WEBVTT 00:00:00.000 --> 00:00:02.877 But, you've been saying recently that, you know, one of the things that's 00:00:02.960 --> 00:00:05.037 inevitable is some kind of an AI 00:00:05.120 --> 00:00:07.877 disaster, and we're hoping that it's 00:00:07.960 --> 00:00:10.037 like 100 people that die, 00:00:10.120 --> 00:00:15.037 and not 1,000 or 10,000 or a million or or even 100 million. 00:00:15.120 --> 00:00:18.717 Uh but, you know, suppose that it does play out that there's a catastrophe 00:00:18.800 --> 00:00:23.157 that's a wake-up call. What do you want to do with that wake-up 00:00:23.240 --> 00:00:26.397 call? Um the one I'm particularly worried about is biological, and it goes 00:00:26.480 --> 00:00:30.397 something like this. You take some existing pathogen, and using biological 00:00:30.480 --> 00:00:34.317 techniques, which I won't discuss, you can modify it enough that it cannot be 00:00:34.400 --> 00:00:37.197 detected, but it's still quite dangerous. That's an example of a 00:00:37.280 --> 00:00:42.117 threat. Guys, Eric Schmidt just gave a pretty shocking interview about AI. In 00:00:42.200 --> 00:00:45.757 this interview, he talked about where AI is headed right now, and how dangerous 00:00:45.840 --> 00:00:49.637 it's it's going to get in the next 2 years. He warned that if governments of 00:00:49.720 --> 00:00:52.077 the US and China don't put any 00:00:52.160 --> 00:00:54.517 safeguards around AI, it's going to take 00:00:54.600 --> 00:00:59.277 over literally everything. Eric also talked about AI disasters, which could 00:00:59.360 --> 00:01:03.597 destroy our world if our governments don't regulate this industry right now. 00:01:03.680 --> 00:01:07.197 Guys, the audio of this interview is a little bit choppy in the first 2 00:01:07.280 --> 00:01:11.277 minutes, but I guarantee that this interview is a real mind-opener. Let's 00:01:11.360 --> 00:01:14.077 watch the interview, and I'll explain everything as we go. 00:01:14.160 --> 00:01:17.197 >> So, are we going to win the AI race, and you know, what are the scenarios where 00:01:17.280 --> 00:01:21.957 we win and lose? It damn looks like we will. And uh let 00:01:22.040 --> 00:01:25.917 me define the way And so, I think the the separate that's 00:01:26.000 --> 00:01:28.837 so consensus, which is I call it, is 00:01:28.920 --> 00:01:33.037 what people in separate that's so believe, which charitably was you. 00:01:33.120 --> 00:01:34.037 Uh 00:01:34.120 --> 00:01:37.637 Kids, that I you only just be built from 00:01:37.720 --> 00:01:42.317 current magnetic computing to various forms of recursive 00:01:42.400 --> 00:01:43.837 self-improvement 00:01:43.920 --> 00:01:47.397 to eventual AGI and superintelligence. 00:01:47.480 --> 00:01:49.597 Um in order to do that, it requires that 00:01:49.680 --> 00:01:52.357 you use the enormous amount of hardware. 00:01:52.440 --> 00:01:56.357 Uh Google TPUs, the biggest chips, and so when everybody in the audience knows 00:01:56.440 --> 00:01:59.677 us that it sure looks like the hardware 00:01:59.760 --> 00:02:02.317 restrictions that the Trump and Biden 00:02:02.400 --> 00:02:05.037 administrations have put on China are 00:02:05.120 --> 00:02:09.477 going to prevent them from competing at that space. I was recently in Shanghai 00:02:09.560 --> 00:02:12.957 for a few days. I have good relationships with the Chinese. 00:02:13.040 --> 00:02:16.477 And my conclusion is they're they're they're fighting a different game. 00:02:16.560 --> 00:02:18.797 They're regular about AI in every 00:02:18.880 --> 00:02:21.877 product, every service, everything, but 00:02:21.960 --> 00:02:26.317 in a more classical way. Whereas America's been a secret for AGI. 00:02:26.400 --> 00:02:29.917 And I was quite worried that we would end up in a superintelligence phrase 00:02:30.000 --> 00:02:35.037 where you would end up with such enormous gains that one side would have 00:02:35.120 --> 00:02:39.397 to actually pre- would have to actually attack the other one. 00:02:39.480 --> 00:02:43.717 Uh and you know, there's a preemptive attack. It looks like that fear from me 00:02:43.800 --> 00:02:48.037 was not well grounded in fact. I think we're going to be okay for a few years. 00:02:48.120 --> 00:02:52.677 Really? What what about robotics? We seem to be pretty far behind in that front. 00:02:52.760 --> 00:02:54.757 Yeah, if if if it's okay for me to be 00:02:54.840 --> 00:02:58.237 completely blunt, um the Chinese are 00:02:58.320 --> 00:03:00.277 doing the same thing in robotics that 00:03:00.360 --> 00:03:02.917 they have done in electronic vehicles. 00:03:03.000 --> 00:03:05.397 Um in case you're confused while the 00:03:05.480 --> 00:03:08.277 current government in the US gets rid of 00:03:08.360 --> 00:03:11.437 solar and wind subsidies and promos, 00:03:11.520 --> 00:03:14.277 China put in 172 00:03:14.360 --> 00:03:16.437 gigawatts of solar last year. I think 00:03:16.520 --> 00:03:19.597 it's the number. It's remarkable. 00:03:19.680 --> 00:03:22.957 So, the the Chinese race around 00:03:23.040 --> 00:03:27.477 solar and EVs, electronic vehicles of one kind or another, looks like they've 00:03:27.560 --> 00:03:32.597 won. They're using all of those technologies, in particular new ways of 00:03:32.680 --> 00:03:36.677 building stepper motors and other very inexpensive, very powerful, you know, 00:03:36.760 --> 00:03:39.157 physical thing. Good example is Unitree 00:03:39.240 --> 00:03:42.957 just launched the R1 uh for $6,000 00:03:43.040 --> 00:03:44.997 uh available in December. I'm I've 00:03:45.080 --> 00:03:47.557 ordered one. We'll see how good it is. 00:03:47.640 --> 00:03:50.437 But the arrival of humanoid robots um is 00:03:50.520 --> 00:03:54.837 likely to be dominated by China. I'm not suggesting there won't be areas where 00:03:54.920 --> 00:03:58.957 the US will play. The US will still have spaces of very 00:03:59.040 --> 00:04:03.317 high-end, very sophisticated stuff. Our software is so much better than the 00:04:03.400 --> 00:04:06.917 Chinese software, but at the hardware level, I think you should assume that 00:04:07.000 --> 00:04:09.877 the world be awash in inexpensive 00:04:09.960 --> 00:04:12.317 Chinese robots. In the same sense that 00:04:12.400 --> 00:04:15.157 they'll be awash in inexpensive Chinese 00:04:15.240 --> 00:04:19.677 electric vehicles. Who knew? And so this is going to be my last question, 00:04:19.760 --> 00:04:23.317 but let me bump it to the front since you just beautifully segued into it. 00:04:23.400 --> 00:04:28.277 In the race for AGI, but then also in parallel robotics, 00:04:28.360 --> 00:04:31.677 China's going to be miles ahead in electricity production. 00:04:31.760 --> 00:04:34.397 And in the very short term, we're all going to be chip constrained, TPU 00:04:34.480 --> 00:04:38.917 constrained, but if you look 3-4 years out, they know that the fabs are running 00:04:39.000 --> 00:04:41.837 at full throttle, the chips are coming out by the millions, 00:04:41.920 --> 00:04:44.797 then you're suddenly electricity constrained. 00:04:44.880 --> 00:04:47.917 Is there a vulnerability for America there? 00:04:48.000 --> 00:04:50.517 It's a huge issue. So, again, let's use 00:04:50.600 --> 00:04:53.157 China versus the US as a metaphor. 00:04:53.240 --> 00:04:57.117 What are China's strengths? And I'm not praising China. I'm just be trying to 00:04:57.200 --> 00:05:01.477 report it. They have solved their electric power problem. 00:05:01.560 --> 00:05:05.077 They also have full control over social media, so they don't have the kind of 00:05:05.160 --> 00:05:07.597 problems that people here complain about. 00:05:07.680 --> 00:05:10.517 They have enormously talented software 00:05:10.600 --> 00:05:14.637 people, and they don't have enough hardware. 00:05:14.720 --> 00:05:16.357 I think that's roughly where they are. 00:05:16.440 --> 00:05:19.197 They're also incredible capitalists. 00:05:19.280 --> 00:05:24.117 They they call it Chinese Chinese socialism with 00:05:24.200 --> 00:05:28.837 socialism with Chinese characteristics, but trust me, it's pure 00:05:28.920 --> 00:05:32.157 raw capitalism. Let's call it what it is. 00:05:32.240 --> 00:05:36.357 In the US, we have the many benefits everybody understands. We do not have 00:05:36.440 --> 00:05:40.717 enough electricity, and our hard and our our at least in our consumer stuff, the 00:05:40.800 --> 00:05:42.917 Chinese are likely to beat us. 00:05:43.000 --> 00:05:44.717 Our hardware architectures are 00:05:44.800 --> 00:05:48.197 fantastic, and I'm including the Amazon 00:05:48.280 --> 00:05:50.437 chip, obviously the Nvidia chips, TPU 00:05:50.520 --> 00:05:55.437 V5, which I'm happy to say I was part of TPU V version 1. 00:05:55.520 --> 00:05:59.037 All of this stuff is incredible. In the following part, Eric talks about the 00:05:59.120 --> 00:06:03.237 lack of energy production in the United States. He says that the US is going to 00:06:03.320 --> 00:06:07.237 fall behind China if we don't generate enough electricity to pursue our AI 00:06:07.320 --> 00:06:09.517 goals. Guys, in my opinion, the way 00:06:09.600 --> 00:06:14.317 China is moving right now, I feel like we've already lost the AI race. I mean, 00:06:14.400 --> 00:06:18.677 the amount of electricity we produce is nowhere close to China. China has been 00:06:18.760 --> 00:06:22.437 upgrading its grids massively just to meet the energy demands of its data 00:06:22.520 --> 00:06:24.757 centers. So, I mean, the real question 00:06:24.840 --> 00:06:27.557 is, can the US build power plants fast 00:06:27.640 --> 00:06:32.277 enough to compete in the AI race? The answer, in my opinion, is no. So, if you 00:06:32.360 --> 00:06:34.397 think about it, and I testified in 00:06:34.480 --> 00:06:39.357 Congress a month or two ago on this, we looked at the amount of 00:06:39.440 --> 00:06:40.997 electricity required in the United 00:06:41.080 --> 00:06:45.757 States to power the expected demand 00:06:45.840 --> 00:06:48.517 of data centers, and we needed, in our 00:06:48.600 --> 00:06:51.837 calculation, by 2030, 00:06:51.920 --> 00:06:55.197 92 gigawatts to be built. And for 00:06:55.280 --> 00:07:00.357 reference, a big nuclear power plant is somewhere between 1 and 1 and 1/2 00:07:00.440 --> 00:07:03.357 gigawatts. To give you a sense of how many nuclear 00:07:03.440 --> 00:07:07.157 power plants are getting started in America, effectively zero. 00:07:07.240 --> 00:07:10.997 So, we had hoped, and I've hoped in my lobbying 00:07:11.080 --> 00:07:15.757 and test testifying that the government would fast track 00:07:15.840 --> 00:07:17.557 availability of all kinds of 00:07:17.640 --> 00:07:20.597 electricity. And indeed, they have 00:07:20.680 --> 00:07:25.077 promoted oil and gas, but they've also hobbled 00:07:25.160 --> 00:07:27.237 solar and wind to a terrible degree, 00:07:27.320 --> 00:07:32.237 which is an error. The country needs more energy. 00:07:32.320 --> 00:07:35.677 And if we don't get more energy, we're not going to be able to fully exploit 00:07:35.760 --> 00:07:39.517 the lead we have in AI and AGI. It's very clear. 00:07:39.600 --> 00:07:42.837 And by the way, the obvious next question is, what do we do? 00:07:42.920 --> 00:07:46.877 Well, there are scenarios. For example, the president went to Saudi and the UAE 00:07:46.960 --> 00:07:49.597 and did huge deals for multiple gigawatts. 00:07:49.680 --> 00:07:53.637 And so we might find ourselves in a situation where our training for our 00:07:53.720 --> 00:07:57.877 most important thing, the thing which are the the essence of America, which is 00:07:57.960 --> 00:08:02.677 American intelligence, is actually being developed in kingdoms. And that may be 00:08:02.760 --> 00:08:06.117 the only fallback we have. That is really weird, and that begs a very 00:08:06.200 --> 00:08:10.117 difficult question. Do you mind if I ask you a a tough one? 00:08:10.200 --> 00:08:15.237 Go ahead. Uh I've been, you know, uh kind of uh inspired by you for most of 00:08:15.320 --> 00:08:20.477 my adult life. And uh I I really have have been 00:08:20.560 --> 00:08:23.677 watching your uh podcast where you're talking about, look, 00:08:23.760 --> 00:08:26.597 when you you remember you when you guest lectured Erik Brynjolfsson's class over 00:08:26.680 --> 00:08:28.957 at Stanford, actually you said when you were running Google you felt like you 00:08:29.040 --> 00:08:32.477 made decisions three times faster than any company on the planet. But then when 00:08:32.560 --> 00:08:36.117 you got into the federal government, you felt like the decision making was 1/3 as 00:08:36.200 --> 00:08:39.877 fast as even a slow company. So from your point of view, it's like, you know, 00:08:39.960 --> 00:08:42.357 a tiny fraction of the pace that we need to move. 00:08:42.440 --> 00:08:44.877 >> Okay, next, Erik talks about some pretty 00:08:44.960 --> 00:08:48.797 serious stuff. He says we're going to see a lot of AI disasters if the 00:08:48.880 --> 00:08:52.797 governments around the world don't regulate AI companies. He warned that 00:08:52.880 --> 00:08:55.437 the pace at which AI is advancing, it 00:08:55.520 --> 00:08:58.437 has the potential to topple democracies. 00:08:58.520 --> 00:09:02.557 Guys, in my opinion, I believe that every major government around the world 00:09:02.640 --> 00:09:05.237 knows that AI is both a risk and a 00:09:05.320 --> 00:09:09.717 competitive advantage. If one country slows down, the other won't. That's why 00:09:09.800 --> 00:09:13.877 major players like the US and China are not putting any guardrails around their 00:09:13.960 --> 00:09:16.557 AI companies. They want that AI edge. 00:09:16.640 --> 00:09:19.477 >> But you've been saying recently that, you know, one of the things that's 00:09:19.560 --> 00:09:21.677 inevitable is some kind of an AI 00:09:21.760 --> 00:09:24.477 disaster, and we're hoping that it's 00:09:24.560 --> 00:09:26.677 like 100 people that die, 00:09:26.760 --> 00:09:31.637 and not 1,000 or 10,000 or a million or or even 100 million. 00:09:31.720 --> 00:09:35.317 Uh but, you know, suppose that it does play out that there's a catastrophe 00:09:35.400 --> 00:09:39.757 that's a wake-up call, what do you want to do with that wake-up 00:09:39.840 --> 00:09:44.397 call? What what's the next move for Eric Schmidt after the wake-up call? 00:09:44.480 --> 00:09:48.797 Well, so so the background here is that uh Dr. Kissinger, Henry, and I spent an 00:09:48.880 --> 00:09:52.197 awful lot of time talking about the period in the 1950s where he was a key 00:09:52.280 --> 00:09:54.757 component of all of these things. 00:09:54.840 --> 00:10:00.037 And the So so what he did was he he used 00:10:00.120 --> 00:10:02.837 the fact that we had used the nuclear bomb 00:10:02.920 --> 00:10:07.037 to negotiate over about a 15-year period 00:10:07.120 --> 00:10:08.837 a set of treaties that restricted 00:10:08.920 --> 00:10:14.557 nuclear proliferation. Those treaties when they were um 00:10:14.640 --> 00:10:17.997 uh when they were negotiated have allowed 00:10:18.080 --> 00:10:22.797 us to be alive today. So these were centrally important. Without but without 00:10:22.880 --> 00:10:26.557 controlling the spread of enriched uranium and the other secrets, we would 00:10:26.640 --> 00:10:28.517 all be toast, literally, because of 00:10:28.600 --> 00:10:31.157 crazy people and so forth. Is there an 00:10:31.240 --> 00:10:34.237 analogous set of things that we can do? 00:10:34.320 --> 00:10:38.197 The problem here is uh I guess the good news is we're not in a war, we haven't 00:10:38.280 --> 00:10:41.877 had a nuclear bomb, we don't have that thing to discuss. So we can talk about 00:10:41.960 --> 00:10:45.877 it, but governments tend to act reactively. So let me talk There are 00:10:45.960 --> 00:10:49.917 three obvious threats right now, which I think are fairly well understood. The 00:10:50.000 --> 00:10:53.557 first is misinformation and the software that we're all 00:10:53.640 --> 00:10:56.197 collectively giving people 00:10:56.280 --> 00:11:00.677 um allows for all sorts of misinformation, fake videos, fake news, 00:11:00.760 --> 00:11:02.997 what have you. We all understand this. 00:11:03.080 --> 00:11:04.917 It's all open source, that's done. 00:11:05.000 --> 00:11:08.477 That's a threat to democracies and maybe to dictatorships, certainly to 00:11:08.560 --> 00:11:11.517 democracies. The second one is cyber. 00:11:11.600 --> 00:11:15.117 And I think one way to understand cyber is that if you can write code, you can 00:11:15.200 --> 00:11:17.757 also write cyber cyber attacks. It's the 00:11:17.840 --> 00:11:19.757 same logic. And you have these 00:11:19.840 --> 00:11:22.837 incredible gains in software. It's 00:11:22.920 --> 00:11:27.077 frightening how good these soft Remember, my career is as a programmer. 00:11:27.160 --> 00:11:28.517 These things program better than I ever 00:11:28.600 --> 00:11:31.797 did. It's like shocking, right? And then 00:11:31.880 --> 00:11:34.117 the third one is bio. And I think most 00:11:34.200 --> 00:11:36.997 people believe that one of those three 00:11:37.080 --> 00:11:40.117 will create some kind of mini crisis 00:11:40.200 --> 00:11:44.637 that will then cause the governments to say, "Hang on. Let's have a conversation 00:11:44.720 --> 00:11:49.117 about how to really deal with the downsize." The upsides are incredible. 00:11:49.200 --> 00:11:52.517 Right? And I want America to win, and I want us to run as fast as we can, and 00:11:52.600 --> 00:11:56.357 we're indeed doing that with the Trump administration, which is great. 00:11:56.440 --> 00:12:00.757 Right? But we have to be aware that these possi- these things are possible. 00:12:00.840 --> 00:12:03.517 Um the one I'm particularly worried about is biological, and it goes 00:12:03.600 --> 00:12:07.477 something like this. You take some existing pathogen, and using biological 00:12:07.560 --> 00:12:11.437 techniques, which I won't discuss, you can modify it enough that it cannot be 00:12:11.520 --> 00:12:14.317 detected, but it's still quite dangerous. That's an example of a 00:12:14.400 --> 00:12:16.157 threat. There are many others. 00:12:16.240 --> 00:12:20.637 Yeah, I that's also the easiest, I think, which is scary. 00:12:20.720 --> 00:12:24.517 Um of the of the immediate, you know, CBRN threats, threats, so cyber, 00:12:24.600 --> 00:12:26.837 biological, radiological, nuclear. 00:12:26.920 --> 00:12:29.957 Uh biological is the one you can kind of do in a basement with three people, but 00:12:30.040 --> 00:12:33.197 it's also the one that's hardest to contain. So, 00:12:33.280 --> 00:12:35.037 anyway, uh I hope the theory is right. 00:12:35.120 --> 00:12:37.557 So, so then uh how do you deal with 00:12:37.640 --> 00:12:42.637 proliferation? You know, the Biden approach was, "Okay, let's contain AGI 00:12:42.720 --> 00:12:46.437 to these five companies, you know, Google being one of them, and then let's 00:12:46.520 --> 00:12:51.237 say any model with over 1E26 training flops has to register with the federal 00:12:51.320 --> 00:12:54.717 government, and then we'll keep it all contained." 00:12:54.800 --> 00:12:57.277 Uh so, that all got scrapped immediately, you know, after the 00:12:57.360 --> 00:13:01.637 election, and got replaced by the new David Sacks document, which the David 00:13:01.720 --> 00:13:06.077 Sacks document is much more about how do we move as quickly as possible and and 00:13:06.160 --> 00:13:09.357 win the race. Um but it doesn't really address 00:13:09.440 --> 00:13:11.357 proliferation, and obviously in America, 00:13:11.440 --> 00:13:13.957 we want startups uh and researchers to 00:13:14.040 --> 00:13:19.037 have incredible access to technology and compute. On the other hand, you know, 00:13:19.120 --> 00:13:23.517 the three people in a basement making a biological weapon is is a real scary 00:13:23.600 --> 00:13:25.837 thing. So, how do you balance those? 00:13:25.920 --> 00:13:29.557 Look, it turns out if you read the David Sacks President Trump announcement, 00:13:29.640 --> 00:13:31.277 they're very clear that they want to 00:13:31.360 --> 00:13:35.117 continue to study the security aspects 00:13:35.200 --> 00:13:37.237 of AI, especially in a geopolitical way. 00:13:37.320 --> 00:13:40.357 And that's code for China versus the US. 00:13:40.440 --> 00:13:44.557 And so they continue in their proposal to fund things involving nation-state 00:13:44.640 --> 00:13:46.797 attacks and so forth and so on. And I 00:13:46.880 --> 00:13:49.317 fully support that kind of stuff. 00:13:49.400 --> 00:13:54.037 Um there's probably a difference on the misinformation social media stuff 00:13:54.120 --> 00:13:57.837 between the two. But for example, the Biden rule was 00:13:57.920 --> 00:14:01.997 simply that you had to report if you're doing a a training run 10 to the 26 or 00:14:02.080 --> 00:14:05.677 greater. I was part of the group that made that number up. Uh we made it up 00:14:05.760 --> 00:14:07.557 cuz we had no better number. Uh I'm not 00:14:07.640 --> 00:14:09.597 suggesting it's the right number. 00:14:09.680 --> 00:14:14.077 And the the conclusion you come to is 00:14:14.160 --> 00:14:18.837 it's probably the case that that we know 00:14:18.920 --> 00:14:22.557 our government, I don't know this, but I'm guessing, knows where the training 00:14:22.640 --> 00:14:25.157 runs are going on in in China and 00:14:25.240 --> 00:14:27.037 because of espionage. And it's probably 00:14:27.120 --> 00:14:30.477 the case that the Chinese have espionage 00:14:30.560 --> 00:14:32.277 on the US knowing where our training 00:14:32.360 --> 00:14:34.757 runs are. So I'm not sure the nature of 00:14:34.840 --> 00:14:37.757 the training run runs is a secret. And 00:14:37.840 --> 00:14:39.877 frankly, everybody knows where the data 00:14:39.960 --> 00:14:42.597 centers are cuz they're immense. So 00:14:42.680 --> 00:14:47.477 uh we will see. Another attack on the 10 26 is that the training is getting more 00:14:47.560 --> 00:14:50.637 um more cost-efficient. If you look at the 00:14:50.720 --> 00:14:54.917 moving they move from something called FP16 to FP8, 00:14:55.000 --> 00:14:58.037 that means eight precision floating point. People are now moving to four 00:14:58.120 --> 00:14:59.357 four bit floating point, which is 00:14:59.440 --> 00:15:02.117 bizarre. It turns out these trainings 00:15:02.200 --> 00:15:07.517 algorithms seem to be quite tolerant for floating point in in precision, which is 00:15:07.600 --> 00:15:10.917 a shock to me. So again, we're getting more efficient 00:15:11.000 --> 00:15:15.277 in training. This creates more of a proliferation problem. Okay, next Eric 00:15:15.360 --> 00:15:20.117 talks about the threat from open-source AI. Guys, I don't agree with Eric here. 00:15:20.200 --> 00:15:24.037 I believe these tech CEOs are going after open-source models because they 00:15:24.120 --> 00:15:25.917 just want to protect their investments. 00:15:26.000 --> 00:15:29.957 Their goal here is to criminalize open-source AI. I don't think 00:15:30.040 --> 00:15:33.797 open-source is the real threat. The real threat is that one of the big 00:15:33.880 --> 00:15:38.037 corporations will eventually train a super-intelligent AI with the wrong 00:15:38.120 --> 00:15:42.237 terminal goal. If I can If I can say in general the proliferation issue, 00:15:42.320 --> 00:15:45.837 I'm not worried about big companies and big countries 00:15:45.920 --> 00:15:49.797 because I can count them. You know, there'll be 10 huge data 00:15:49.880 --> 00:15:52.677 centers, 10 huge training runs around the world. 00:15:52.760 --> 00:15:55.877 I'm much more worried about the open-source groups 00:15:55.960 --> 00:15:59.077 which can operate in the shadows. They don't have to solve every problem. They 00:15:59.160 --> 00:16:02.997 just have to do one thing well. They can patch together the open-source which is 00:16:03.080 --> 00:16:05.237 generally available and is good enough. 00:16:05.320 --> 00:16:07.357 If you look at the quality of uh deep 00:16:07.440 --> 00:16:11.637 seek R R2 R1 and now R2 coming, sure 00:16:11.720 --> 00:16:16.677 looks like it's in 80% or 90% of these top closed models. The closed models 00:16:16.760 --> 00:16:20.717 people, which obviously includes Google, get very defensive over this because 00:16:20.800 --> 00:16:23.917 they say, "Look, those are kind of synthetic. They have diffused the 00:16:24.000 --> 00:16:27.677 models. They've trained our best information." They're not All of that is 00:16:27.760 --> 00:16:32.077 true. They're correct. But they are nevertheless useful for specific things 00:16:32.160 --> 00:16:34.197 and they could be used in a 00:16:34.280 --> 00:16:39.237 proliferation issue to do various forms of cyber and biological attacks. Well, 00:16:39.320 --> 00:16:42.437 yeah, the current rule of thumb is that distillation and transfer learning is 00:16:42.520 --> 00:16:46.717 about 1% of the cost of the original training but gets you to the same 00:16:46.800 --> 00:16:50.277 destination. So, if you take, you know, FP4 which you mentioned, you know, you 00:16:50.360 --> 00:16:55.357 get an 8x performance boost over FP32. So, you get an 8x on the on the, 00:16:55.440 --> 00:16:58.197 you know, the the weights and then you've got another 00:16:58.280 --> 00:17:02.357 100x on the transfer learning and it's really hard to draw an analogy to 00:17:02.440 --> 00:17:06.957 nuclear or chemical weapons in the past cuz you couldn't kind of shrink them a 00:17:07.040 --> 00:17:09.237 thousand x under the covers and get the 00:17:09.320 --> 00:17:12.077 same result. But AI is really weird that 00:17:12.160 --> 00:17:14.317 way. It's It's very compressible, very 00:17:14.400 --> 00:17:16.317 fluid. 00:17:16.400 --> 00:17:19.197 Uh well, again, we we we will 00:17:19.280 --> 00:17:21.717 It does not look like we have a very 00:17:21.800 --> 00:17:24.517 good solution for distillation. 00:17:24.600 --> 00:17:27.917 It looks like opponent of a company 00:17:28.000 --> 00:17:32.837 can mask their queries to look like normal sets of users and then distill 00:17:32.920 --> 00:17:36.397 the models. One of my friends has thought about this 00:17:36.480 --> 00:17:40.437 a lot and he thinks that the eventual state in the United 00:17:40.520 --> 00:17:43.317 States is that the biggest models will 00:17:43.400 --> 00:17:45.317 never be released. 00:17:45.400 --> 00:17:47.557 And that the companies will distill 00:17:47.640 --> 00:17:50.717 their own models down for that reason. 00:17:50.800 --> 00:17:53.757 That's his opinion. We'll see if that's true. 00:17:53.840 --> 00:17:58.477 And this produces a bizarre outcome where the biggest models in the United 00:17:58.560 --> 00:18:03.237 States are closed source and the biggest models in China are open 00:18:03.320 --> 00:18:05.997 source. And the geopolitical issue there, of 00:18:06.080 --> 00:18:08.597 course, is that open source is free and 00:18:08.680 --> 00:18:10.877 the closed source models are not free. 00:18:10.960 --> 00:18:14.597 And so the vast majority of governments and countries who don't have the kind of 00:18:14.680 --> 00:18:18.397 money that the West does and so forth will end up standardizing on Chinese 00:18:18.480 --> 00:18:22.197 models, not because they're better, but because they're free. And so we'll see 00:18:22.280 --> 00:18:25.317 if that's true or not, but I do worry about that. All right, guys. That's it 00:18:25.400 --> 00:18:30.320 for today's video. I'll see you guys again next week with another one.