:04:19.997 which as I understand it 00:04:20.080 --> 00:04:24.478 means they go faster and higher and much harder to intercept. 00:04:24.562 --> 00:04:27.197 How close is Ukraine to a solution to 00:04:27.280 --> 00:04:32.294 that? There are at least five companies in Ukraine that have come up with solutions 00:04:32.377 --> 00:04:33.277 that are in test 00:04:33.360 --> 00:04:39.890 now. It should be soon. But this was a major event in the war and I didn't fully 00:04:39.973 --> 00:04:40.717 foresee it 00:04:40.800 --> 00:04:45.570 which I hold myself responsible. We knew that the Russians were going to take 00:04:45.653 --> 00:04:46.957 their shades, make the 00:04:47.040 --> 00:04:50.895 back stronger, add these small jet engines which you can get from China. 00:04:50.978 --> 00:04:52.317 We thought they were going 00:04:52.400 --> 00:04:57.095 to use about 10% of the supply and we thought that that was overcomeable. 00:04:57.178 --> 00:04:58.797 Russia accelerated the use 00:04:58.880 --> 00:05:03.760 of these to more than half of the shahads. But they did one other thing. 00:05:03.843 --> 00:05:05.277 They don't have access 00:05:05.360 --> 00:05:09.924 to the equivalent of Starlink for Russian use in Ukraine, thank goodness. 00:05:10.007 --> 00:05:11.197 So they built a mesh 00:05:11.280 --> 00:05:15.496 network out of Chinese parts and they typically have two shahads that are 00:05:15.580 --> 00:05:16.557 sitting in a going 00:05:16.640 --> 00:05:22.385 around in a circle between Bellarusia and Ke for example and those are relays 00:05:22.469 --> 00:05:23.597 and those relays 00:05:23.680 --> 00:05:28.526 allow them to see literally 8K video of whatever their target is. 00:05:28.609 --> 00:05:30.877 So right next door they managed 00:05:30.960 --> 00:05:35.086 to go right through the central central window of the SBU headquarters which 00:05:35.169 --> 00:05:38.804 is like the CIA and that's a good warning of the power of these things. 00:05:38.888 --> 00:05:40.957 There are reports that there are plans to 00:05:41.040 --> 00:05:46.451 use up to 80,000 such jet shahads against Keev and the other towns between now 00:05:46.534 --> 00:05:47.437 and the end of 00:05:47.520 --> 00:05:52.903 December. This is a major point. Um the government here is extremely focused 00:05:52.986 --> 00:05:54.557 on this. Um I know of a 00:05:54.640 --> 00:05:57.883 number of efforts that are likely to work. You mentioned that the Russians, 00:05:57.966 --> 00:05:58.637 and this would be 00:05:58.720 --> 00:06:03.913 even worse, are close to or potentially close to building their version of Starink 00:06:03.997 --> 00:06:04.557 Rasat. Um, 00:06:04.640 --> 00:06:07.999 how close are they? Because it seems to me, and I'm no technician on this, 00:06:08.082 --> 00:06:08.557 but it seems 00:06:08.640 --> 00:06:12.557 that that would be a real gamecher. Um, there's evidence that they are have 00:06:12.640 --> 00:06:16.286 started launching. There's evidence that it doesn't work yet. Uh, 00:06:16.369 --> 00:06:19.357 my own opinion is it's probably a year away because 00:06:19.440 --> 00:06:24.012 these networks are hard. Once that network is in place, 00:06:24.096 --> 00:06:26.637 it is more or less unblockable, 00:06:26.720 --> 00:06:32.682 which means the Russians will be able to again hunt anywhere inside of Ukraine. 00:06:32.765 --> 00:06:37.600 Ukraine today does not have the ability to use Starlink to hunt inside of Russia. 00:06:37.683 --> 00:06:39.357 So you can imagine a situation 00:06:39.440 --> 00:06:43.936 if both countries had that capability, the level of damage that they could do to each 00:06:44.020 --> 00:06:44.637 other. In the 00:06:44.720 --> 00:06:48.555 case of Ukraine, one of their highest priorities is hitting the oil deriffs 00:06:48.638 --> 00:06:49.757 and so forth and so on. 00:06:49.840 --> 00:06:53.456 They do that without really being able to see what they're doing. 00:06:53.539 --> 00:06:55.277 Imagine how much more deadly the 00:06:55.360 --> 00:06:59.297 Ukrainians could be against Russian oil if this if this happen. Well, 00:06:59.380 --> 00:07:00.637 it this begs a question 00:07:00.720 --> 00:07:05.040 which many of the Ukrainian commanders I've spoken to mention. 00:07:05.123 --> 00:07:07.597 Why can't we use Starlink in Russian 00:07:07.680 --> 00:07:11.319 territory? And it's usually come up in the context of ballistic missiles because 00:07:11.402 --> 00:07:12.157 the other, and you 00:07:12.240 --> 00:07:16.263 mentioned it, the other thing that is terrorizing Ukraine is Russian launches 00:07:16.347 --> 00:07:17.437 of ballistic missiles, 00:07:17.520 --> 00:07:20.873 which are, you know, dramatically more powerful, huge damage, 00:07:20.956 --> 00:07:22.957 terrorizing the cities. And they say, 00:07:23.040 --> 00:07:27.023 if you can't hit the arrow, why can't we go after the archer? 00:07:27.107 --> 00:07:29.357 And why can't someone persuade Elon 00:07:29.440 --> 00:07:32.376 Musk to let us for that purpose use Starink in Russia? 00:07:32.459 --> 00:07:34.557 Do you think that would be a good idea? 00:07:34.640 --> 00:07:40.335 Um, I'm in favor of giving Ukraine the tools to allow them to end the war 00:07:40.419 --> 00:07:42.077 by force, literally to 00:07:42.160 --> 00:07:46.309 stop the Russian aggression. Um, the question you're asking is really 00:07:46.392 --> 00:07:47.597 a political decision. 00:07:47.680 --> 00:07:52.209 It's really a government decision. And the US government needs to grapple with 00:07:52.292 --> 00:07:53.037 this question. 00:07:53.120 --> 00:07:59.037 The most important thing right now is to get the innovation cycle in Ukraine equal 00:07:59.120 --> 00:08:05.038 or better than the innovation cycle now present in Russia and also have the scale. 00:08:05.122 --> 00:08:08.277 Ukraine does not have effective anti-bballistic missiles. 00:08:08.360 --> 00:08:10.637 It has one effective ballistic missile which 00:08:10.720 --> 00:08:14.132 is the firepoint of so-called flamingo missile which has done well. 00:08:14.216 --> 00:08:15.437 They can't make that many 00:08:15.520 --> 00:08:19.048 of them for various scaling reasons. In the case of jet shahads, 00:08:19.131 --> 00:08:20.797 there's no Ukrainian equivalent 00:08:20.880 --> 00:08:26.010 of a jet shahad. So again the same problem people are working on this just I mean 00:08:26.094 --> 00:08:26.397 that's 00:08:26.480 --> 00:08:30.049 interesting you say it's a political decision so this is not that something 00:08:30.132 --> 00:08:34.730 that is up to Mr. Musk himself. It would be a political decision in the US. Again, 00:08:34.814 --> 00:08:36.477 I'm speculating. I can answer 00:08:36.560 --> 00:08:40.403 the question precisely. There's no technical restriction that prevents 00:08:40.486 --> 00:08:44.855 the use of any of these satellite services over any country. That's interesting. 00:08:44.938 --> 00:08:46.717 It's it's it's a it's a question. 00:08:46.800 --> 00:08:51.272 It's not it's not like the satellites stop working over Russia because of some 00:08:51.355 --> 00:08:52.557 geographic nature. You 00:08:52.640 --> 00:08:57.496 can get the same outcome if you choose to fly the satellites over Russia and you 00:08:57.579 --> 00:09:03.151 enable them. As I understand it, it is a policy of SpaceX that it does not do that. 00:09:03.234 --> 00:09:05.117 How you would change that is 00:09:05.200 --> 00:09:09.617 a political question. Yeah. So either SpaceX could or the it could be asked 00:09:09.700 --> 00:09:11.117 to do that were there the 00:09:11.200 --> 00:09:15.071 political decision or it could be ordered to do so. In the American system, 00:09:15.155 --> 00:09:15.757 the president 00:09:15.840 --> 00:09:20.797 can order a national security action on a private company and the private company 00:09:20.880 --> 00:09:24.530 pretty much has to obey. And you would advocate that? Uh in general, 00:09:24.613 --> 00:09:27.197 let's talk about what's coming next. I mean that's 00:09:27.280 --> 00:09:31.912 all pretty pretty alarming and that's these are the challenges Ukraine faces now 00:09:31.996 --> 00:09:32.797 but coming down 00:09:32.880 --> 00:09:38.372 the road is you know even greater technical innovation and particularly AI 00:09:38.456 --> 00:09:42.405 and you know in most contexts you and I have conversations about what is happening 00:09:42.488 --> 00:09:43.357 in Silicon Valley not 00:09:43.440 --> 00:09:46.883 what is happening in Ukraine and I wanted to sort of stand back because I think 00:09:46.966 --> 00:09:47.597 there's a bit of 00:09:47.680 --> 00:09:53.680 misunderstanding about what's going on with AI and this war right now people talk 00:09:53.763 --> 00:09:54.957 about is AI being 00:09:55.040 --> 00:10:00.312 used but as I understand it what's really being used is quite well-known technology 00:10:00.395 --> 00:10:00.957 of machine 00:10:01.040 --> 00:10:04.177 vision for final targeting. Could you explain that because I think it's 00:10:04.261 --> 00:10:05.357 an important thing to make 00:10:05.440 --> 00:10:11.781 clear. So in the progression of war you start with the the person with the soldier 00:10:11.865 --> 00:10:15.602 with the bullet with the gun and the bullet and they aim. Um, 00:10:15.685 --> 00:10:18.717 at some point it should be relatively straightforward 00:10:18.800 --> 00:10:23.935 to have the gun be able to see the subject with a laser or what have you, 00:10:24.019 --> 00:10:29.273 identify it, and then the human pulls the button and it always hits. 00:10:29.357 --> 00:10:33.277 That analogy scales up in terms of autonomy. So, 00:10:33.360 --> 00:10:39.771 it's reasonable to expect that every weapon has a final autonomous seg segment 00:10:39.855 --> 00:10:41.437 where basically it's 00:10:41.520 --> 00:10:45.031 close enough, it's near the target, it's going to roughly hit the target. 00:10:45.114 --> 00:10:48.583 You press a button and it hits it precisely. All of these are not AI. 00:10:48.667 --> 00:10:50.717 They're just good computer science, good 00:10:50.800 --> 00:10:55.662 computer engineering. Um, artificial intelligence is systems that learn. 00:10:55.745 --> 00:10:59.512 And artificial intelligence learn like we do. You know, we sort of learn, 00:10:59.596 --> 00:11:01.597 you get smarter and smarter and smarter. You're 00:11:01.680 --> 00:11:05.551 self-improving. The technical term is recursive self-improvement. 00:11:05.634 --> 00:11:07.437 And we're just beginning to see 00:11:07.520 --> 00:11:12.892 the use of AI in war. There's a lengthy study on Russian use of the Chinese models 00:11:12.975 --> 00:11:17.045 to do some of their decision-making. Um, these are all from leaked memos, 00:11:17.128 --> 00:11:18.717 so I don't know what the truth 00:11:18.800 --> 00:11:21.745 is. In Ukraine, there's relatively little use of that. 00:11:21.828 --> 00:11:24.157 There is a project which has been discussed 00:11:24.240 --> 00:11:29.423 by the MOD to try to help automate the planning of of war and planning of where 00:11:29.507 --> 00:11:30.557 people go. Again, 00:11:30.640 --> 00:11:34.238 it's just at the beginning. So, speaking as a computer scientist, 00:11:34.321 --> 00:11:35.597 it's pretty obvious what 00:11:35.680 --> 00:11:39.801 it should look like. Take all the information that you have all the open 00:11:39.885 --> 00:11:41.437 source, all the proprietary, 00:11:41.520 --> 00:11:46.781 all the super secret stuff, feed it into one large computer and say what's going 00:11:46.864 --> 00:11:51.778 on and have these large language models use tools to do reasoning. To me, 00:11:51.861 --> 00:11:54.797 it's it's fantastic. They can think at this 00:11:54.880 --> 00:11:57.586 level. They don't think like humans, by the way. 00:11:57.669 --> 00:12:00.317 They're reasoning around an objective. It makes 00:12:00.400 --> 00:12:05.321 sense to me that eventually you'll say, "How can I best position my assets 00:12:05.405 --> 00:12:10.036 to achieve my objective such as winning the war, defending this city, so forth?" 00:12:10.120 --> 00:12:12.237 But the more interesting questions are, 00:12:12.320 --> 00:12:16.399 why am I losing? Why is this group not as good as the other group? 00:12:16.482 --> 00:12:17.597 You want the system 00:12:17.680 --> 00:12:23.153 to anticipate that and you want the system to say, "General, you have a problem. 00:12:23.236 --> 00:12:27.602 Here is our recommended solution. Can we help you?" That's I think the future. 00:12:27.685 --> 00:12:29.277 That's where we're going. Ju I 00:12:29.360 --> 00:12:32.905 I wanted to delve into that but just before one more question on this sort 00:12:32.988 --> 00:12:33.837 of what's it called 00:12:33.920 --> 00:12:38.664 last mile machine vision which is that in another in an earlier show this year uh 00:12:38.747 --> 00:12:39.677 with my colleague 00:12:39.760 --> 00:12:45.171 Shashank Jooshi Jack um Watling from Russi said that and it was interesting he 00:12:45.254 --> 00:12:46.157 that many many 00:12:46.240 --> 00:12:50.237 um soldiers in Ukraine were turning this capability off because it actually wasn't 00:12:50.320 --> 00:12:57.094 as good as having a a human operator kind of why is that that's called a bug um so I 00:12:57.177 --> 00:12:58.320 fly my own jet 00:12:59.040 --> 00:13:04.666 And uh there's a general rule that you should fly the last segment of your jet 00:13:04.749 --> 00:13:05.837 in coupled mode, 00:13:05.920 --> 00:13:10.170 which means the airplane is flying all the way down to the landing point. 00:13:10.253 --> 00:13:14.870 But I prefer to turn it off because it's more entertaining for me to fly my jet, 00:13:14.954 --> 00:13:16.717 but I'm not in I'm not in a But 00:13:16.800 --> 00:13:20.567 you're not in a foxhole in Ukraine. I'm not in a war. Yeah. Right. 00:13:20.650 --> 00:13:22.317 So in a war, you'd want to use 00:13:22.400 --> 00:13:28.479 a guaranteed solution. So any situation where the terminal guidance as it's called 00:13:28.562 --> 00:13:33.422 is not as good as a human is a bug. And now why? So I'll give you an example. 00:13:33.506 --> 00:13:35.597 I was in a operation center a while 00:13:35.680 --> 00:13:42.680 ago in um very near the nuclear reactor and there was a military target that 00:13:44.118 --> 00:13:46.477 the Ukrainians decided 00:13:46.560 --> 00:13:51.326 to hit. They used a American product called a Switchblade 600. 00:13:51.410 --> 00:13:53.517 We watched it. We saw target 00:13:53.600 --> 00:13:58.054 lock and it missed. And I go, "Wow." And they said, "Well, 00:13:58.137 --> 00:14:00.557 we have one high Mars we can try 00:14:00.640 --> 00:14:05.175 per day." And I said, "Okay." So they launch the high Mars and you wait and then 00:14:05.259 --> 00:14:10.025 they said, "Lock, hit," and it missed. Why? Because of jamming. 00:14:10.108 --> 00:14:14.237 So it may very well be that the anecdotal example 00:14:14.320 --> 00:14:17.671 that you're using is because the human could see the target quite well, 00:14:17.754 --> 00:14:18.397 but the jamming 00:14:18.480 --> 00:14:23.677 confus and with jamming the confus, right? Okay. And we know how to fix that now. So, 00:14:23.760 --> 00:14:26.574 so the important point is that terminal guidance, 00:14:26.657 --> 00:14:29.117 which sort of think of it as the last mile, 00:14:29.200 --> 00:14:33.864 is crucial for having a precision attack. One of the things that's interesting about 00:14:33.948 --> 00:14:34.317 this new 00:14:34.400 --> 00:14:39.634 form of war is there's almost no civilian damage that's because the weapons are so 00:14:39.717 --> 00:14:40.477 accurate. And 00:14:40.560 --> 00:14:45.442 another horrifying fact is that there's almost no survivable injuries. 00:14:45.525 --> 00:14:47.357 The drones are so dangerous 00:14:47.440 --> 00:14:51.781 that the the soldiers that are being killed are being attacked are in fact 00:14:51.864 --> 00:14:53.037 killed. So let's talk 00:14:53.120 --> 00:14:57.238 about what the consequences will be when you have real AI when you have 00:14:57.321 --> 00:14:58.717 this ability to reason to 00:14:58.800 --> 00:15:04.202 infer to learn that that we talk about constantly in Silicon Valley coming here. 00:15:04.286 --> 00:15:08.623 Firstly, what does it mean for how autonomous warfare becomes? 00:15:08.707 --> 00:15:12.477 Um because now you can have this last mile guidance, 00:15:12.560 --> 00:15:17.597 but then presumably you can have the AI not only giving the general some lessons 00:15:17.680 --> 00:15:18.557 but providing a 00:15:18.640 --> 00:15:22.428 battle plan and and down the road, you know, executing them autonomously. 00:15:22.511 --> 00:15:23.277 So let's imagine 00:15:23.360 --> 00:15:29.227 I'm an evil dictator and I wake up one day and I have a perfect AI system and I say, 00:15:29.310 --> 00:15:34.700 "Attack this country. Tell me when you're done." Right? That's my command. 00:15:34.783 --> 00:15:37.360 Notice I didn't say anything else. 00:15:37.920 --> 00:15:41.597 the system would be then free to do whatever it thought was appropriate, 00:15:41.680 --> 00:15:45.694 which could include massive war crimes, right? All sorts of terrible things. 00:15:45.777 --> 00:15:48.752 It could destroy civilians. It could destroy infrastructure. 00:15:48.836 --> 00:15:51.357 It might attack another country that was uninvolved 00:15:51.440 --> 00:15:56.198 in order to annoy the first country. It it could have all sorts of reasoning that 00:15:56.281 --> 00:16:01.030 is not appropriate. It's really important we establish some basic rules about AI 00:16:01.113 --> 00:16:02.317 and war. And the most 00:16:02.400 --> 00:16:06.443 important one is that there's a human involved in every military decision. 00:16:06.526 --> 00:16:07.837 So in other words, attack 00:16:07.920 --> 00:16:11.862 this country, call me when you're done. That's not an appropriate command. 00:16:11.946 --> 00:16:13.277 The appropriate command is 00:16:13.360 --> 00:16:18.101 here as a military target defined by the laws of war in a time bounded time 00:16:18.185 --> 00:16:19.517 and space where I have 00:16:19.600 --> 00:16:25.983 approval to do that. Any other solution in my view becomes scalably unimaginably 00:16:26.066 --> 00:16:27.357 difficult because 00:16:27.440 --> 00:16:31.851 once you start letting the computer make the targeting decisions on its own, 00:16:31.934 --> 00:16:35.420 you have no real quality control on how it makes its decisions. 00:16:35.503 --> 00:16:38.157 You're watching an interview from The Economist. 00:16:38.240 --> 00:16:42.157 If you want to watch our other interviews, including with Steve Bannon and Benjamin 00:16:42.240 --> 00:16:46.614 Netanyahu, subscribe to The Economist by clicking the link above. Now, 00:16:46.697 --> 00:16:48.397 back to the interview. There 00:16:48.480 --> 00:16:55.023 was a recent example um that alarmed our industry where an open AI essentially 00:16:55.106 --> 00:16:56.637 a large set of bots 00:16:56.720 --> 00:17:00.880 that were doing something involving testing took their instructions a bit too 00:17:00.964 --> 00:17:01.597 literally and 00:17:01.680 --> 00:17:06.118 they started to actually attack the system that they were supposed to help and they 00:17:06.201 --> 00:17:06.717 manipulated 00:17:06.800 --> 00:17:11.128 it by lying, cheating and stealing. And everyone studied this as a good example 00:17:11.211 --> 00:17:12.077 as a good sort of 00:17:12.160 --> 00:17:18.011 marker of how AI systems can fail. So, another way to say this is if I were 00:17:18.095 --> 00:17:21.813 an evil dictator, which I hopefully am not, and I said, "Attack this country. 00:17:21.896 --> 00:17:23.437 Let me know what you've done." I'd also 00:17:23.520 --> 00:17:28.339 then have to add, "Follow all the laws of war. Don't hit any civilians, 00:17:28.423 --> 00:17:29.997 use appropriate caution, 00:17:30.080 --> 00:17:34.141 and all the things that human judgment require. Without the human, 00:17:34.225 --> 00:17:35.837 the AI is not going to have 00:17:35.920 --> 00:17:38.319 that judgment." So, you you make that point very powerfully. 00:17:38.403 --> 00:17:39.437 You made it very powerfully 00:17:39.520 --> 00:17:42.475 on a panel at the conference you and I are both at this morning and you said if we go 00:17:42.558 --> 00:17:46.977 down here we'll end up in a Doctor Strange love situation. However, 00:17:47.060 --> 00:17:49.837 we are in a country that is at war with 00:17:49.920 --> 00:17:56.495 a country with a dictator who does seem to show no respect for any rules of law, 00:17:56.579 --> 00:17:58.077 any law at all. And 00:17:58.160 --> 00:18:05.021 so I I wonder whether in the real world of this war it is possible to say we're going 00:18:05.105 --> 00:18:05.757 to follow 00:18:05.840 --> 00:18:09.191 the rules of war because I I was speaking to a couple of the commanders earlier 00:18:09.275 --> 00:18:09.757 this week and 00:18:09.840 --> 00:18:14.677 I asked them about this and they're all really worried about what very powerful AI 00:18:14.760 --> 00:18:15.277 will mean. 00:18:15.360 --> 00:18:20.374 They're worried about completely complete autonomy but they basically say if 00:18:20.457 --> 00:18:26.038 the enemy is doing this we can't not. So evidence is that that's not true. 00:18:26.121 --> 00:18:30.077 evidence in historical context are that there are 00:18:30.160 --> 00:18:33.918 some restraints on behavior in war and you can only look up the history of 00:18:34.002 --> 00:18:34.957 the Nuremberg trials 00:18:35.040 --> 00:18:39.112 to understand what the limits are. There are people held accountable. 00:18:39.195 --> 00:18:43.765 Those people were hanged um after a long trial and so forth and so on. 00:18:43.849 --> 00:18:47.117 I would assume that that principle will survive 00:18:47.200 --> 00:18:53.099 our lives and our really yes I really do. There's this simple logic which says 00:18:53.183 --> 00:18:59.357 that and people say it all the time. If if the two are fighting in a death grip, 00:18:59.440 --> 00:19:01.837 there will be no constraints on 00:19:01.920 --> 00:19:05.582 either. Historically, that has not been true. Historically, 00:19:05.665 --> 00:19:07.677 there has been justice afterwards 00:19:07.760 --> 00:19:12.171 and the people who are participating it are at least aware of their personal 00:19:12.254 --> 00:19:16.706 liability and so forth. There are so obviously sociopaths, but in general, 00:19:16.790 --> 00:19:19.037 people are subject to the law even in 00:19:19.120 --> 00:19:23.299 war. But there are, let's do it in the context of Vladimir Putin in Ukraine. 00:19:23.382 --> 00:19:24.477 Look at what happened 00:19:24.560 --> 00:19:29.595 in BHA at the beginning. Look at the way he is targeting civilians in cities across 00:19:29.678 --> 00:19:30.397 Ukraine. Yes. 00:19:30.480 --> 00:19:34.892 But they're not stopping him. Well, that's because he hasn't been arrested yet. 00:19:34.976 --> 00:19:35.917 But but I guess my 00:19:36.000 --> 00:19:40.614 my point is about the logic of this escalation because presumably the more 00:19:40.697 --> 00:19:44.922 autonomy you allow an AI as we go up from where we are right now up this scale, 00:19:45.005 --> 00:19:46.557 the more effective you will be. 00:19:46.640 --> 00:19:49.573 there is a first mover advantage because AI thinks faster. 00:19:49.656 --> 00:19:51.757 It's able to come up with a better quicker 00:19:51.840 --> 00:19:58.206 solution. So if the Russians have fewer limits and go faster up that escalatory 00:19:58.289 --> 00:20:02.480 ladder, Ukraine may have, you know, international law on its side, 00:20:02.563 --> 00:20:05.677 but it will be in the here and now less able to defeat 00:20:05.760 --> 00:20:09.158 them. The solution will not be to commit war crimes. 00:20:09.242 --> 00:20:11.837 The solution will be is to help Ukraine. 00:20:11.920 --> 00:20:17.683 I I'm not going to yield on my core point that there is a boundary to what humans 00:20:17.766 --> 00:20:18.477 can do. So, 00:20:18.560 --> 00:20:22.325 let's use a thought experiment. Let's imagine that Ukraine had some secret 00:20:22.408 --> 00:20:23.677 biological weapon, which I 00:20:23.760 --> 00:20:26.527 don't think they do, but let's say they did. I'm making that up. 00:20:26.610 --> 00:20:27.997 And furthermore, they were forced 00:20:28.080 --> 00:20:33.638 to use it against some Russian attack. Would they be prosecuted? Absolutely. 00:20:33.721 --> 00:20:35.197 And I'm sure they are 00:20:35.280 --> 00:20:38.813 aware of it. Yeah. I I'm not I'm not um I'm not asking you to yield on 00:20:38.897 --> 00:20:40.157 a principle. I'm just I am 00:20:40.240 --> 00:20:45.770 very struck particularly on this trip at the pace of innovation escalation and so I 00:20:45.853 --> 00:20:49.142 think there is a risk that this develops a logic of its own. 00:20:49.225 --> 00:20:51.837 So so that's why it's important that humans debate 00:20:51.920 --> 00:20:56.539 this stuff. We you and I agree very strongly that the rate of innovation 00:20:56.622 --> 00:20:58.237 in this war is faster than 00:20:58.320 --> 00:21:02.531 has ever occurred. It's faster than the rate of innovation that I have seen in 00:21:02.614 --> 00:21:06.449 my many decades in tech. And the reason is because lives are at stake. 00:21:06.532 --> 00:21:08.797 And it turns out it's true on both sides. 00:21:08.880 --> 00:21:13.469 The other side is frankly more evil and better funded and has more resources 00:21:13.552 --> 00:21:14.637 and they're willing 00:21:14.720 --> 00:21:19.444 to accept more death than the Ukrainian side. But it's still a fight. 00:21:19.528 --> 00:21:24.751 And it's very important, I'll say again, that we help Ukraine fight this war. 00:21:24.835 --> 00:21:27.757 What they need is money and help in scaling 00:21:27.840 --> 00:21:31.460 and technology. They can fight the war. We don't need to invade. 00:21:31.544 --> 00:21:33.197 We don't need to send a lot of 00:21:33.280 --> 00:21:36.680 troops and all that kind of stuff people, but they need some help. 00:21:36.764 --> 00:21:41.255 They won't make it without some level of help, at least economic. Yeah. So, 00:21:41.338 --> 00:21:44.877 the innovations in drone warfare have gone beyond the 00:21:44.960 --> 00:21:48.050 theater of this war. I mean, just in the last few days, we've seen, you know, 00:21:48.133 --> 00:21:48.957 we've seen the Houthis 00:21:49.040 --> 00:21:54.396 effectively shut down uh another exit route for oil products in the Gulf. Um, 00:21:54.479 --> 00:21:55.597 do you think that 00:21:55.680 --> 00:22:02.337 the US military recognizes the degree to which the nature of war has shifted and 00:22:02.420 --> 00:22:05.876 is acting fast enough to transform itself to prevail? Well, 00:22:05.960 --> 00:22:08.397 I worked for the US military for eight years in 00:22:08.480 --> 00:22:14.180 various under various presidents and there are plenty of people who understand it 00:22:14.263 --> 00:22:15.037 and they can 00:22:15.120 --> 00:22:19.685 articulate it. They understand this conversation and they're worried about it. 00:22:19.768 --> 00:22:20.877 The system is almost 00:22:20.960 --> 00:22:25.572 unchangeable. I was talking to one president about it and he said not even 00:22:25.655 --> 00:22:26.397 the president 00:22:26.480 --> 00:22:31.477 can fix the system and I said well that's a very bad answer and the reason is 00:22:31.560 --> 00:22:32.797 that the military is 00:22:32.880 --> 00:22:37.648 not run by a single person not even the president although you think they are it's 00:22:37.732 --> 00:22:38.477 run by the way 00:22:38.560 --> 00:22:42.877 the congress works the way the special interest works the way the lobbying works 00:22:42.960 --> 00:22:43.757 so my conclusion 00:22:43.840 --> 00:22:48.456 is the only way to address this problem in the US military is to allow the US 00:22:48.539 --> 00:22:49.677 military to continue 00:22:49.760 --> 00:22:53.723 to spend money on the things that have been organized in the various politically 00:22:53.806 --> 00:22:54.077 correct 00:22:54.160 --> 00:22:59.154 states and uh and so forth and just let that happen while a relatively small 00:22:59.238 --> 00:23:00.157 amount of money 00:23:00.240 --> 00:23:04.921 um is allocated to this new stuff. The new stuff is far more effective, 00:23:05.004 --> 00:23:06.397 far less expensive and 00:23:06.480 --> 00:23:11.948 I think will work and that over time even the kerogens on the old side will say look 00:23:12.031 --> 00:23:16.331 I really want that new thing because they are sincerely all of them sincerely want 00:23:16.414 --> 00:23:17.757 to protect America. There's 00:23:17.840 --> 00:23:21.419 no one who want who who's a a traitor in the military. 00:23:21.503 --> 00:23:23.997 They just disagree on the tactics. And 00:23:24.080 --> 00:23:29.638 what I've learned about militaries is that they're incredibly uh change resistant, 00:23:29.722 --> 00:23:30.877 riskresistant. And 00:23:30.960 --> 00:23:35.091 the most one possibility is they didn't grow up in an innovative environment, 00:23:35.175 --> 00:23:36.077 which is possible. 00:23:36.160 --> 00:23:40.146 A there's a long paper that analyzed this that I read and it basically said at 00:23:40.230 --> 00:23:40.877 the end of the 00:23:40.960 --> 00:23:45.857 day people in the military are more conservative because their loss aversion 00:23:45.940 --> 00:23:47.037 is greater because 00:23:47.120 --> 00:23:51.388 people die. So if you think about it, if you have a pretty certain model of 00:23:51.472 --> 00:23:52.317 an expensive way 00:23:52.400 --> 00:23:56.367 to do something and Eric shows up with a cheaper way to do the same thing, 00:23:56.451 --> 00:23:57.517 you would never adopt 00:23:57.600 --> 00:24:01.550 it because you're certain the more expensive one works and you can't take 00:24:01.634 --> 00:24:02.877 the risk. I think that's 00:24:02.960 --> 00:24:06.515 the best way to understand what's going on in addition to the political maybe it 00:24:06.598 --> 00:24:07.197 takes and maybe 00:24:07.280 --> 00:24:12.152 the lesson of Ukraine is you know first of all it happened in after an invasion here 00:24:12.235 --> 00:24:13.037 but secondly it 00:24:13.120 --> 00:24:18.067 takes people from outside the military I mean the striking thing all of these drone 00:24:18.150 --> 00:24:19.037 brigades are led 00:24:19.120 --> 00:24:22.857 by people who in a former life were civilians tech you know tech leaders 00:24:22.940 --> 00:24:24.237 business leaders they took 00:24:24.320 --> 00:24:28.557 all the traditional Soviet generals and they all retired them for the same reason 00:24:28.640 --> 00:24:29.357 so that because 00:24:29.440 --> 00:24:32.175 it it it I mean you've given a very good answer about the United states. 00:24:32.259 --> 00:24:33.037 I'm wondering what you 00:24:33.120 --> 00:24:37.122 think about Europe's recognition of the scale of which war change because in some 00:24:37.205 --> 00:24:40.396 ways Europe is more vulnerable. I mean it's right here on our doorstep. 00:24:40.479 --> 00:24:41.917 So the Europeans I've spoken with 00:24:42.000 --> 00:24:45.410 all have the same problem that the US people do. They're very smart. 00:24:45.493 --> 00:24:46.797 They understand the problem 00:24:46.880 --> 00:24:52.418 but the system is not organized around change. In most European countries, 00:24:52.501 --> 00:24:54.317 the militaries operate as 00:24:54.400 --> 00:24:57.218 jobs programs. And those are important jobs. I'm not taking away. 00:24:57.301 --> 00:24:58.557 But they're building the wrong 00:24:58.640 --> 00:25:03.613 thing. So here's my deal. You can keep the exact budget that you have. In fact, 00:25:03.696 --> 00:25:04.637 you can increase 00:25:04.720 --> 00:25:08.987 it. Just build more effective things because you're going to need them. 00:25:09.071 --> 00:25:10.397 For example, in Europe, 00:25:10.480 --> 00:25:14.323 there's a huge problem with missile defense. And for all sorts of reasons, 00:25:14.406 --> 00:25:17.745 the Patriot missiles, which they rely on, are not available enough. 00:25:17.829 --> 00:25:20.157 We now understand looking at missile stocks and 00:25:20.240 --> 00:25:23.771 so forth on the Western side, we don't have enough of them. Now, 00:25:23.854 --> 00:25:24.957 since we don't have a 00:25:25.040 --> 00:25:28.682 lot more money to spend because these economies are struggling with debt and so 00:25:28.765 --> 00:25:29.437 forth and so on, 00:25:29.520 --> 00:25:34.747 you have to come up with new solutions. I show up with an argument that I have 00:25:34.831 --> 00:25:35.837 cheaper, better, 00:25:35.920 --> 00:25:42.282 faster, just use my architecture with all the companies that represent new war 00:25:42.366 --> 00:25:43.357 and spend the 00:25:43.440 --> 00:25:46.768 same amount of money. The only reason that doesn't occur is because the incumbents 00:25:46.851 --> 00:25:49.997 don't want to make the change. I think that's one powerful reason. 00:25:50.081 --> 00:25:52.637 I think the other reason is that you're, you know, 00:25:52.720 --> 00:25:55.170 we're now a bit skeptical of American companies. 00:25:55.254 --> 00:25:57.757 We're a bit worried about that. I understand. And 00:25:57.840 --> 00:26:03.654 um it doesn't employ many people. Um I can assure you that if you were to take all 00:26:03.737 --> 00:26:07.553 the people in the militaries in Germany, the UK, and so forth, 00:26:07.637 --> 00:26:10.717 and you simply reassign them to an architecture that 00:26:10.800 --> 00:26:13.729 was AI, automatic targeting, missiles and drones, 00:26:13.812 --> 00:26:16.557 and some ships to move them around, you'd have 00:26:16.640 --> 00:26:20.876 full employment. Every job would change, but every job would still exist. 00:26:20.959 --> 00:26:24.986 And European labor laws don't allow you to lay off people anyway. So trust me, 00:26:25.069 --> 00:26:27.277 those people will still have jobs. What about 00:26:27.360 --> 00:26:31.327 the idea that at least for Europe, Ukraine becomes the vehicle through this which 00:26:31.410 --> 00:26:32.077 this happens? I 00:26:32.160 --> 00:26:35.133 mean, that struck me so much on this trip that we have the solution for Europe 00:26:35.216 --> 00:26:35.917 sitting here staring 00:26:36.000 --> 00:26:39.314 us in the face. I have said that publicly and I've said that to my Ukrainian friends 00:26:39.397 --> 00:26:42.463 because I want to encourage them to sell to Europe. 00:26:42.547 --> 00:26:45.997 Every European government I talk to says, "Your ideas 00:26:46.080 --> 00:26:51.772 are great. it needs to be built here. So the most likely scenario in that scenario 00:26:51.856 --> 00:26:55.512 is Ukraine will license their technology to the European countries, 00:26:55.595 --> 00:26:57.517 which they should. They'll get a nice 00:26:57.600 --> 00:27:02.415 fee and the Europeans will build them at twice the price of eur of Ukrainians. 00:27:02.498 --> 00:27:06.822 And that's okay in a rich democracy, which we can debate whether it's rich or not. 00:27:06.905 --> 00:27:08.397 But the part of the reason the 00:27:08.480 --> 00:27:11.462 Ukrainian stuff is so effective is the people who work so hard. 00:27:11.546 --> 00:27:13.117 When you go to the companies here, 00:27:13.200 --> 00:27:16.576 so and I've invested and worked with them pretty closely. 00:27:16.659 --> 00:27:21.418 There are it looks like China, you have and I mean that in the work ethic sense. 00:27:21.502 --> 00:27:24.317 You have these long pe long buildings which are 00:27:24.400 --> 00:27:28.678 often underground now of people in white jackets sitting shoulders shoulder. 00:27:28.762 --> 00:27:32.608 They work for 12 hours a day, six or seven days a week, which is a Chinese. 00:27:32.691 --> 00:27:34.477 That's more than China. China's 966, 00:27:34.560 --> 00:27:37.340 right? So it's the work ethic is extraordinary. And they're all men. 00:27:37.423 --> 00:27:42.157 And do you know why they're all men? because if you work in such a factory, 00:27:42.240 --> 00:27:45.837 you get an exemption from the draft. I I didn't 00:27:45.920 --> 00:27:49.423 know they were all men. Um there are a lot of women involved in in elements of 00:27:49.506 --> 00:27:53.757 this too, but but what's noticeable is I said there why and they're all, you know, 00:27:53.840 --> 00:27:57.289 men of military age. And the answer is why do they work so hard? 00:27:57.373 --> 00:27:59.277 If they get fired, they go to the go 00:27:59.360 --> 00:28:04.151 to the front. Interesting. You mentioned China and I wanted to to ask you about how 00:28:04.234 --> 00:28:04.797 what you've 00:28:04.880 --> 00:28:11.504 seen happen here in terms of techled innovations in warfare has changed 00:28:11.587 --> 00:28:13.677 your view of how the US 00:28:13.760 --> 00:28:18.487 and China should or how this race technological race would will will 00:28:18.570 --> 00:28:20.397 develop. I mean you've been 00:28:20.480 --> 00:28:24.873 you played a huge role in this thinking through how to ensure that the US stays 00:28:24.957 --> 00:28:25.837 ahead in AI. Does 00:28:25.920 --> 00:28:30.136 what you've seen here change your perception of where that will go, 00:28:30.219 --> 00:28:31.997 what China will end up doing, 00:28:32.080 --> 00:28:37.626 and who will be ahead? And my general belief is that we're extremely fortunate 00:28:37.710 --> 00:28:38.637 that China has 00:28:38.720 --> 00:28:43.117 been copying what we're doing because what we're doing needs to change and therefore 00:28:43.200 --> 00:28:43.757 they'll have 00:28:43.840 --> 00:28:48.856 hysteresis in terms of changing. Um, if you think about it for a while, 00:28:48.939 --> 00:28:53.212 China already has the world's largest drone manufacturer, which is DJI, 00:28:53.296 --> 00:28:55.757 which America has just recently banned. Um, 00:28:55.840 --> 00:28:58.112 there's every reason to think, and I don't know this, 00:28:58.196 --> 00:29:02.490 that they have a military equivalent, or maybe there's a military sub subdivision. 00:29:02.574 --> 00:29:05.597 The reason to think that we are better off than we 00:29:05.680 --> 00:29:08.439 think is that China's not fought a war in 50 years. 00:29:08.522 --> 00:29:10.557 There's no reason to believe that they 00:29:10.640 --> 00:29:14.591 understand the very subtle tactics. When you met with the commanders and when I 00:29:14.674 --> 00:29:19.155 talked to them, when you actually ask them what are they actually doing, 00:29:19.238 --> 00:29:21.597 their reasoning is worldclass subtle 00:29:21.680 --> 00:29:25.498 thinking. I want to do this, this, this, this will happen and so forth and so on. 00:29:25.581 --> 00:29:26.317 You learn that by 00:29:26.400 --> 00:29:30.053 doing. There's every reason to think that China has no experience with that. 00:29:30.137 --> 00:29:35.713 There's no doctrine. There's no books. You can't study it. So, we may be okay. Um, 00:29:35.797 --> 00:29:38.397 with respect to the uh the Pacific, 00:29:38.480 --> 00:29:43.522 the Pacific area and so forth, there's a publicly discussed plan to protect 00:29:43.605 --> 00:29:44.957 the Taiwan Strait and 00:29:45.040 --> 00:29:49.997 so forth. I don't know, but I'm assuming that that will ultimately be a drone war 00:29:50.080 --> 00:29:50.557 and not a 00:29:50.640 --> 00:29:55.117 ship war. If you think about it, you don't really want to put a ship that close 00:29:55.201 --> 00:29:56.157 to China. You want 00:29:56.240 --> 00:29:59.727 to put a drone ship with drones in it because you don't want the people to be 00:29:59.810 --> 00:30:03.767 killed. So, it makes sense to me that all the arguments that we see here are also 00:30:03.850 --> 00:30:04.637 applicable to the 00:30:04.720 --> 00:30:10.637 Pacific. Big ships, big aircraft carriers are easy to hit, easy to blow up, 00:30:10.720 --> 00:30:11.997 and so forth. Um, 00:30:12.080 --> 00:30:17.437 drone ships are are harder. Drones are are harder to detect. You see this in Iran 00:30:17.520 --> 00:30:21.075 um with the problems we're having keeping the straight open. 00:30:21.158 --> 00:30:23.197 Um I think that uh my understanding 00:30:23.280 --> 00:30:28.219 again this is public information that China has a very large set of missiles 00:30:28.303 --> 00:30:29.277 that are land to 00:30:29.360 --> 00:30:33.934 sea. We would need to have anti-missiles. I don't know if we do or not that could 00:30:34.018 --> 00:30:34.797 that could deal 00:30:34.880 --> 00:30:39.632 deal with that threat. So if you it maps pretty well to the current belief in 00:30:39.715 --> 00:30:40.720 the China fight. 00:30:41.440 --> 00:30:47.223 What about could could the military area be an area where you might between China 00:30:47.306 --> 00:30:48.237 and the United 00:30:48.320 --> 00:30:52.835 States be able to reach some agreement over norms to go back to where you were 00:30:52.918 --> 00:30:53.837 earlier? I wonder 00:30:53.920 --> 00:30:58.155 because there's a lot of talk now given the pace at which AI capabilities are 00:30:58.238 --> 00:30:59.277 increasing very real 00:30:59.360 --> 00:31:03.144 fears of you know 100% recursive self-improvement and outofc control AI 00:31:03.227 --> 00:31:04.397 and people are debating 00:31:04.480 --> 00:31:08.825 whether you know the US should pause whether or companies should pause whether 00:31:08.908 --> 00:31:09.677 there should be 00:31:09.760 --> 00:31:12.881 some regulation and there are many who say we can't do that because China will be 00:31:12.964 --> 00:31:13.277 ahead well 00:31:13.360 --> 00:31:17.849 first place we're not going to do a pause because the entire incentive system 00:31:17.932 --> 00:31:19.037 in America is not to 00:31:19.120 --> 00:31:24.051 pause and furthermore if you do a pause you do really open it up to competitive 00:31:24.135 --> 00:31:28.098 either local, you know, a local competitor violates the rules or the global. 00:31:28.181 --> 00:31:29.917 So there's not going to be a pause. 00:31:30.000 --> 00:31:34.077 The right thing to do is to invent this future while also invent the solutions 00:31:34.160 --> 00:31:38.127 to the problems that you're creating. It's just occurring quickly. 00:31:38.210 --> 00:31:41.117 Um so an example here is there's some evidence 00:31:41.200 --> 00:31:48.200 that you can take a new model that you um shows up and use the previous one which 00:31:48.516 --> 00:31:49.517 is typically 00:31:49.600 --> 00:31:54.592 four months older to supervise the new model. It looks like that works. 00:31:54.676 --> 00:31:56.237 People are trying that. 00:31:56.320 --> 00:32:01.087 We'll see if that's technically true at scale. Um, if you look at China, 00:32:01.171 --> 00:32:02.637 there is no question we 00:32:02.720 --> 00:32:09.029 are in an an AI race that is profound between the Chinese approach and 00:32:09.112 --> 00:32:11.677 the American approach. You're 00:32:11.760 --> 00:32:15.868 watching an interview from The Economist. If you want to watch our other interviews, 00:32:15.951 --> 00:32:16.317 including 00:32:16.400 --> 00:32:20.824 with Mark Carney and Dario Amade, subscribe to The Economist by clicking 00:32:20.908 --> 00:32:22.077 the link above. Now, 00:32:22.160 --> 00:32:25.906 back to the interview. And the easiest way to understand that is would you prefer 00:32:25.989 --> 00:32:31.267 to have your day-to-day life intermediated by an American model or a Chinese model? 00:32:31.350 --> 00:32:35.877 I think as westerners we prefer an American or American and British model. 00:32:35.960 --> 00:32:38.877 So that to me is the interesting fight. There's a 00:32:38.960 --> 00:32:41.990 series of uh track two and track one dialogues. 00:32:42.073 --> 00:32:45.037 I'm participated in one track two. There was a 00:32:45.120 --> 00:32:49.551 big meeting in Vienna where both sides are now talking this through. 00:32:49.634 --> 00:32:51.277 The simplest thing that we 00:32:51.360 --> 00:32:55.943 need to do is to avoid surprises. Henry Kissinger, our mutual friend, 00:32:56.026 --> 00:32:57.837 who we miss dearly, was very 00:32:57.920 --> 00:33:03.576 concerned that a single event would spiral in the same way that World War I did 00:33:03.660 --> 00:33:04.957 from the killing of 00:33:05.040 --> 00:33:10.823 the Arch Duke into a total conflration that that once the process started because 00:33:10.906 --> 00:33:11.837 of nationalism 00:33:11.920 --> 00:33:17.099 and tensions and lack of communications, you can end up in a horrible, 00:33:17.183 --> 00:33:21.397 horrible war. A war with China would be just terrible, right? 00:33:21.480 --> 00:33:24.717 It would destroy so much of our modern lives in 00:33:24.800 --> 00:33:30.163 and all of those kinds of But should there be I mean I obviously completely agree 00:33:30.246 --> 00:33:34.758 with you that a war with China would be horrific, possibly existential, 00:33:34.841 --> 00:33:36.717 but do you think there is room 00:33:36.800 --> 00:33:43.800 for some kind of limitations, mutually agreed limitations on AI between China 00:33:44.043 --> 00:33:45.277 and the United 00:33:45.360 --> 00:33:49.887 States? I mean, we have the meeting coming up between President Xi and President 00:33:49.971 --> 00:33:50.637 Trump. And A, 00:33:50.720 --> 00:33:54.903 do you think it's sensible? And B, do you think that the military area in a sort 00:33:54.987 --> 00:33:59.574 of modern-day equivalent of nuclear arms agreements would be an area to start? So, 00:33:59.658 --> 00:34:04.585 we all believe that there needs to be a set of treaties in this area. 00:34:04.669 --> 00:34:08.877 The first treaty needs to be one about no surprises. 00:34:08.960 --> 00:34:15.427 So imagine a situation where an agent bot manages to get out of its testing range 00:34:15.510 --> 00:34:16.317 by accident 00:34:16.400 --> 00:34:21.091 um and but ultimately triggers a nuclear response because it signals that there's 00:34:21.175 --> 00:34:24.353 one coming and then the nuclear response happens automatically. 00:34:24.436 --> 00:34:26.637 There there are plenty of such scenarios that 00:34:26.720 --> 00:34:31.455 analysts have come up with. So the single most important thing is a no surprises 00:34:31.539 --> 00:34:36.127 rule, right? And the next one would be to attack the question of cyber. 00:34:36.210 --> 00:34:38.077 Neither side has been willing 00:34:38.160 --> 00:34:42.142 to limit its use of cyber attack, offensive or defensive capability. 00:34:42.226 --> 00:34:46.686 If you can't agree to that, it's going to be hard to agree to much more. And so, 00:34:46.769 --> 00:34:49.277 and you don't think even in principle a mutual 00:34:49.360 --> 00:34:56.360 agreement on pausing would make sense? I think it's the problem with pausing 00:34:57.074 --> 00:35:02.047 is first place um it's counter to the incentives of all the players of course. 00:35:02.131 --> 00:35:04.157 The second is how do you verify? Okay. 00:35:04.240 --> 00:35:09.414 So, for purposes of argument, the UN will in will administer the pause. So, 00:35:09.498 --> 00:35:10.957 that that's now you're 00:35:11.040 --> 00:35:14.305 really condemning it to failure. No, but we have an international body and they're 00:35:14.388 --> 00:35:18.103 supposed to be helping and both are on the security council. Okay. So, 00:35:18.187 --> 00:35:20.157 now how do they verify it? So, then 00:35:20.240 --> 00:35:24.047 they have to go to the secret labs which includes the military secret labs and they 00:35:24.130 --> 00:35:24.797 have to read all 00:35:24.880 --> 00:35:28.240 the code. Well, first you can't read all the code because it's all written 00:35:28.323 --> 00:35:29.357 by computers. So now you 00:35:29.440 --> 00:35:33.797 have to take your computer system from the UN and you have to apply it to the Chinese 00:35:33.881 --> 00:35:34.477 system. Well, 00:35:34.560 --> 00:35:40.494 the UN's system is not as fast or modern if you will than the Chinese or the US one 00:35:40.578 --> 00:35:43.905 by definition because they don't have all the resources. 00:35:43.989 --> 00:35:47.197 So you see how um I wrote a paper a while ago which said 00:35:47.280 --> 00:35:52.556 there is a solution to this and the solution is to have what we called mutual 00:35:52.639 --> 00:35:58.557 AI destruction. And the idea was that as we go up and up this competition list, 00:35:58.640 --> 00:36:01.997 you have to give the other side a chance to 00:36:02.080 --> 00:36:06.589 say you've gone too far. So we proposed that rather than bombing each other's data 00:36:06.672 --> 00:36:07.037 centers, 00:36:07.120 --> 00:36:10.884 we would just have a computer that would shut down the power to the data center 00:36:10.967 --> 00:36:11.517 in such a way 00:36:11.600 --> 00:36:15.654 it couldn't be restored for a while. So when when the other side got farther 00:36:15.737 --> 00:36:20.694 along, you could shut off its and and they could obviously do the same thing to you. 00:36:20.777 --> 00:36:22.957 This was intended as a serious point 00:36:23.040 --> 00:36:26.216 and but it's also sort of far sounds like a modern day Doctor Strange level. 00:36:26.299 --> 00:36:27.117 But but the important 00:36:27.200 --> 00:36:31.832 point is that it in this particular formulation it would allow when somebody 00:36:31.915 --> 00:36:33.197 got well ahead of you, 00:36:33.280 --> 00:36:36.919 you could shut them down and then they could shut you down and then you restart. 00:36:37.002 --> 00:36:37.757 That's another way 00:36:37.840 --> 00:36:42.957 of addressing this pause issue. I mean, we need a solution that's implementable for a 00:36:43.040 --> 00:36:46.902 pause and I don't see it. We we do. That's the alarming thing you just said. 00:36:46.986 --> 00:36:50.727 We need a solution that's implementable and you don't see it. 00:36:50.810 --> 00:36:54.237 And I and to conclude I I as you know I'm we know I 00:36:54.320 --> 00:36:57.477 know we know each other reasonably well. I'm basically an optimist and I'm always 00:36:57.560 --> 00:36:57.757 looking 00:36:57.840 --> 00:37:02.002 on the upside. But I have to say in the last few weeks my general optimism has 00:37:02.086 --> 00:37:03.037 been a bit knocked. 00:37:03.120 --> 00:37:08.278 firstly by what's been happening in Silicon Valley and the the fact that there 00:37:08.362 --> 00:37:09.757 are growing numbers of 00:37:09.840 --> 00:37:14.890 researchers in these companies themselves alarmed about the dynamic of not being 00:37:14.973 --> 00:37:15.917 able to pause as 00:37:16.000 --> 00:37:19.744 you say and secondly by what I've seen here which is even we're well we're 00:37:19.827 --> 00:37:21.037 nowhere near the frontier 00:37:21.120 --> 00:37:25.773 of AI that we could be at in a few years and yet you see this escalation 00:37:25.856 --> 00:37:26.957 of innovation that 00:37:27.040 --> 00:37:31.453 seems to me to be ever more lethal. So so the counterargument is to be optimistic. 00:37:31.536 --> 00:37:31.837 Give me 00:37:31.920 --> 00:37:37.677 some reasons for optimism. You're building a kind of intelligence that's nonhuman but 00:37:37.760 --> 00:37:41.977 under our control. Can we keep it under control? I believe we can. 00:37:42.060 --> 00:37:43.997 Partly because we at some level 00:37:44.080 --> 00:37:47.489 can pull the plug. Uh you always have the issue of a terrorist group. 00:37:47.572 --> 00:37:48.957 Let's assume terrorists don't 00:37:49.040 --> 00:37:52.483 have access to this because they can't get enough data centers to do it. 00:37:52.567 --> 00:37:53.757 Then diplomacy between the 00:37:53.840 --> 00:37:59.328 countries should allow us to keep effective human control on these systems. 00:37:59.411 --> 00:38:04.173 They will move quickly, but fundamentally they'll be under human control. 00:38:04.256 --> 00:38:07.117 I honestly believe that as long as AI is under 00:38:07.200 --> 00:38:11.597 human control and under human alignment, it will be okay. That is my opinion. Well, 00:38:11.680 --> 00:38:17.553 I have to say I really, really hope you're right. Eric Schmidt, 00:38:17.636 --> 00:38:21.040 thank you very much. Thank you, Zan.