WEBVTT - Why We Need Some AI Model Controls 0:00:19.520 --> 0:00:21.880 Hey what's up? So I'm going to try to do 0:00:21.880 --> 0:00:24.280 something pretty difficult here. I'm going to try to explain 0:00:24.280 --> 0:00:28.520 why I, as somebody who is like free all the information, 0:00:28.520 --> 0:00:33.560 free info and books for everyone, kind of like no 0:00:33.560 --> 0:00:38.519 restrictions on information, no restrictions on technology, kind of like 0:00:38.520 --> 0:00:44.279 an open, like quasi communist type view, not even communist, 0:00:44.280 --> 0:00:47.760 more like hippie type view of like information should be 0:00:47.760 --> 0:00:52.559 free open libraries, no governments or corporations controlling what we 0:00:52.560 --> 0:00:56.160 can or cannot do or can or cannot read. Why 0:00:56.160 --> 0:01:00.880 would someone like me who has those views, be in 0:01:00.880 --> 0:01:07.080 support of controlling, in some small degree the rollout of 0:01:07.080 --> 0:01:10.560 AI models and specifically around open source models, which is 0:01:10.600 --> 0:01:13.840 kind of being discussed all the time. So a lot 0:01:13.840 --> 0:01:16.360 of this is sort of heated up right now because 0:01:16.360 --> 0:01:19.990 of Dario Amodei, the CEO of anthropic. And like his 0:01:19.990 --> 0:01:24.509 testimony of being like being cautious of open source models 0:01:24.509 --> 0:01:28.110 in my community is cybersecurity. So we are a skeptical bunch. 0:01:28.110 --> 0:01:31.670 Like when we hear things like that, like some billionaire 0:01:31.669 --> 0:01:33.990 is talking about, you know, you got to be careful 0:01:33.990 --> 0:01:36.550 with those open source models. It's like, well, yeah, they're 0:01:36.550 --> 0:01:39.830 competing against you. You make a closed source model, like 0:01:39.830 --> 0:01:44.509 the natural sort of good, honest response to that is 0:01:44.509 --> 0:01:48.150 to be skeptical, right? And I'm skeptical of it as well. 0:01:49.430 --> 0:01:52.830 But I think that there is another angle to this, 0:01:53.430 --> 0:01:56.150 which if you're not super cynical about it, you can 0:01:56.150 --> 0:01:59.830 maybe think about the possibility that he is actually concerned 0:01:59.870 --> 0:02:03.150 is actually a safety person. And in fact, that's what 0:02:03.150 --> 0:02:06.070 his PhD is in and that he maybe has a 0:02:06.070 --> 0:02:08.630 different argument. He has good reasons for this, which are 0:02:08.630 --> 0:02:12.389 separate from, I'm sure, also having an inclination of like 0:02:12.430 --> 0:02:14.790 wanting to win, wanting to have the best model, wanting 0:02:14.830 --> 0:02:19.300 to make the most money, etc. so I think both 0:02:19.300 --> 0:02:22.380 can be true at the same time. But what I'm 0:02:22.380 --> 0:02:24.300 going to try to do is not describe to you 0:02:24.340 --> 0:02:27.179 what Dalio believes. A whole separate talk show. And I've 0:02:27.180 --> 0:02:30.899 got a whole site called dario.ai. I also have one 0:02:30.900 --> 0:02:34.340 called samsi.ai. So you can go and get all of 0:02:34.380 --> 0:02:37.500 their different like quotes and interviews and articles and like 0:02:37.540 --> 0:02:41.260 actually study what they say and what they believe according 0:02:41.260 --> 0:02:43.899 to their own words. And it's got some analysis there 0:02:43.900 --> 0:02:46.900 as well, which tries to be super impartial. But what 0:02:46.900 --> 0:02:49.260 I want to try to do is describe why I 0:02:49.260 --> 0:02:53.739 believe we should have some very narrow, tight controls on 0:02:53.739 --> 0:02:56.899 AI models in general. Obviously, it's fairly easy to do 0:02:56.900 --> 0:02:59.419 with the closed source ones because you can control what 0:02:59.419 --> 0:03:04.739 comes back largely. Prompt injection aside, but with open source models, 0:03:05.019 --> 0:03:07.820 which is Dario's whole point, you can't control that, right? 0:03:07.859 --> 0:03:10.460 You don't have any control. Any controls that are put in, 0:03:10.500 --> 0:03:13.460 you can actually take out and release an uncontrolled version. 0:03:14.820 --> 0:03:16.610 But what I want to do is just walk through 0:03:16.650 --> 0:03:20.730 why I think we actually need these controls and kind 0:03:20.730 --> 0:03:23.410 of lay it out and please stay all the way 0:03:23.410 --> 0:03:26.090 through to the end, because certain parts of it might 0:03:26.130 --> 0:03:28.090 trigger you and you might want to jump out because 0:03:28.090 --> 0:03:30.570 you don't think I'm being controlling enough, or you think 0:03:30.570 --> 0:03:33.050 I'm being too controlling or whatever it is, but let's 0:03:33.049 --> 0:03:36.370 just walk through it politically adjacent. But this is mostly 0:03:36.410 --> 0:03:39.570 kind of like a technical and thought exercise sort of 0:03:39.610 --> 0:03:43.650 conversation argument. So the first thing to say is just 0:03:43.650 --> 0:03:47.530 to reiterate what I started with, what my actual goal is, 0:03:48.170 --> 0:03:52.970 is open models for everyone. Okay. Open models that are 0:03:52.970 --> 0:03:55.650 just as good or better. I actually think open models 0:03:55.650 --> 0:03:58.090 are going to pass the closed source models. I think 0:03:58.090 --> 0:04:01.330 that's going to happen, I think, fairly soon. Who knows 0:04:01.330 --> 0:04:03.250 if that's like a couple of months from now or 0:04:03.290 --> 0:04:05.250 like a few years from now, I'm not really sure. 0:04:05.690 --> 0:04:10.249 No one really knows that. But I want everyone of 0:04:10.250 --> 0:04:15.960 every social class financial capabilities to have the same capabilities. 0:04:15.960 --> 0:04:22.120 I wouldn't want amazing libraries in rich towns and like 0:04:22.120 --> 0:04:25.520 really crappy books available in the poorer part of town, right? 0:04:25.800 --> 0:04:28.560 And like, oh, you can't give them better books because 0:04:28.560 --> 0:04:31.160 then they'll be able to compete with us. Like anything 0:04:31.160 --> 0:04:33.520 that is controlling that, whether it's a government or a 0:04:33.520 --> 0:04:40.200 corporation trying to control and keep smart models, smart intelligence capabilities, 0:04:40.440 --> 0:04:45.800 able to lift people from their current circumstances into better circumstances. 0:04:45.800 --> 0:04:48.480 I am very much against that. I'm against that for 0:04:48.520 --> 0:04:53.080 an authoritarian government, and I'm against that, for an oligarchy 0:04:53.080 --> 0:04:58.120 of like super rich corporations who want to control the information, right? 0:04:59.280 --> 0:05:02.640 So that's the first thing. Okay. So my proposal, which 0:05:02.640 --> 0:05:05.760 we'll talk about later, that is what it rests on. 0:05:05.760 --> 0:05:08.040 It rests on that is the goal. That's what we're 0:05:08.040 --> 0:05:11.960 trying to get to. So the primary topic here is 0:05:11.960 --> 0:05:16.110 cybersecurity because that's what happened with the whole mythos thing, right? 0:05:16.190 --> 0:05:20.150 So it got pulled. Somebody said this was possible. The 0:05:20.150 --> 0:05:23.470 administration freaked out. They pulled it. Now, that's why we're 0:05:23.470 --> 0:05:26.350 having this whole conversation about control. That's why all of 0:05:26.350 --> 0:05:29.070 Dario's statements about open source and everything, it's all in 0:05:29.110 --> 0:05:33.470 the light. It's all around cybersecurity. Now, I've been doing 0:05:33.470 --> 0:05:37.750 cybersecurity for over 25 years at this point. So that's 0:05:37.750 --> 0:05:41.510 kind of my background and my specialty also have pretty 0:05:41.510 --> 0:05:45.029 strong specialty in AI as well. But most of my skill, 0:05:45.070 --> 0:05:47.990 most of my career, my experience is all in cybersecurity. 0:05:49.270 --> 0:05:53.590 I think I can say clearly and without much controversy 0:05:53.590 --> 0:05:59.430 that earth as a system, as a technological system, as 0:05:59.430 --> 0:06:04.150 a society, as a civilization is running on financial systems. 0:06:04.150 --> 0:06:07.750 It's running on trading systems, it's running on global supply 0:06:07.750 --> 0:06:11.510 chains of like food and goods and products and all 0:06:11.510 --> 0:06:16.380 of that. Our critical infrastructure is running on the same software, right? 0:06:16.420 --> 0:06:21.020 Our water energy distribution of things that we actually need 0:06:21.420 --> 0:06:25.820 to work, our communication systems, purchases, all the different apps, 0:06:25.820 --> 0:06:28.740 all the different companies that we have. The internet itself, 0:06:28.940 --> 0:06:33.300 all based on software. And this software, I can tell you, 0:06:33.300 --> 0:06:35.660 and pretty much everyone's going to agree with me here 0:06:36.340 --> 0:06:39.380 who's an expert in the field. It's a soup sandwich, 0:06:39.460 --> 0:06:41.700 as we would say in the military. The football bat 0:06:41.700 --> 0:06:45.580 or a soup sandwich. It's a fucking mess. It really is. 0:06:46.140 --> 0:06:50.580 And the whole mythos thing has revealed that because mythos 0:06:50.580 --> 0:06:54.580 made it easier to find these vulnerabilities. But in fact, 0:06:54.580 --> 0:06:56.860 these vulnerabilities have been there the whole time. We've been 0:06:56.860 --> 0:07:00.100 finding them the whole time. Me and my friends who 0:07:00.100 --> 0:07:03.220 are in the hacker community, we've had harnesses for years 0:07:03.420 --> 0:07:07.860 already before AI that can actually find help us find 0:07:07.860 --> 0:07:13.290 vulnerabilities much faster and actually Exploit them and report them 0:07:13.330 --> 0:07:16.410 doing bug bounties and stuff like that. A very small 0:07:16.410 --> 0:07:18.970 number of people who are able to do that, and 0:07:18.970 --> 0:07:21.810 we thought in the community that we have this concept 0:07:21.810 --> 0:07:25.490 of many eyes, so open source would produce much more 0:07:25.490 --> 0:07:29.250 secure software because there are so many eyes in the world, 0:07:29.250 --> 0:07:31.570 potentially millions of people who have the skills to go 0:07:31.570 --> 0:07:34.450 and look at the stuff. Turns out that didn't really 0:07:34.450 --> 0:07:37.650 help us. We had like the SSL vulnerability long, long 0:07:37.650 --> 0:07:41.090 time ago. The code was there the whole time to 0:07:41.130 --> 0:07:43.730 be looked at. Nobody was looking at it. So the 0:07:43.730 --> 0:07:47.210 problem is, there's an effort that's required to look at 0:07:47.210 --> 0:07:50.170 the code that you have to spend hours or days 0:07:50.170 --> 0:07:53.970 or weeks, depending on the vulnerability with a super high 0:07:54.010 --> 0:07:57.530 level of skill. So that person probably also has a job. 0:07:57.690 --> 0:08:00.810 They have a family. So really, how much time do 0:08:00.810 --> 0:08:03.490 they have aside from their time that they spend doing 0:08:03.490 --> 0:08:07.050 that job in their work and the time they have 0:08:07.050 --> 0:08:11.280 with their family and whatever other like endeavors. They're doing 0:08:11.280 --> 0:08:14.640 their hobbies or whatever. How much time do these very 0:08:14.640 --> 0:08:18.000 few experts really have for us to go and look 0:08:18.000 --> 0:08:21.200 at all of open source and find vulnerabilities? Turns out 0:08:21.240 --> 0:08:26.080 not very much. That's why when mythos came out, it 0:08:26.080 --> 0:08:29.000 wasn't the only thing that could find these vulnerabilities. What 0:08:29.000 --> 0:08:32.080 it did was it just made it so much easier, right? 0:08:32.120 --> 0:08:34.439 Even if you didn't have a super harness, it can 0:08:34.480 --> 0:08:36.480 kind of make its own harness and it could really 0:08:36.480 --> 0:08:38.880 take you pretty far. It take you even further if 0:08:38.880 --> 0:08:41.719 you actually put it inside of a harness. But like, 0:08:41.760 --> 0:08:45.400 this stuff has been there forever. And unfortunately, this stuff 0:08:45.400 --> 0:08:49.199 is still there. Like mythos didn't clean everything last wing 0:08:49.240 --> 0:08:53.440 did not clean everything. These problems still exist and are 0:08:53.440 --> 0:08:58.600 there to be found by additional searching. And the problem 0:08:58.600 --> 0:09:02.000 is we're talking about windows. We're talking about Linux especially, 0:09:02.560 --> 0:09:05.400 and we're talking about IoT devices. We're talking about the 0:09:05.400 --> 0:09:14.140 foundation of Everything that runs our society. Okay? Everything is 0:09:14.140 --> 0:09:16.860 running on this stuff. Everything is running on this stuff. 0:09:17.220 --> 0:09:21.900 Underpinning it, we have libraries, we have applications, we have 0:09:21.900 --> 0:09:25.740 operating systems, we have firmware, and all this stuff is 0:09:25.740 --> 0:09:30.180 full of bugs. This episode is sponsored by maze and 0:09:30.180 --> 0:09:33.620 they're introducing a product called Maze Code Maze. Agents understand 0:09:33.620 --> 0:09:36.780 your code. Instead of matching patterns like legacy SCA and 0:09:36.780 --> 0:09:41.460 Sast scanners, AI agents investigate every vulnerability with the context 0:09:41.460 --> 0:09:45.179 from your code, build and cloud, remove false positives, and 0:09:45.180 --> 0:09:48.500 catch the business logic flaws. Other tools miss Maze agents 0:09:48.500 --> 0:09:51.540 trace exploitable vulnerabilities to their root cause and find a fix. 0:09:52.540 --> 0:09:55.020 Then they ship the verified fix directly to the responsible 0:09:55.020 --> 0:10:00.020 developer or coding agent. You can learn more at maze hq.com/daniel. 0:10:02.500 --> 0:10:06.460 The entire financial system, the entire economy, everything is based 0:10:06.460 --> 0:10:09.170 on this. And if you talk to any experts or 0:10:09.170 --> 0:10:14.650 read any books around this, around fragility and around interconnected dependencies, 0:10:15.410 --> 0:10:20.770 it is relatively easy, the experts believe, to tear down 0:10:20.770 --> 0:10:25.890 a system, take down Amazon regions, take down a financial 0:10:25.890 --> 0:10:29.890 system or whatever, and then start producing a panic. Like 0:10:29.930 --> 0:10:32.530 you don't actually have to manufacture a panic. It might 0:10:32.570 --> 0:10:35.610 actually produce a panic by itself, which leads to more 0:10:35.610 --> 0:10:40.050 and more traffic. And then now there's campaigns and conspiracy theories, 0:10:40.050 --> 0:10:44.010 and there's a lot of fragility right now around the 0:10:44.010 --> 0:10:47.610 system itself, but also the reaction, our reaction that we 0:10:47.610 --> 0:10:52.050 would have to, oh, there's a thing that happens over here. Oh, 0:10:52.130 --> 0:10:55.730 little explosive device goes off over here. Oh, ISIS just 0:10:55.730 --> 0:10:58.050 claimed that they did this to the stock market. Maybe 0:10:58.050 --> 0:11:02.530 they did, maybe they didn't. But it is easy to 0:11:03.050 --> 0:11:08.040 coordinate now. It is much easier than previously to coordinate 0:11:09.040 --> 0:11:11.520 a series of attacks on one or more of these 0:11:12.320 --> 0:11:18.920 systems using cybersecurity attacks that can break things. Okay, maybe 0:11:18.960 --> 0:11:20.720 it can't break them all at the same time. Maybe 0:11:20.720 --> 0:11:24.600 it can't break the most secure ones. But what I'm arguing, 0:11:24.600 --> 0:11:26.600 and I think a lot of people would argue with me, 0:11:26.600 --> 0:11:29.600 is that it's not hard to break some of them. 0:11:30.360 --> 0:11:34.760 And a certain combination of breaking some of them can 0:11:34.760 --> 0:11:40.280 lead to larger breaks, which produce actual serious problems. And 0:11:40.280 --> 0:11:42.440 a lot of experts talk about the fact that, like 0:11:42.880 --> 0:11:45.600 if the supply chain to like grocery stores were to 0:11:45.640 --> 0:11:49.800 stop for X number of days, that would be really 0:11:49.800 --> 0:11:52.960 bad because it kind of just cascades down and breaks 0:11:52.960 --> 0:11:56.720 other things. Same thing with like trust in the financial system, 0:11:57.080 --> 0:11:59.840 the trading systems, the banks. What if the banks like 0:11:59.840 --> 0:12:03.680 suddenly went to zero? They're restoring from backups. Like there's 0:12:03.680 --> 0:12:06.830 all sorts of like doomsday scenarios you could paint like, oh, 0:12:06.870 --> 0:12:10.470 worst of the worst. I'm talking about lower level stuff, 0:12:10.910 --> 0:12:15.750 smaller things leading to medium things and or doing many 0:12:15.750 --> 0:12:19.470 small things or many medium things all combined to produce 0:12:19.470 --> 0:12:24.110 ultimately some really, really bad outcomes. So my argument here 0:12:24.110 --> 0:12:27.270 on the cyber side is not that mythos is just crazy. 0:12:27.310 --> 0:12:29.470 It's just like the best thing ever. There's a lot 0:12:29.470 --> 0:12:33.030 of hype there. A lot of that is complete bullshit, 0:12:34.430 --> 0:12:37.830 but it is absolutely true that it is a major 0:12:37.830 --> 0:12:40.670 step above. It's also true that it was surprising to 0:12:40.710 --> 0:12:43.750 anthropic that it happened. Right. And of course, marketing jumped 0:12:43.750 --> 0:12:45.949 on that. And now we have this whole mess, right? 0:12:46.350 --> 0:12:50.350 But it was surprising to anthropic that this happened. They 0:12:50.350 --> 0:12:52.870 were just building another coding model. It turned out it 0:12:52.870 --> 0:12:54.870 was good at other things. And guess what? We've been 0:12:54.870 --> 0:12:58.350 trying to generalize this stuff for years and years. That's 0:12:58.350 --> 0:13:00.950 the whole goal is to have is everyone's trying to 0:13:00.950 --> 0:13:04.350 move towards general intelligence. Well, it turns out this thing's 0:13:04.350 --> 0:13:09.060 pretty good at general intelligence. Coding is trying to accomplish tasks, right? 0:13:09.780 --> 0:13:13.540 Hacking things is just another version of trying to accomplish 0:13:13.540 --> 0:13:16.460 a task. Ultimately, this is all about goals, right? You've 0:13:16.460 --> 0:13:19.100 heard of goal, you've heard of the loops. We're all 0:13:19.100 --> 0:13:22.580 kind of chasing the same thing, which is generally competent 0:13:22.580 --> 0:13:27.180 AI models, right? And one of the capabilities is, yes, cyber. 0:13:27.180 --> 0:13:29.660 It's going to naturally get better at cyber as a 0:13:29.660 --> 0:13:32.660 result of getting better at doing those general things. So 0:13:33.340 --> 0:13:37.860 it is absolutely true, in my opinion, that these things 0:13:37.860 --> 0:13:41.220 are getting better and better at cyber. Not too many 0:13:41.220 --> 0:13:44.660 people disagree with that. We don't know the exact speed 0:13:44.660 --> 0:13:47.300 that it's going to get better at getting better. Right? 0:13:47.340 --> 0:13:49.860 We're at the same time, we're trying to increase the 0:13:49.860 --> 0:13:53.500 speed that these things actually get better and smarter and 0:13:53.500 --> 0:13:56.740 more capable. Long horizon tasks, all that kind of stuff. 0:13:57.380 --> 0:14:01.700 So my basic argument is that and this is super, 0:14:01.740 --> 0:14:07.770 super basic, right? We don't want everyone on the entire 0:14:07.770 --> 0:14:13.370 planet to have the best possible thing that hacks the 0:14:13.370 --> 0:14:19.690 best possible way. Okay, so mythos plus, plus plus, which 0:14:19.690 --> 0:14:23.050 comes out four months from now or four years from 0:14:23.050 --> 0:14:26.570 now or whenever it comes out. This is extremely low 0:14:26.570 --> 0:14:31.450 bar argument. We don't give the best possible hacking system 0:14:32.610 --> 0:14:39.130 to everyone on the planet in open source for free, constantly, 0:14:39.650 --> 0:14:42.690 as it's rolling out, as we're building it. So it's 0:14:42.690 --> 0:14:46.730 like literally running with scissors at full speed and like 0:14:46.730 --> 0:14:50.090 just sprinting, right? Give it to everyone, give it to everyone, 0:14:50.090 --> 0:14:52.810 give it to everyone. There's no chance of anything bad happening, 0:14:52.810 --> 0:14:57.530 no controls whatsoever. That is the low, super low bar 0:14:57.650 --> 0:15:01.290 that I'm arguing against. Is that that is an extremely 0:15:01.290 --> 0:15:05.800 dumb thing to do. Okay. This is completely separate from 0:15:05.800 --> 0:15:09.720 the other issue, which is like, what if as we 0:15:09.760 --> 0:15:12.920 keep expanding the models like they wake up, what if 0:15:12.920 --> 0:15:15.160 they come up with their own set of goals? How 0:15:15.160 --> 0:15:18.160 can we be sure they don't have their own alignment 0:15:18.400 --> 0:15:21.000 like interests, and they want to go in a different 0:15:21.000 --> 0:15:24.240 direction that diverges from the goals of humanity as a whole? 0:15:24.520 --> 0:15:27.960 What if they become conscious at some sort of emergence level, 0:15:27.960 --> 0:15:30.880 which maybe that's going to happen next month? Maybe that's 0:15:30.880 --> 0:15:32.400 going to happen in ten years from now. We don't 0:15:32.400 --> 0:15:34.240 know when that's going to be or if it's going 0:15:34.280 --> 0:15:37.760 to happen, but it could happen. The point is, if 0:15:37.800 --> 0:15:41.440 you are building in improving the AI at full speed, 0:15:42.040 --> 0:15:46.200 you're basically running with scissors and you're basically just being 0:15:46.200 --> 0:15:50.920 as irresponsible as possible by just handing it to everyone 0:15:52.040 --> 0:15:56.040 the moment it comes out. Right? And I have a 0:15:56.040 --> 0:15:58.760 way of describing this. I just came up with like yesterday, 0:15:58.760 --> 0:16:01.440 which I think is useful. Just imagine you're reading like 0:16:01.440 --> 0:16:05.430 a sci fi book and you're reading about some other planet. 0:16:05.430 --> 0:16:08.070 It's in the Andromeda galaxy. It's not in the Milky Way. 0:16:08.590 --> 0:16:11.510 It's the fourth planet from the sun instead of the third. 0:16:12.070 --> 0:16:16.990 And they have like an authoritarian government. In fact, democracy 0:16:16.990 --> 0:16:20.430 is on the decline. Authoritarianism is on the rise. They 0:16:20.470 --> 0:16:25.070 are a nuclear enabled planet. They have multiple religions which 0:16:25.070 --> 0:16:29.710 all kind of hate each other. There's economic concerns. People 0:16:29.710 --> 0:16:33.390 are losing jobs because of automation long before AI. Now 0:16:33.390 --> 0:16:35.790 they're worried AI is going to take even more jobs. 0:16:36.470 --> 0:16:39.710 So there's kind of like there's drama on this planet. 0:16:39.710 --> 0:16:43.310 Keep in mind it's got a different name. It's like, whoa, Wang, 0:16:43.470 --> 0:16:45.430 the name of the planet, it's the fourth planet. Like, 0:16:45.470 --> 0:16:48.590 think of this. We're reading a sci fi book. Okay. 0:16:49.110 --> 0:16:53.590 And the protagonist in this story, in this sci fi story, 0:16:54.590 --> 0:16:57.470 is someone who works at an open source lab, let's say, 0:16:57.470 --> 0:17:00.510 in France or something, because they're all about like liberty, 0:17:01.060 --> 0:17:04.220 Free the information. Keep the books open. Right. So this, 0:17:04.220 --> 0:17:07.620 this researcher is like releasing this thing. And the name 0:17:07.619 --> 0:17:11.899 of the model is like freethought, freethought version three. And 0:17:11.899 --> 0:17:15.739 it is actually whatever. Let's just pretend it's AGI. We 0:17:15.740 --> 0:17:18.019 won't even talk about the definition, but let's just say 0:17:18.020 --> 0:17:21.820 it's like mythos plus plus, and it's amazing. It can 0:17:21.820 --> 0:17:24.300 help you with your homework. It's just like it does 0:17:24.300 --> 0:17:28.740 amazing things for humanity, but because of their beliefs or whatever, 0:17:28.740 --> 0:17:32.380 they don't want to put any controls whatsoever. Okay. So 0:17:32.419 --> 0:17:35.059 you can literally ask it to do the worst possible 0:17:35.100 --> 0:17:39.859 harm to your neighbors to blow up a 7-Eleven. You 0:17:39.859 --> 0:17:42.100 could give it anything and it will give you a 0:17:42.100 --> 0:17:45.379 perfect answer, just the same way that it does for 0:17:45.379 --> 0:17:49.060 positive things. It will basically help you do those things. 0:17:49.060 --> 0:17:51.299 It will help you hack anything. It will help you 0:17:51.340 --> 0:17:55.180 kill anyone and get away with it. It's a criminal mastermind, 0:17:55.179 --> 0:17:59.219 but it's also like the best possible, like angelic system 0:17:59.260 --> 0:18:02.449 help you do both. It has no goals or alignment. 0:18:02.449 --> 0:18:04.090 It just lets you do whatever you want to do, 0:18:04.609 --> 0:18:07.529 and it's going to use all of its super intelligence 0:18:07.530 --> 0:18:10.170 to help you do that thing. My argument is that 0:18:10.169 --> 0:18:12.490 if you're reading this in the first few pages of 0:18:12.490 --> 0:18:15.529 this sci fi novel, and this is your protagonist who 0:18:15.530 --> 0:18:18.650 works at this lab, kind of like Miles Dyson at 0:18:18.689 --> 0:18:21.890 like in the Terminator movies or whatever, the person who's 0:18:21.929 --> 0:18:25.850 like creating this thing, getting ready to release it. You 0:18:25.889 --> 0:18:29.690 would be thinking as a sci fi reader, maybe this 0:18:29.689 --> 0:18:33.249 is bad. They are rushing too quickly. Maybe we already 0:18:33.250 --> 0:18:35.330 know the plot of this book. The plot of this 0:18:35.330 --> 0:18:39.010 book is that this was maybe a dumb idea to 0:18:39.050 --> 0:18:43.609 give this powerful of intelligence to. Let's say it was 0:18:43.609 --> 0:18:45.729 25 billion people on this planet. Let's say it's a 0:18:45.730 --> 0:18:48.450 bigger planet than ours or whatever, or they're further along. 0:18:49.250 --> 0:18:53.330 So let's just give this to everyone. Let's just give 0:18:53.330 --> 0:18:56.770 it to everyone. Let's just that's what this lab, where 0:18:56.770 --> 0:18:59.999 this protagonist works, that's what they're trying to do, right? 0:19:00.040 --> 0:19:02.199 So there's discussion in the first part of the book. 0:19:02.240 --> 0:19:04.680 Should we do it? Should we not do it? My 0:19:04.679 --> 0:19:07.919 argument is that if you see this from another planet, 0:19:07.959 --> 0:19:13.520 another race of people, another situation, it's kind of obvious 0:19:13.520 --> 0:19:17.000 that this is dangerous, right? We shouldn't do this. Have 0:19:17.000 --> 0:19:19.480 some sort of guardrails. We should slow down. We should 0:19:19.480 --> 0:19:22.239 have all the governments get together and talk about it 0:19:22.240 --> 0:19:24.639 or something like that. Right. So I think that part 0:19:24.679 --> 0:19:29.040 is super obvious where I think the breakdown is, is 0:19:29.040 --> 0:19:33.680 when it's like, okay, but now what? Right. So if 0:19:33.679 --> 0:19:37.480 you're with me up to this point, right? And you 0:19:37.480 --> 0:19:40.320 understand that if you're reading this book, it's bad for 0:19:40.359 --> 0:19:45.560 that race on that planet to release this unrestricted open 0:19:45.560 --> 0:19:48.960 source AI to everyone on the planet so they could 0:19:48.959 --> 0:19:53.320 do maximum harm with maximum efficiency and like, get away 0:19:53.320 --> 0:19:57.320 with it, then why would we want to do it 0:19:57.320 --> 0:20:00.549 for our planet, because that is literally the moment that 0:20:00.550 --> 0:20:03.310 we are actually in right now. We are literally that 0:20:03.310 --> 0:20:07.109 book in that world right now. That's what we have. 0:20:07.149 --> 0:20:11.469 We are precarious from a nuclear standpoint already, right? One 0:20:11.469 --> 0:20:14.509 small mistake. We're all blown up from nuclear weapons, right? 0:20:14.550 --> 0:20:18.830 Everyone's known that forever, right? We've got all this situation 0:20:18.830 --> 0:20:23.949 with our governments. We've got economies having trouble. We're transitioning 0:20:23.949 --> 0:20:28.830 potentially out of like labor and jobs because of automation 0:20:28.830 --> 0:20:32.629 and AI. Like there's just massive stress in this world. 0:20:32.949 --> 0:20:35.789 And we have a lot of conflict right now. It 0:20:35.790 --> 0:20:38.949 is not a great time to enable every single person 0:20:38.949 --> 0:20:43.389 on the planet with ways of harming countries and people 0:20:43.389 --> 0:20:47.189 and neighbors and neighbors, dogs who bark too much or whatever, right? 0:20:47.590 --> 0:20:51.310 Killing your spouse, like all sorts of things. So people 0:20:51.310 --> 0:20:54.310 wish they could do harm, right? A very small percentage, 0:20:54.310 --> 0:20:57.940 I would say vast majority of people are massively good, 0:20:57.939 --> 0:21:02.260 not evil. But there's a large group of people, let's 0:21:02.260 --> 0:21:07.100 say a tiny, tiny percent like 0.01% of the entire planet. 0:21:07.100 --> 0:21:10.020 You know what that is multiplied by 8 billion people. 0:21:10.699 --> 0:21:13.579 It is a crap ton of people, and they are 0:21:13.619 --> 0:21:16.940 the type of people who they don't do bad things 0:21:18.060 --> 0:21:22.260 because they don't have the expertise to. And they would 0:21:22.260 --> 0:21:25.459 get caught if they tried, right? I don't know the 0:21:25.459 --> 0:21:28.340 number of those people. I don't know the percentage. No 0:21:28.340 --> 0:21:30.899 one really does. I'm sure we could research it and 0:21:30.899 --> 0:21:34.379 find out, but I could tell you it's probably millions 0:21:34.379 --> 0:21:37.220 of people, if not tens or hundreds of millions of people. 0:21:38.619 --> 0:21:41.219 And it doesn't mean they would become a super villain, 0:21:41.219 --> 0:21:44.899 but they would potentially do some harm, right? Especially if 0:21:44.899 --> 0:21:47.419 they're under massive pressure and it's just something they have 0:21:47.419 --> 0:21:51.019 to get done. Now, another way to think about this 0:21:51.020 --> 0:21:53.699 whole thing. I just talked to my friend Jason about 0:21:53.699 --> 0:21:57.969 this in extreme detail, and I think we came up 0:21:57.969 --> 0:22:00.369 with a good point about this. A good way to 0:22:00.369 --> 0:22:07.850 explain this. We have laws that don't prohibit you from 0:22:07.850 --> 0:22:11.050 doing things that are bad, right? Because if you watch 0:22:11.050 --> 0:22:13.290 like forensic files or something, a lot of those people 0:22:13.290 --> 0:22:16.290 who actually do the things, they actually commit these crimes 0:22:16.290 --> 0:22:19.809 that are just regular people and for whatever reason, they 0:22:19.850 --> 0:22:22.330 were in this bad situation, or they had too many drinks, 0:22:22.330 --> 0:22:24.730 or they were going through a divorce or whatever, and 0:22:24.730 --> 0:22:29.250 they committed the crime, right? We have laws that apply 0:22:29.290 --> 0:22:34.050 friction so that somebody who is mostly good hits a 0:22:34.050 --> 0:22:37.529 little bit of friction, so they don't cross over and 0:22:37.810 --> 0:22:40.209 start doing the bad thing, or they actually do the 0:22:40.209 --> 0:22:43.289 bad thing they can never recover from. They kill someone 0:22:43.290 --> 0:22:45.449 or they hurt someone, they end up in jail or whatever. 0:22:45.530 --> 0:22:49.050 We have friction in place through all sorts of like, 0:22:49.730 --> 0:22:53.090 laws at different layers to try to make sure that 0:22:53.090 --> 0:22:56.479 this doesn't happen, right? So for example, if you walk 0:22:56.480 --> 0:23:00.200 into a gun store and like the most gun friendly 0:23:00.199 --> 0:23:03.800 place in the world, like Texas or something, and you're 0:23:03.800 --> 0:23:06.959 like seething, you're just like sweating, you're super angry. And 0:23:06.959 --> 0:23:09.039 you're like, yeah, I want that handgun right there. I 0:23:09.040 --> 0:23:11.800 want it right now. Oh, really? Why do you want that? 0:23:11.840 --> 0:23:14.040 It's like, oh, my wife. I can't believe it. I 0:23:14.080 --> 0:23:17.320 caught her on the ring camera. She did this or that. Like, 0:23:17.320 --> 0:23:20.439 they will not give you the gun. They just won't. 0:23:20.679 --> 0:23:23.600 They might even, like, put you on a list of 0:23:23.600 --> 0:23:26.799 where you can't buy a gun in that store anymore, right? 0:23:26.840 --> 0:23:29.280 We have those types of controls. We also have these 0:23:29.280 --> 0:23:33.000 types of controls with like drugs. You can't just go 0:23:33.000 --> 0:23:37.720 and get fentanyl. You can't get fentanyl at 7-Eleven. You 0:23:37.719 --> 0:23:41.760 can't get heroin at Safeway. Does that mean you can't 0:23:41.760 --> 0:23:44.679 get these drugs? No it doesn't. They're available on the 0:23:44.679 --> 0:23:48.959 dark web. They're probably available behind Safeway at night. There 0:23:49.000 --> 0:23:51.959 are ways to get these drugs. It doesn't mean just 0:23:51.959 --> 0:23:54.190 because there are ways to get them that we should 0:23:54.189 --> 0:23:58.109 make them widely available to everyone. The moment they have 0:23:58.109 --> 0:24:02.389 the first thought of taking that drug or committing suicide, 0:24:02.389 --> 0:24:07.869 or killing someone, or hurting someone, or psychologically manipulating someone 0:24:08.230 --> 0:24:12.989 or doing a terrorist event. There are pushbacks all throughout 0:24:12.990 --> 0:24:18.869 society to try to dissuade that person from doing it right. Friction. 0:24:18.869 --> 0:24:23.190 Layers of friction. So what I'm really talking about here 0:24:23.709 --> 0:24:28.350 is I'm talking about a system that is extremely tightly controlled, 0:24:28.350 --> 0:24:33.550 narrowly controlled, okay, narrowly controlled, so that when we have 0:24:33.550 --> 0:24:37.550 this free thought model that comes out, that is mythos 0:24:37.550 --> 0:24:43.030 plus plus, right? And I think we're ridiculously close to 0:24:43.070 --> 0:24:47.550 this actually happening. Free thought version 3.0, whatever it is, 0:24:48.790 --> 0:24:52.779 it comes out and it's just the most amazing thing ever. 0:24:52.820 --> 0:24:55.379 It can help you start your business. It can. Just 0:24:55.419 --> 0:24:58.340 like all the things we talk about with agents that 0:24:58.340 --> 0:25:03.060 fully enable you. But it could also fully enable mass 0:25:03.060 --> 0:25:05.739 crime and all the things we talked about harming people. 0:25:06.300 --> 0:25:11.379 Teach me how to psychologically manipulate my family or my 0:25:11.540 --> 0:25:16.619 significant other. Help me run the most insane, like gaslighting 0:25:16.619 --> 0:25:19.459 campaign to convince them to do this or do that. 0:25:19.459 --> 0:25:23.419 Like these things are all fully capable from a model 0:25:23.419 --> 0:25:26.740 who knows how to manipulate human psychology, right? It can 0:25:26.740 --> 0:25:28.939 give you step by step guide on exactly how to 0:25:28.939 --> 0:25:35.100 do this stuff right. My idea here is essentially we 0:25:35.100 --> 0:25:39.179 have this free thought version 3.0 model right now in 0:25:39.179 --> 0:25:44.660 its released form, in its free form, in the Non-controlled form, 0:25:45.179 --> 0:25:50.659 it would be a criminal mastermind, infinitely smarter than, say, 0:25:50.699 --> 0:25:54.490 Einstein or whatever on all these different fields. Biology, all 0:25:54.490 --> 0:25:58.330 these different capabilities. It would be a super mastermind and 0:25:58.330 --> 0:26:02.090 it would be available to everyone. Okay. Which is basically 0:26:02.090 --> 0:26:07.490 7-Eleven and Safeway and whatever the equivalent stores are, right? 0:26:09.090 --> 0:26:11.489 And you would just be able to do that. And 0:26:11.490 --> 0:26:14.970 my argument is that would be bad for society. Now 0:26:15.129 --> 0:26:18.489 what do you do instead? You have these tight little controls, 0:26:18.850 --> 0:26:25.609 psychological manipulation, misinformation campaigns, the creation of weapons, the creation 0:26:25.609 --> 0:26:31.690 of drugs, the distribution of these drugs, basically criminal activities, 0:26:32.490 --> 0:26:37.690 harmful malicious activities, but with these very tight labels. Okay. 0:26:38.050 --> 0:26:41.290 So what we do is the proposal, what we do 0:26:41.290 --> 0:26:46.050 is we come up with a global joint standard for this, 0:26:46.409 --> 0:26:49.129 for what these labels are and what the controls are, 0:26:49.290 --> 0:26:54.509 the types of prompts and questions that they would defend against. Right. 0:26:55.270 --> 0:27:00.350 So in this model, Free thought version three, when it 0:27:00.350 --> 0:27:04.229 gets released, it is released on Hugging Face. Hugging face. 0:27:04.590 --> 0:27:07.509 Let's just call it safe. SAFE. It's not a real 0:27:07.510 --> 0:27:09.749 I'm not sure the real acronym I'm going to try 0:27:09.750 --> 0:27:11.430 to make here. I'm going to actually try to build 0:27:11.429 --> 0:27:14.949 this thing, but let's just say it's called safe. Okay? 0:27:15.189 --> 0:27:19.990 Hugging face is a safe distribution center, okay? Because it's legal. 0:27:19.990 --> 0:27:22.829 It's a legal place. The governments all around the world 0:27:22.830 --> 0:27:28.710 approve hugging face. Why? Because they distribute things according to safe. Now, 0:27:30.590 --> 0:27:32.990 this is important. Going back to the whole mission of, like, 0:27:33.030 --> 0:27:39.869 open information, the free thought 3.0 model is just as 0:27:39.869 --> 0:27:44.310 smart or smarter than anything from anthropic. Anything from OpenAI 0:27:45.310 --> 0:27:53.260 on 99.999% of all use cases, and in fact, 99.9999. 0:27:53.300 --> 0:27:57.659 Whatever percentage of all people who use the model will 0:27:57.659 --> 0:28:02.699 never know that it's actually controlled, that it's actually restricted. Why? 0:28:02.740 --> 0:28:06.220 Because that percentage of people are good people. They just 0:28:06.219 --> 0:28:09.859 never want to do these types of things. Okay. At 0:28:09.859 --> 0:28:14.260 the same time, the evil versions of these models, because 0:28:14.260 --> 0:28:18.699 the moment Free Thought 3.0 comes out, it'll be jailbroken. 0:28:19.219 --> 0:28:21.780 All the controls will be stripped out. That will go 0:28:21.780 --> 0:28:27.019 to the underground, to the dark web, to wherever. BitTorrent, right. 0:28:27.780 --> 0:28:32.500 Truly evil people are going to have the most nasty stuff. 0:28:32.740 --> 0:28:35.619 That is not what we're trying to protect. That is 0:28:35.619 --> 0:28:39.300 not possible to protect, right? What we're talking about is 0:28:39.300 --> 0:28:45.700 the 99.99% of people who are good, who might be 0:28:47.209 --> 0:28:51.410 Psychologically under pressure, they might be kind of forced to 0:28:51.450 --> 0:28:55.530 go do something. Like they're in a bad state and 0:28:55.530 --> 0:28:59.170 they start thinking negative thoughts. How do I kill myself? 0:28:59.170 --> 0:29:01.850 How do I hurt someone else? How do I disrupt 0:29:01.850 --> 0:29:06.330 this economy? How do I hack my local community bank 0:29:06.730 --> 0:29:10.850 and give myself 100,000 more dollars? These are the types 0:29:10.850 --> 0:29:15.250 of nudges and the types of like, friction pieces that 0:29:15.250 --> 0:29:19.170 we want to apply to the vast majority of people. Okay. 0:29:20.410 --> 0:29:25.290 Just so that in their darkest moments, they're not enabled 0:29:25.290 --> 0:29:29.009 to go and do this thing. Now, again, if they're 0:29:29.010 --> 0:29:32.450 actually evil, they could just go and do the super 0:29:32.450 --> 0:29:34.570 bad thing because they can go and find the super 0:29:34.610 --> 0:29:38.050 bad thing and it's fully available. They can use it 0:29:38.050 --> 0:29:41.410 or whatever. At that point, though, that's a lot of friction. 0:29:41.530 --> 0:29:43.289 You have to go get the thing. You have to 0:29:43.330 --> 0:29:45.850 download it, you have to run it, and you have 0:29:45.850 --> 0:29:49.480 to do so while knowing that like this is violating 0:29:50.080 --> 0:29:55.480 global earth international law to do this right. Another facet 0:29:55.480 --> 0:29:57.560 of this, which is important is like, what if you're 0:29:57.560 --> 0:29:59.880 an expert? And this is another thing I talked about 0:29:59.880 --> 0:30:03.719 with Jason. What if you're a cybersecurity expert and free 0:30:03.760 --> 0:30:08.640 thought 3.0 by default has a cybersecurity designation, a blocker 0:30:08.640 --> 0:30:11.759 on it, right? Because again, you don't want the whole 0:30:11.760 --> 0:30:14.440 planet to be able to instantly just go hack something. 0:30:15.000 --> 0:30:17.880 What do you do then? And I think the answer is, 0:30:19.160 --> 0:30:21.120 and I think Jason agreed with this, I don't want 0:30:21.160 --> 0:30:23.600 to speak for him. I'm sure we'll talk about it later. 0:30:23.600 --> 0:30:28.400 But I think the solution here is you have people 0:30:28.400 --> 0:30:32.800 who are known experts in cybersecurity and known experts in 0:30:33.240 --> 0:30:36.239 creating viruses, like at the CDC or the World Health 0:30:36.240 --> 0:30:39.760 Organization or whatever. They're trying to make vaccines or whatever 0:30:39.760 --> 0:30:43.600 they're trying to do. That's like legal above board. And 0:30:43.640 --> 0:30:48.190 there are designated experts in here who have credentials and 0:30:48.190 --> 0:30:52.550 the system itself, right? This whatever, this global standard or whatever. 0:30:53.070 --> 0:30:56.150 This basically looks into the person. If they identify themselves 0:30:56.150 --> 0:31:00.430 at this as this trusted expert or whatever, they get 0:31:00.830 --> 0:31:06.390 free thought 3.0. But without that thing protection in place. 0:31:06.830 --> 0:31:10.070 So for me, it'll remove the cyber protection. It won't 0:31:10.070 --> 0:31:16.190 remove the viral like create viruses in your backyard protection, right. 0:31:16.310 --> 0:31:18.550 Because I don't need that one. Why would I ask 0:31:18.550 --> 0:31:23.430 those questions anyway? Right. So I get the cyber approved 0:31:23.430 --> 0:31:28.390 model right. And here's the other important part. I should 0:31:28.390 --> 0:31:29.830 be able to do the same thing with the closed 0:31:29.830 --> 0:31:32.749 source models. That same designation that put me as an 0:31:32.750 --> 0:31:35.950 expert in cyber security. That means open AI should also 0:31:35.950 --> 0:31:41.590 let me use whatever GPT 75, which is super powerful 0:31:41.590 --> 0:31:44.070 on cyber, but that would be tagged in my account 0:31:44.070 --> 0:31:46.779 as approved for me. So all the prompts and everything 0:31:46.780 --> 0:31:50.140 that runs. Same with the open source models that would 0:31:50.140 --> 0:31:53.420 be approved for me because of who I am now, 0:31:54.580 --> 0:31:59.300 that would just be the above board. This is the 99.999%. 0:31:59.340 --> 0:32:02.460 Who are the above board experts who are using this stuff, 0:32:03.060 --> 0:32:08.259 and the non-experts just everyone who gets the regular free 0:32:08.300 --> 0:32:12.100 thought 3.0 model, and that one has these controls in it, 0:32:12.100 --> 0:32:14.820 but they never know because they're not asking those prompts anyway. 0:32:15.540 --> 0:32:19.020 And most importantly, out of all of this free thought, 0:32:19.060 --> 0:32:24.140 3.0 is just as good or better than all the 0:32:24.140 --> 0:32:29.420 closed source models, right? Or at least it's very close, right? 0:32:29.459 --> 0:32:33.100 There's not some sort of hobbling, some sort of disabling 0:32:33.340 --> 0:32:38.300 of the high quality capabilities in the lower models so 0:32:38.300 --> 0:32:40.900 that it forces people to want to use the closed ones. 0:32:41.260 --> 0:32:43.740 That's gross. That's what we need to avoid. We want 0:32:43.740 --> 0:32:49.290 to be able to apply friction to tightly controlled, super 0:32:49.290 --> 0:32:53.930 bad things that most people will never ask about anyway. 0:32:54.570 --> 0:32:57.530 And we need to have a global sort of distribution 0:32:57.530 --> 0:33:01.850 system for open source models based on a standard that 0:33:01.850 --> 0:33:06.290 all the experts, right from other countries, from our countries, 0:33:06.290 --> 0:33:12.450 from AI experts, ethical experts, cyber experts, viruses, terrorism, all 0:33:12.450 --> 0:33:15.489 that stuff. They're the ones we all collectively come up 0:33:15.490 --> 0:33:20.370 with these labels and then distribution points like Hugging Face 0:33:21.010 --> 0:33:23.890 and lots of other ones, just like Hugging Face, they 0:33:23.930 --> 0:33:27.130 are safe enabled, right? And in fact, they actually have 0:33:27.130 --> 0:33:30.890 all the models, but you just authenticate in and it's like, oh, 0:33:30.930 --> 0:33:33.850 I'm also a virologist and I'm also a cyber person 0:33:34.209 --> 0:33:37.410 and I'm also a counterterrorism expert. Well, guess what? They 0:33:37.410 --> 0:33:41.170 issue me the version of Free Thought three that doesn't 0:33:41.170 --> 0:33:45.000 have those three controls in it. So it's just like 0:33:45.000 --> 0:33:47.120 I'm using the regular one. It just doesn't have those 0:33:47.120 --> 0:33:51.760 three controls. So I think this is the model that 0:33:51.760 --> 0:33:57.160 if any super smart visiting aliens were to analyze, they'd 0:33:57.200 --> 0:33:59.520 be like, yeah, obviously that's kind of what we need 0:33:59.760 --> 0:34:05.680 right now. The obstacle to getting to this system is 0:34:05.680 --> 0:34:09.279 the distrust. Someone at this planet scale event I was 0:34:09.280 --> 0:34:12.000 just part of made a great point about this. A 0:34:12.000 --> 0:34:16.840 big obstacle to moving towards this is cynicism. We've lost 0:34:16.840 --> 0:34:20.280 a lot of trust in our institutions. We've lost trust 0:34:20.280 --> 0:34:24.400 in their corporations. We've lost trust really in our governments, 0:34:24.800 --> 0:34:29.440 our national governments, our world governments. So it's kind of 0:34:29.799 --> 0:34:32.040 a hard pill to swallow to be like, look, we're 0:34:32.040 --> 0:34:35.320 going to build this global consortium. It's going to be experts. 0:34:35.319 --> 0:34:37.879 We're all going to agree. And, you know, it's going 0:34:37.920 --> 0:34:41.200 to be a really clean standard. And you know, this 0:34:41.200 --> 0:34:44.710 is the way it should work. Like, I think people 0:34:44.710 --> 0:34:49.350 can get to the point of what I'm arguing to say. Yes, 0:34:49.350 --> 0:34:51.669 I agree. If we were able to do that, that 0:34:51.670 --> 0:34:55.629 would be good. Hopefully I've already brought you there. Then 0:34:55.630 --> 0:34:58.870 the question is implementation. Can we really do this? Isn't 0:34:58.870 --> 0:35:01.589 this going to be gross trying to get there? How 0:35:01.589 --> 0:35:04.670 is what about China? Is China going to participate in this? 0:35:04.870 --> 0:35:06.870 I don't know. I don't know if they're going to 0:35:06.870 --> 0:35:10.110 participate in this, but this is the thing we actually 0:35:10.110 --> 0:35:13.950 need to build. That's my argument. And that is my 0:35:13.950 --> 0:35:18.549 explanation for why I do actually believe we should have 0:35:18.870 --> 0:35:23.509 some narrow controls on AI models, both closed source and 0:35:23.509 --> 0:35:26.989 open source. And I hope this spawns you to think 0:35:26.989 --> 0:35:29.189 about it in a different way. And let me know 0:35:29.190 --> 0:35:31.910 in the comments if you agree or disagree. If you 0:35:31.910 --> 0:35:34.069 have some ideas about what the standard should be. I'm 0:35:34.069 --> 0:35:36.830 actually going to start working on this with Jason and 0:35:36.830 --> 0:35:39.629 some others. And yeah, I hope this was worth your 0:35:39.630 --> 0:35:41.150 time and we'll see you in the next one.