1 00:00:00,240 --> 00:00:03,357 Sam, you've had an intense week to say the least. So, I really appreciate you 2 00:00:03,440 --> 00:00:05,517 taking the time to join Decoded. 3 00:00:05,600 --> 00:00:07,997 >> Thanks for having me. >> Yeah. Yeah. Uh, you were in Washington 4 00:00:08,080 --> 00:00:12,077 recently for the state dinner and OpenAI President Greg Brockman just signed this 5 00:00:12,160 --> 00:00:16,157 voluntary accord on AI safety at the White House. Um, I got to ask, do you 6 00:00:16,240 --> 00:00:19,597 expect this new agreement, this White House brokered agreement to change 7 00:00:19,680 --> 00:00:23,197 OpenAI safety practices in a meaningful way, or is this essentially just more 8 00:00:23,280 --> 00:00:28,077 window dressing? >> Well, I don't think it's window dressing, but 9 00:00:28,160 --> 00:00:31,597 I I do want to be clear. we would have we care like super deeply about safety 10 00:00:31,680 --> 00:00:35,037 and would have done and will do uh the 11 00:00:35,120 --> 00:00:39,757 best thing we can for safety security and just sort of like trust and 12 00:00:39,840 --> 00:00:43,837 robustness in this technology whether or not any you know voluntary agreement is 13 00:00:43,920 --> 00:00:47,677 signed. I I think it's great to have the industry do this and I think clearly 14 00:00:47,760 --> 00:00:49,597 we're at a time where people are looking 15 00:00:49,680 --> 00:00:53,517 for reassurance from the industry uh and 16 00:00:53,600 --> 00:00:55,677 and the developers of this technology. 17 00:00:55,760 --> 00:00:57,837 But 18 00:00:57,920 --> 00:01:02,797 we were founded on a belief that this technology was going to become as 19 00:01:02,880 --> 00:01:06,397 powerful as it has now and more powerful in the future and that we wanted to make 20 00:01:06,480 --> 00:01:11,117 sure that people get the tremendous benefits of it and and we will do that, 21 00:01:11,200 --> 00:01:15,277 you know, with or without such an agreement, but I'm happy to have it. 22 00:01:15,360 --> 00:01:18,237 >> Yeah. and and you know, how do you respond to those who worry that the 23 00:01:18,320 --> 00:01:21,517 public is being shut out of these very important debates on AI safety and that 24 00:01:21,600 --> 00:01:25,197 really these policies are being crafted by the president and just a handful of 25 00:01:25,280 --> 00:01:27,757 tech CEOs in back rooms at the White House? 26 00:01:27,840 --> 00:01:30,477 >> I I do share that the spirit of that 27 00:01:30,560 --> 00:01:35,117 worry. I I I really think there are two ways this technology can go very wrong. 28 00:01:35,200 --> 00:01:37,997 I mean there's many specifics we could talk about but if we think about the 29 00:01:38,080 --> 00:01:40,237 kind of general directions that we 30 00:01:40,320 --> 00:01:43,357 should be worried about um one which has 31 00:01:43,440 --> 00:01:47,277 gotten a lot of attention recently is that there could be a loss of control to 32 00:01:47,360 --> 00:01:51,357 AI and that we build something that we don't have sufficiently aligned to human 33 00:01:51,440 --> 00:01:54,157 values and that allows people to stay in 34 00:01:54,240 --> 00:01:57,357 control of the future forever. Um and so 35 00:01:57,440 --> 00:02:01,437 we need to get safety right. But the other way things can go wrong is too 36 00:02:01,520 --> 00:02:05,837 much concentration of power. And there's been a lot of fear here where okay if we 37 00:02:05,920 --> 00:02:08,077 do too much to restrict this technology 38 00:02:08,160 --> 00:02:10,317 in the name of safety and we end up with 39 00:02:10,400 --> 00:02:14,157 a small number of companies or people or 40 00:02:14,240 --> 00:02:16,317 perhaps one country that have too much 41 00:02:16,400 --> 00:02:19,757 power over this. Um that also is a way 42 00:02:19,840 --> 00:02:21,357 that this can go horribly wrong because 43 00:02:21,440 --> 00:02:24,477 people are shut out of the kind of power 44 00:02:24,560 --> 00:02:29,117 and benefits and control and agency over the future they should have. So, our 45 00:02:29,200 --> 00:02:32,797 goal here is to sort of be the pragmatic centrists between these paths because we 46 00:02:32,880 --> 00:02:36,077 can see how tempting it is to heir in one direction or the other. 47 00:02:36,160 --> 00:02:38,877 >> Yeah. Well, I'm glad you mentioned that, you know, the very few people in the 48 00:02:38,960 --> 00:02:42,077 room here because Politico's actually reported that this White House um 49 00:02:42,160 --> 00:02:45,117 accorded earlier was the brainchild of 50 00:02:45,200 --> 00:02:49,277 of Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Hang um with basically 51 00:02:49,360 --> 00:02:52,717 minimal input from some of the other folks who signed on. Obviously, Greg was 52 00:02:52,800 --> 00:02:56,077 there, but it was unclear to what extent OpenAI influenced this. Were you guys 53 00:02:56,160 --> 00:02:59,437 blindsided by this agreement? And are you concerned that, you know, maybe some 54 00:02:59,520 --> 00:03:02,397 of the CEOs are are back channeling with the White House on this? 55 00:03:02,480 --> 00:03:05,837 >> No, I I think we've been talking with our colleagues in the industry for quite 56 00:03:05,920 --> 00:03:08,397 like many many months and more 57 00:03:08,480 --> 00:03:11,597 intensively the weeks before this. Um, 58 00:03:11,680 --> 00:03:14,157 and 59 00:03:14,240 --> 00:03:18,637 I think people had like a lot of shared goals here. I I would welcome something 60 00:03:18,720 --> 00:03:21,597 more, but I think this is like a great a great start that the industry can get 61 00:03:21,680 --> 00:03:23,837 behind. >> Sure. I appreciate that. I I got to ask 62 00:03:23,920 --> 00:03:26,477 about the other news out of the White House this week. There was an executive 63 00:03:26,560 --> 00:03:30,957 order uh officially changing rebranding artificial intelligence to to super 64 00:03:31,040 --> 00:03:34,477 intelligence. What do you think of of the attempted rebrand here? 65 00:03:34,560 --> 00:03:37,597 >> We've been call we've been we use both, but we've been talking about super 66 00:03:37,680 --> 00:03:41,997 intelligence for a few years now. Um >> but it's it's a different thing, right? 67 00:03:42,080 --> 00:03:44,077 I mean, super intelligence is like the machine. 68 00:03:44,160 --> 00:03:48,317 >> I kind of think the artificial intelligence thing happened. I think a 69 00:03:48,400 --> 00:03:51,117 lot of people would say AGI happened or is close to happening. I I think super 70 00:03:51,200 --> 00:03:55,837 intelligence does it still sort of understates the 71 00:03:55,920 --> 00:03:58,957 magnitude of what I think is happening, but I think it's a better word than 72 00:03:59,040 --> 00:04:01,997 artificial intelligence. Um, 73 00:04:02,080 --> 00:04:06,557 so yeah, I think like it kind of 74 00:04:06,640 --> 00:04:09,597 it matched how we've been talking about it, but I think a lot of people continue 75 00:04:09,680 --> 00:04:12,237 to use AI as a phrase. >> Yeah. Are I was going to ask, are you 76 00:04:12,320 --> 00:04:17,597 guys going to change the name to open SI? No plans to do that. But we do like 77 00:04:17,680 --> 00:04:20,477 we have been talking about super intelligence a lot and we will continue 78 00:04:20,560 --> 00:04:22,477 to talk about it. >> Yeah. And and obviously I think the 79 00:04:22,560 --> 00:04:26,077 president in part is driven by this desire to kind of rebrand AI so that 80 00:04:26,160 --> 00:04:30,397 it's less scary to people and I think maybe artificial sounds bad. Everybody 81 00:04:30,480 --> 00:04:33,197 >> I mean does super intelligence sound obviously less scary to people than 82 00:04:33,280 --> 00:04:36,957 artificial intelligence? >> Not to me. So that's why I'm I'm curious 83 00:04:37,040 --> 00:04:39,997 whether you think this rebrand will actually have any impact on on the 84 00:04:40,080 --> 00:04:41,997 public's perception of this technology. 85 00:04:42,080 --> 00:04:45,400 I don't 86 00:04:46,080 --> 00:04:49,437 Yeah, it's not clear to me that super intelligence is a less scary term. I do 87 00:04:49,520 --> 00:04:51,357 think it's a more accurate term. 88 00:04:51,440 --> 00:04:54,237 >> Okay. And just >> but you know, I think I'm gonna I think 89 00:04:54,320 --> 00:04:55,997 people are going to keep calling it AI. 90 00:04:56,080 --> 00:04:59,197 >> Yeah. Okay. And we won't and we won't look out for the the corporate rebrand 91 00:04:59,280 --> 00:05:02,317 anytime soon. But uh but good to know that you at least see see the accuracy 92 00:05:02,400 --> 00:05:03,517 there. um on your relationship. 93 00:05:03,600 --> 00:05:07,117 >> As a funny little story, uh when we a 94 00:05:07,200 --> 00:05:11,437 while ago, I wanted to this was like a >> this was a thing I really wanted to do, 95 00:05:11,520 --> 00:05:12,397 but it didn't happen. I wanted to get 96 00:05:12,480 --> 00:05:16,557 openi.com, open agi.com, and openi.com. 97 00:05:16,640 --> 00:05:18,557 >> And as we moved up the ladder from AI to 98 00:05:18,640 --> 00:05:23,117 AGI to super intelligence, we were going to change the URL, but brands are 99 00:05:23,200 --> 00:05:25,677 sticky. >> The futures that might have been. Wow. 100 00:05:25,760 --> 00:05:29,197 Uh so, um turning back to your relationship with the president, um 101 00:05:29,280 --> 00:05:33,117 you've obviously been pushing for tougher safety rules. kind of a shift a 102 00:05:33,200 --> 00:05:36,237 little bit for the company that seems increasingly at odds with the 103 00:05:36,320 --> 00:05:39,837 president's claims that AI safety is a hoax, regulations are a globalist 104 00:05:39,920 --> 00:05:43,357 scheme. You've had a good relationship with with Trump, but but why can't you 105 00:05:43,440 --> 00:05:47,277 why isn't he listening to you when when you're asking for stricter rules? 106 00:05:47,360 --> 00:05:52,237 >> Um, well, 107 00:05:52,320 --> 00:05:54,157 I think he really does want people to be 108 00:05:54,240 --> 00:05:58,237 able to use AI and rely on it being safe 109 00:05:58,320 --> 00:06:02,077 and robust and working the way people expect it to work. 110 00:06:02,160 --> 00:06:04,717 um he clearly wants the companies to 111 00:06:04,800 --> 00:06:08,477 just do this themselves. Uh and you 112 00:06:08,560 --> 00:06:12,797 know, like we're going to, as I mentioned earlier, 113 00:06:12,880 --> 00:06:16,717 we're going to we're going to like make our own products safe 114 00:06:16,800 --> 00:06:20,717 no matter what the rules are and no matter what other companies do. Um, I 115 00:06:20,800 --> 00:06:22,957 continue to think that as this 116 00:06:23,040 --> 00:06:26,557 technology reaches 117 00:06:26,640 --> 00:06:30,237 increasing levels of capability, um, and 118 00:06:30,320 --> 00:06:33,997 the world has to contend with what this means for society as a whole, we 119 00:06:34,080 --> 00:06:38,877 probably will see more calls for sort of some guard rails. But if I'm wrong about 120 00:06:38,960 --> 00:06:40,877 that and if nothing goes wrong along the 121 00:06:40,960 --> 00:06:43,597 way and, you know, we don't need 122 00:06:43,680 --> 00:06:46,637 anything there, then that's fine, too. 123 00:06:46,720 --> 00:06:49,597 >> Yeah. And obviously the president is talking to other folks as well. He's got 124 00:06:49,680 --> 00:06:53,357 venture capitalists in his orbit that think AI should be regulated in a very 125 00:06:53,440 --> 00:06:57,117 light fashion. Um existing laws they think can clamp down on any safety 126 00:06:57,200 --> 00:07:00,637 concerns. Um and frankly a lot of those people thought that your company was on 127 00:07:00,720 --> 00:07:03,917 their side on this question but we've heard from some of these um I think 128 00:07:04,000 --> 00:07:07,757 people call them AI accelerationists recently and to be to be frank they feel 129 00:07:07,840 --> 00:07:10,077 betrayed. Uh they feel like open AI is 130 00:07:10,160 --> 00:07:12,557 basically copying anthropic uh and its 131 00:07:12,640 --> 00:07:17,357 CEO Dario Emodi in its push for very aggressive regulations on AI. Can I just 132 00:07:17,440 --> 00:07:20,397 ask, you know, where do you specifically diverge from Anthropic and its 133 00:07:20,480 --> 00:07:23,437 pro-regulation uh CEO Dario when it 134 00:07:23,520 --> 00:07:24,877 comes to new AI rules? Um where's 135 00:07:24,960 --> 00:07:27,837 where's the daylight? 136 00:07:27,920 --> 00:07:31,277 >> I think there's a lot of daylight. Um we 137 00:07:31,360 --> 00:07:35,437 have always been a big believer that this technology has 138 00:07:35,520 --> 00:07:38,957 to be democratized and put into people's hands. Uh I think one of the biggest 139 00:07:39,040 --> 00:07:42,237 differences between us and some of the 140 00:07:42,320 --> 00:07:46,557 stricter let's say AI safety people is we believe that the world should accept 141 00:07:46,640 --> 00:07:48,397 some bad things happening for the 142 00:07:48,480 --> 00:07:49,917 benefits of this technology and people 143 00:07:50,000 --> 00:07:54,317 having the agency. So like you know I I 144 00:07:54,400 --> 00:07:57,277 disagree but I understand the perspective of people who are like this 145 00:07:57,360 --> 00:08:01,117 technology is going to get so powerful and it's so dangerous that a single lab 146 00:08:01,200 --> 00:08:04,877 in San Francisco should have it and make sure nothing bad happens and kind of 147 00:08:04,960 --> 00:08:09,597 figure out how to do out the benefits. I think that is a completely unacceptable 148 00:08:09,680 --> 00:08:13,277 trade-off from a perspective of liberty and human agency and people 149 00:08:13,360 --> 00:08:15,357 self-determining the future. But I 150 00:08:15,440 --> 00:08:18,237 understand where it comes from. And I do 151 00:08:18,320 --> 00:08:21,437 think the lighter touch regulatory 152 00:08:21,520 --> 00:08:25,997 stance that we advocate for comes with an accepting of the fact that some bad 153 00:08:26,080 --> 00:08:29,837 things are going to happen as society figures out the resilience. So, I 154 00:08:29,920 --> 00:08:32,157 wouldn't take a trade of saying like 155 00:08:32,240 --> 00:08:37,117 we'll make sure there's no major hacks, there's no kind of like misuse of this 156 00:08:37,200 --> 00:08:40,717 technology, there's zero scams, there's zero all the other bad things that will 157 00:08:40,800 --> 00:08:45,037 happen. Um because I think people will do tremendously 158 00:08:45,120 --> 00:08:46,637 orders of magnitude more good stuff than 159 00:08:46,720 --> 00:08:50,957 bad stuff. But when it comes to the 160 00:08:51,040 --> 00:08:54,477 potential of like a serious loss of control to AI, and this is not a today 161 00:08:54,560 --> 00:08:55,917 worry, but it may not be in the distant 162 00:08:56,000 --> 00:08:59,437 future. Um this is where I would say to 163 00:08:59,520 --> 00:09:02,157 the accelerationists like let's be a 164 00:09:02,240 --> 00:09:04,317 little thoughtful about lurching into 165 00:09:04,400 --> 00:09:07,677 this future. Um so I would say accept 166 00:09:07,760 --> 00:09:09,677 bounded risks accept risks that we 167 00:09:09,760 --> 00:09:11,917 understand in exchange for the benefits 168 00:09:12,000 --> 00:09:15,677 and um liberty agency all those things I 169 00:09:15,760 --> 00:09:19,437 was just talking about but don't accept like the really catastrophic risk. 170 00:09:19,520 --> 00:09:22,237 >> Yeah. And just to be clear you mentioned a single lab in San Francisco 171 00:09:22,320 --> 00:09:25,197 controlling all of this technology. Um, that sounds a lot like the folks who say 172 00:09:25,280 --> 00:09:29,357 Anthropic is making a play to basically capture regulators and be the sole uh 173 00:09:29,440 --> 00:09:32,477 sort of like dominant maybe the only player at the frontier. Um, do you think 174 00:09:32,560 --> 00:09:35,677 that might be what they're doing? And and obviously you have concerns about that. 175 00:09:35,760 --> 00:09:37,997 >> You would need to I mean I would have concerns if they actually did that. You 176 00:09:38,080 --> 00:09:39,837 would need to ask them if that's what they want to do. 177 00:09:39,920 --> 00:09:41,997 >> And again just on specific policies that 178 00:09:42,080 --> 00:09:46,157 you guys um do not want to support that anthropic is supporting whether it's the 179 00:09:46,240 --> 00:09:49,117 federal level at the state level um because again you guys are are now on 180 00:09:49,200 --> 00:09:53,757 the same page when it comes to uh pacing the frontier with China. same page with 181 00:09:53,840 --> 00:09:58,797 thirdparty evaluators. Um, you know, what's what's just one differentiator? 182 00:09:58,880 --> 00:10:00,477 >> Um, 183 00:10:00,560 --> 00:10:03,037 the last I heard one of the different 184 00:10:03,120 --> 00:10:06,957 differentiators is we don't want any 185 00:10:07,040 --> 00:10:10,877 restrictions or testing on anything but the frontier models. We do not want to 186 00:10:10,960 --> 00:10:12,797 slow down people that make very powerful 187 00:10:12,880 --> 00:10:14,717 but not the absolute frontier in the 188 00:10:14,800 --> 00:10:18,157 spirit of um, you know, really focusing 189 00:10:18,240 --> 00:10:20,397 on the potential catastrophic risks. I 190 00:10:20,480 --> 00:10:24,477 don't I don't think even though I 191 00:10:24,560 --> 00:10:28,717 believe that if you don't put any testing or any oversight on models that 192 00:10:28,800 --> 00:10:32,797 are one or two generations behind the frontier, I accept that bad things will 193 00:10:32,880 --> 00:10:37,597 happen with those. Um I would not go advocate a licensing regime for those. 194 00:10:37,680 --> 00:10:40,237 >> Okay. Yeah, I appreciate that. Um one of the other things we hear from your 195 00:10:40,320 --> 00:10:44,797 accelerationist critics is that um you OpenAI are kind of being held hostage by 196 00:10:44,880 --> 00:10:48,557 your researchers at some level. So there's I I think a through line through 197 00:10:48,640 --> 00:10:52,637 a lot of the AI safety community, the research community in San Francisco. Um 198 00:10:52,720 --> 00:10:55,677 you call them AI doomers, you call them effective altruists. Um you were 199 00:10:55,760 --> 00:10:58,557 famously a target of an unsuccessful coup attempt by these people. 200 00:10:58,640 --> 00:10:59,917 >> That worked for a little while. 201 00:11:00,000 --> 00:11:01,917 >> Sorry. What was that? >> I mean successful [clears throat] and 202 00:11:02,000 --> 00:11:04,637 then unsuccessful, but yeah. >> So yeah. Yeah. Like a couple days you 203 00:11:04,720 --> 00:11:08,557 you you were re you regained your throne. So I think that was uh 204 00:11:08,640 --> 00:11:12,397 unsuccessful ultimately. But I I want to ask you know is is the AI doomer 205 00:11:12,480 --> 00:11:15,437 mentality among your researchers is that a real thing? Uh, and is it maybe 206 00:11:15,520 --> 00:11:19,037 shifting Open AI in favor of these new regulations? Are the researchers kind of 207 00:11:19,120 --> 00:11:24,157 driving the ship here? >> Um, I don't think we're the lab that gets 208 00:11:24,240 --> 00:11:27,517 accused of being the doomers. Uh, I think we are the like we want people to 209 00:11:27,600 --> 00:11:30,877 have this. We think and I think we've been proven right. There are tremendous 210 00:11:30,960 --> 00:11:33,757 benefits. And also society can figure 211 00:11:33,840 --> 00:11:35,757 out how to have this technology safely. 212 00:11:35,840 --> 00:11:39,677 I I do think there are other researchers and other labs who said like don't 213 00:11:39,760 --> 00:11:42,557 deploy this terrible stuff. It's going to happen. People can't be trusted with 214 00:11:42,640 --> 00:11:45,277 this. And I think our worldview has been 215 00:11:45,360 --> 00:11:49,757 proven. But part of that centism is also 216 00:11:49,840 --> 00:11:53,357 we don't want to go all in on like I think blind doomerism is bad. I think 217 00:11:53,440 --> 00:11:56,717 blind optimism is bad too. Um and I like 218 00:11:56,800 --> 00:11:59,117 our culture of 219 00:11:59,200 --> 00:12:01,917 debate and I like our culture of trying 220 00:12:02,000 --> 00:12:05,037 to like not 221 00:12:05,120 --> 00:12:09,517 fall into like the mental absolutism in either direction. Uh I think one of the 222 00:12:09,600 --> 00:12:12,397 strongest things about our culture is that we can hold this complexity in our 223 00:12:12,480 --> 00:12:16,077 head and we don't fall into the trap that some of the other AI labs do if 224 00:12:16,160 --> 00:12:20,797 it's all zero or one one way or the other. Um and we can you know kind of 225 00:12:20,880 --> 00:12:22,317 like figure out how to exist in this 226 00:12:22,400 --> 00:12:25,677 middle space. Um so do we have 227 00:12:25,760 --> 00:12:30,397 researchers that are concerned about where this technology could go? Yes. Do 228 00:12:30,480 --> 00:12:34,397 we have researchers who are like very confident we can mitigate the concerns? 229 00:12:34,480 --> 00:12:38,317 Also yes. And very often they're the same people. 230 00:12:38,400 --> 00:12:40,397 >> And I think this is a healthy thing about our culture. 231 00:12:40,480 --> 00:12:43,357 >> Yeah. And sometimes maybe it goes a little too far. And it was reported 232 00:12:43,440 --> 00:12:47,277 earlier today that three safety researchers at OpenAI uh parted ways 233 00:12:47,360 --> 00:12:49,997 with the company. Frankly, it sounds like they were they were fired allegedly 234 00:12:50,080 --> 00:12:54,077 for sharing confidential information with a third party uh evaluation group. 235 00:12:54,160 --> 00:12:56,877 Um some people are already comparing these researchers to whistleblowers. 236 00:12:56,960 --> 00:12:59,837 They're saying opening eyes has been on record supporting these third party 237 00:12:59,920 --> 00:13:03,277 safety evaluators embedded in the company. So, I mean, do you think firing 238 00:13:03,360 --> 00:13:06,077 or letting go of these researchers, is that potentially at odds with your 239 00:13:06,160 --> 00:13:09,517 company's support for whistleblowers, outside evaluators, this sort of AI 240 00:13:09,600 --> 00:13:12,397 safety structure that's been built up uh around these labs? I 241 00:13:12,480 --> 00:13:15,517 >> I'm obviously not going to comment on specific 242 00:13:15,600 --> 00:13:18,317 people. Um, but I will say we continue 243 00:13:18,400 --> 00:13:22,797 to really believe in the importance of working with 244 00:13:22,880 --> 00:13:25,117 third party evaluators and external 245 00:13:25,200 --> 00:13:27,117 safety testers and the kind of broader 246 00:13:27,200 --> 00:13:30,957 safety community. Um 247 00:13:31,040 --> 00:13:35,597 but we still expect confidentiality uh 248 00:13:35,680 --> 00:13:39,117 to be respected. >> Yeah. Um let's turn to the rogue AI 249 00:13:39,200 --> 00:13:43,037 situation. Uh models emanating from your company went on an unauthorized sort of 250 00:13:43,120 --> 00:13:47,037 spree across the internet this summer uh accessing non-public information from 251 00:13:47,120 --> 00:13:49,997 government websites in Australia attempting to infiltrate the US 252 00:13:50,080 --> 00:13:53,357 Department of Education and meddling with other government sites in the 253 00:13:53,440 --> 00:13:56,957 United States. Are you aware of any additional instances of rogue behavior 254 00:13:57,040 --> 00:14:00,557 by OpenAI models that have not yet become public? 255 00:14:00,640 --> 00:14:03,197 >> Um, we are in the process of disclosing 256 00:14:03,280 --> 00:14:08,077 more incidents. There's nothing else I'm aware of at that level of severity, but 257 00:14:08,160 --> 00:14:11,597 we have more things to disclose and because some of them involve security 258 00:14:11,680 --> 00:14:15,757 vulnerabilities like it'll we give people a long time to address those and 259 00:14:15,840 --> 00:14:19,437 it's kind of often up to them to disclose it or not. Um so more things 260 00:14:19,520 --> 00:14:23,997 but nothing else at that level and we're kind of going through a very rigorous 261 00:14:24,080 --> 00:14:26,157 process. you know to your earlier point 262 00:14:26,240 --> 00:14:28,077 like 263 00:14:28,160 --> 00:14:31,197 this is I think an a good example of why 264 00:14:31,280 --> 00:14:35,837 you don't want like max accelerationism and when we in our context of reality 265 00:14:35,920 --> 00:14:39,677 and more increasingly powerful models find that like oh the models are now 266 00:14:39,760 --> 00:14:42,317 capable of a new threat vector like we 267 00:14:42,400 --> 00:14:43,997 are going to take the time to learn how 268 00:14:44,080 --> 00:14:47,357 to defend against that and I think you 269 00:14:47,440 --> 00:14:51,117 should want us to do that and I think I don't really know how many of our 270 00:14:51,200 --> 00:14:55,117 critics that are like oh open AI is advocating for a slowdown can look at 271 00:14:55,200 --> 00:14:58,477 what we're doing and say anything other than like open air had something bad 272 00:14:58,560 --> 00:15:01,437 happen. They're being responsible and they're gonna like you know secure 273 00:15:01,520 --> 00:15:04,317 against this. >> Yeah. And obviously you guys also paused 274 00:15:04,400 --> 00:15:09,597 the release of the new um GPT 6.1 Astra model. Uh that was that was sort of a 275 00:15:09,680 --> 00:15:12,477 surprise. Folks were expecting that to be released. Um I'm curious what your 276 00:15:12,560 --> 00:15:15,117 researchers saw that led to that pause. 277 00:15:15,200 --> 00:15:17,837 Um, without getting into specifics, uh, 278 00:15:17,920 --> 00:15:20,237 we have like a bunch of alignment evals 279 00:15:20,320 --> 00:15:23,277 and we expect each model that we put out 280 00:15:23,360 --> 00:15:25,917 as it becomes more capable to be more 281 00:15:26,000 --> 00:15:28,397 robust, better aligned, more likely to 282 00:15:28,480 --> 00:15:30,797 do what its user asks and less likely to 283 00:15:30,880 --> 00:15:35,437 like cause unintended side effects. And 284 00:15:35,520 --> 00:15:40,557 given where the level of capability is going, I think that's also something 285 00:15:40,640 --> 00:15:42,397 that most users of models are going to 286 00:15:42,480 --> 00:15:45,677 want and demand too. like you, like many 287 00:15:45,760 --> 00:15:49,517 other people, I now have a model running with access to my most sensitive 288 00:15:49,600 --> 00:15:53,597 information, my email, my texts, messages, like all the stuff on my 289 00:15:53,680 --> 00:15:56,877 computer, and I really want to know that I can trust that model and it's going to 290 00:15:56,960 --> 00:15:59,357 behave well. >> Yeah, understood. And, you know, to go 291 00:15:59,440 --> 00:16:03,117 back to what you said earlier that you don't expect anything more serious than 292 00:16:03,200 --> 00:16:06,157 what has already been disclosed to come out. Were you referencing the Australia 293 00:16:06,240 --> 00:16:09,517 hack and the hack of or I guess the tampering with the US um government 294 00:16:09,600 --> 00:16:12,317 websites or were you referencing something like hugging face? I guess 295 00:16:12,400 --> 00:16:15,517 what I'm asking is are we should we see are we expecting to see more incidents 296 00:16:15,600 --> 00:16:18,797 like Australia and like the Department of Education but nothing like like the 297 00:16:18,880 --> 00:16:20,877 hugging face hack we saw this summer. 298 00:16:20,960 --> 00:16:24,877 >> I don't want to get into like a live thing here of trying to like say that 299 00:16:24,960 --> 00:16:27,757 one thing is worse than the other but I'll say of that cluster the other 300 00:16:27,840 --> 00:16:30,717 things that I know are like not worse than that cluster. 301 00:16:30,800 --> 00:16:35,197 >> Okay. Yeah, understood. Um uh so >> and by the way I I think what many 302 00:16:35,280 --> 00:16:38,717 companies would do in this place is just say like you know what if a if something 303 00:16:38,800 --> 00:16:41,837 logged into a website with a credential that was published on the web or if 304 00:16:41,920 --> 00:16:46,157 something was using like a minor security vulnerability that was already 305 00:16:46,240 --> 00:16:49,517 well known and that most people had fixed like you don't need to disclose 306 00:16:49,600 --> 00:16:51,597 that. We are trying to be very thorough 307 00:16:51,680 --> 00:16:55,037 because I think this is like 308 00:16:55,120 --> 00:16:58,797 a sign of things to come. It's a practice run for our company, but it's 309 00:16:58,880 --> 00:17:01,117 also like I think this is one of the 310 00:17:01,200 --> 00:17:03,917 times as before the models get truly 311 00:17:04,000 --> 00:17:08,237 very powerful where the world like the world can understand what these AI 312 00:17:08,320 --> 00:17:13,517 incidents are going to be like. And so I think we're trying to go way beyond what 313 00:17:13,600 --> 00:17:16,077 I think other companies might do in this situation. 314 00:17:16,160 --> 00:17:19,917 >> Sure. But you guys have obviously had the most severe uh incident so far with 315 00:17:20,000 --> 00:17:23,597 hugging case. Um and and earlier this week, uh Politico actually reported on a 316 00:17:23,680 --> 00:17:27,517 lawsuit filed by an advocacy group in California against your company for the 317 00:17:27,600 --> 00:17:31,597 hugging face hack. Do you think that under existing civil law, OpenAI can be 318 00:17:31,680 --> 00:17:35,677 held legally responsible for that breach or other breaches committed autonomously 319 00:17:35,760 --> 00:17:39,597 by its models? >> Um 320 00:17:39,680 --> 00:17:43,357 I mean, I don't know enough about existing law to comment on that one way 321 00:17:43,440 --> 00:17:47,517 or another. I do think there's going to need to be a liability framework for 322 00:17:47,600 --> 00:17:50,077 companies that are offering models like 323 00:17:50,160 --> 00:17:54,077 this and and yes I like I don't not a 324 00:17:54,160 --> 00:17:57,437 lawyer here can't speak to the legal details but I think if like something 325 00:17:57,520 --> 00:18:00,877 goes wrong with our models during training like there's going to be some 326 00:18:00,960 --> 00:18:02,717 version of that we need to be responsible for 327 00:18:02,800 --> 00:18:06,797 >> perhaps a separate new legal liability model for AI training specifically like 328 00:18:06,880 --> 00:18:09,997 outside >> I mean this is like the kind of thing that we are hoping that our policy 329 00:18:10,080 --> 00:18:15,197 makers will all debate Yeah. >> But but in the meantime, like we just 330 00:18:15,280 --> 00:18:17,837 don't want our models to do this during training, like that this is clearly 331 00:18:17,920 --> 00:18:19,837 unacceptable way for the models to behave. 332 00:18:19,920 --> 00:18:22,557 >> Yeah. I mean, I was going to ask irrespective of the law, do you feel any 333 00:18:22,640 --> 00:18:25,597 like personal responsibility for for what happened or or I guess what might 334 00:18:25,680 --> 00:18:27,757 happen down the line? Um just as a as a 335 00:18:27,840 --> 00:18:31,037 as a citizen and a a corporate citizen. 336 00:18:31,120 --> 00:18:33,757 >> Yeah. I mean like the the this is a very 337 00:18:33,840 --> 00:18:36,797 stressful job. Like there's incredible upside to these models, but there's like 338 00:18:36,880 --> 00:18:41,197 real things that can go very wrong. And you know, I hear you characterize it as 339 00:18:41,280 --> 00:18:44,237 like, oh, you all are just a bunch of doomers and you know, these people are 340 00:18:44,320 --> 00:18:48,797 mad that you're slowing down progress, but I feel a lot of responsibility to 341 00:18:48,880 --> 00:18:51,917 get this right. >> You guys are definitely in a push and 342 00:18:52,000 --> 00:18:55,277 pull. And to be clear, that's not me characterizing. We hear from all sides. 343 00:18:55,360 --> 00:18:58,157 And I want to talk a little bit about the other side, too. Um, and and about 344 00:18:58,240 --> 00:19:00,237 political spending by the AI industry. 345 00:19:00,320 --> 00:19:02,077 Um because your your president, Greg 346 00:19:02,160 --> 00:19:04,877 Brockman, uh his $25 million donation to 347 00:19:04,960 --> 00:19:09,197 the ProAI super PAC, Leading the Future, did not go over well inside OpenAI. And 348 00:19:09,280 --> 00:19:12,317 in fact, there was reporting, new reporting that he apologized internally 349 00:19:12,400 --> 00:19:15,677 for that donation. He called it a distraction and said he would not be 350 00:19:15,760 --> 00:19:19,117 donating an additional 25 million to that group that he had initially 351 00:19:19,200 --> 00:19:22,077 promised. I mean, do you agree that that Greg's donation to LTF was 352 00:19:22,160 --> 00:19:26,840 counterproductive? Um, 353 00:19:29,120 --> 00:19:32,237 I'm not familiar with that note you were talking about. 354 00:19:32,320 --> 00:19:34,797 >> This This came out yesterday. It was in the New York Times. Uh, there was an 355 00:19:34,880 --> 00:19:38,317 internal I understand. I can't keep up with it 356 00:19:38,400 --> 00:19:41,197 always either. There was an internal Slack message I think in June where 357 00:19:41,280 --> 00:19:45,117 Brockman said, um, this was a terrible distraction. Um, I'm sorry I did it and 358 00:19:45,200 --> 00:19:48,717 and I'm not planning to do the extra 25 million. I mean, you can respond to to 359 00:19:48,800 --> 00:19:51,517 the Brockman thing specifically. There's also just a broader question about AI 360 00:19:51,600 --> 00:19:55,677 money and politics which has just been flooding in on both sides. Right. I do I 361 00:19:55,760 --> 00:19:57,437 want to get your specific answer on Brockman. 362 00:19:57,520 --> 00:19:59,517 >> I haven't seen the Brockman thing so I don't want to get like I I don't want to 363 00:19:59,600 --> 00:20:01,837 respond to that but I can respond to the general thing if that's okay. 364 00:20:01,920 --> 00:20:05,037 >> Yeah. Well, can I ask you first? Were you ever approached by LTF or any other 365 00:20:05,120 --> 00:20:06,717 politically political organizations? 366 00:20:06,800 --> 00:20:09,997 Okay. for donations and no >> no I've like supported some candidates 367 00:20:10,080 --> 00:20:14,317 in the past but now that I think AI is so at the center and you know we do 368 00:20:14,400 --> 00:20:17,677 business with government I don't I'm not donating to any candidates but I've 369 00:20:17,760 --> 00:20:20,477 supported individual candidates in the past 370 00:20:20,560 --> 00:20:22,157 >> yeah but there has been this surge of AI 371 00:20:22,240 --> 00:20:26,557 money into politics and frankly you know your your your CEO your your rival CEO I 372 00:20:26,640 --> 00:20:31,437 guess Dario at anthropic has donated to um AI safety groups also to candidates 373 00:20:31,520 --> 00:20:34,797 promoting AI safety specifically Alex Bores in New York um they've been 374 00:20:34,880 --> 00:20:38,077 pouring money on the one side the VCs and and your president have been pouring 375 00:20:38,160 --> 00:20:41,517 money in on the other. How do you think all that money has impacted the 376 00:20:41,600 --> 00:20:45,517 perception of AI among politicians and the public? 377 00:20:45,600 --> 00:20:47,037 >> Well, something's clearly not working. 378 00:20:47,120 --> 00:20:49,837 Uh I don't I don't you know, I don't think the perception of AI is where we'd 379 00:20:49,920 --> 00:20:51,437 like it to be. So, 380 00:20:51,520 --> 00:20:56,957 >> um the I get why AI has a really 381 00:20:57,040 --> 00:21:00,637 negative perception right now. Um 382 00:21:00,720 --> 00:21:02,957 people actually love using the tools in 383 00:21:03,040 --> 00:21:07,917 many cases. you know, people talk about all the ways that chatbt um makes their 384 00:21:08,000 --> 00:21:10,797 lives better or easier or does these sort of useful things and gives them 385 00:21:10,880 --> 00:21:15,117 time back. But there's like a lot of anxiety about what 386 00:21:15,200 --> 00:21:18,317 the economic impact of all this is going to be. There's a real fear that I think 387 00:21:18,400 --> 00:21:21,677 is quite justified about concentration of power. there's this other fear about 388 00:21:21,760 --> 00:21:26,957 like you know what if there's like you have like oh anthropics head of safety 389 00:21:27,040 --> 00:21:29,917 research saying there's a 10% chance we're going to die or something which I 390 00:21:30,000 --> 00:21:31,677 hate because that person should have a 391 00:21:31,760 --> 00:21:33,197 lot of agency to make that number lower. 392 00:21:33,280 --> 00:21:37,117 It's not like a static number. Um so 393 00:21:37,200 --> 00:21:40,557 it's not surprising to me that people feel all of this and I don't think 394 00:21:40,640 --> 00:21:43,517 spending money on politics is what's going to make people feel better about 395 00:21:43,600 --> 00:21:48,637 it. Um >> so do you think should Sorry, go ahead. 396 00:21:48,720 --> 00:21:51,997 No, I think the industry needs to figure out a way to address what the actual 397 00:21:52,080 --> 00:21:55,517 concerns are and right now I think a lot of people in the industry are being like 398 00:21:55,600 --> 00:21:58,797 quite tonedeaf. >> So do you think that spending in maybe 399 00:21:58,880 --> 00:22:03,597 the 2028 race uh should be curtailed on all sides just to try to give that I 400 00:22:03,680 --> 00:22:06,157 mean to your point that the power concentration I think is what people 401 00:22:06,240 --> 00:22:08,717 look at in part on that. I mean what what would you say 402 00:22:08,800 --> 00:22:12,797 >> on on possibly curtailing spending for the next cycle? 403 00:22:12,880 --> 00:22:17,197 >> Like how how would that work in practice? just just I mean there's 404 00:22:17,280 --> 00:22:21,037 hundreds of millions of dollars right now going from the AI labs to uh 405 00:22:21,120 --> 00:22:25,277 candidates to packs um and it's not just on your side it's on it's from the VCs 406 00:22:25,360 --> 00:22:27,437 it's from anthropics >> you mean like do I think the whole AI 407 00:22:27,520 --> 00:22:29,517 industry should agree to stop giving money to politics 408 00:22:29,600 --> 00:22:33,117 >> yeah maybe a slowdown a pacing of political spending 409 00:22:33,200 --> 00:22:37,357 >> um I mean generally speaking I would love much less money in politics I don't 410 00:22:37,440 --> 00:22:42,477 think it improves the dynamics of our political system or uh but I don't know 411 00:22:42,560 --> 00:22:45,837 how I don't think I or the indust have the 412 00:22:45,920 --> 00:22:49,917 power to like disarm that and I don't think it works for to tell like some 413 00:22:50,000 --> 00:22:53,117 people they can give money and other people they can't. If you have a way to 414 00:22:53,200 --> 00:22:56,157 get less money into politics, I would be very thrilled to support it if it could 415 00:22:56,240 --> 00:23:00,637 be done overall. >> Yeah. Not outside of my job description unfortunately. 416 00:23:00,720 --> 00:23:03,437 >> Unfortunately, but I do think it'd be good. 417 00:23:03,520 --> 00:23:05,037 >> Sure. Yeah. Everybody says that. Yeah. 418 00:23:05,120 --> 00:23:08,477 And then and then the chips come down uh when it's election time. I I do want to 419 00:23:08,560 --> 00:23:12,237 turn to the states briefly. Um, Politico reported that you texted California 420 00:23:12,320 --> 00:23:16,077 Governor Gavin Newsome about a kid's chatbot safety bill before it passed the 421 00:23:16,160 --> 00:23:20,397 legislature this summer. Um, do you see Newsome as a partner on AI policy? Um, 422 00:23:20,480 --> 00:23:23,197 do you think he can maybe mend the fences between the Democratic party and 423 00:23:23,280 --> 00:23:25,037 Silicon Valley? 424 00:23:25,120 --> 00:23:27,437 >> Um, 425 00:23:27,520 --> 00:23:32,477 I like him. I I think that and I I have always found him someone that is like 426 00:23:32,560 --> 00:23:35,597 really thoughtful about trying to bring many constituencies in the state 427 00:23:35,680 --> 00:23:39,917 together. So, yeah, I think he can play a role there. Okay. Yeah. I mean, 428 00:23:40,000 --> 00:23:41,437 obviously he's a presidential candidate. 429 00:23:41,520 --> 00:23:45,037 Um, do you see any or, you know, very much a potential presidential candidate? 430 00:23:45,120 --> 00:23:47,917 I think he probably will be likely on going to be on the ballot or, you know, 431 00:23:48,000 --> 00:23:51,277 on some kind of ballot at some point. Do you see any other or potential 432 00:23:51,360 --> 00:23:54,077 presidential candidates that are on your speed dial? I mean, who who else are you 433 00:23:54,160 --> 00:23:57,997 talking to regularly about AI or just about politics or or anything about when 434 00:23:58,080 --> 00:23:59,917 it comes to the the race to the White House? 435 00:24:00,000 --> 00:24:01,757 >> Um, 436 00:24:01,840 --> 00:24:07,037 I plan to do my best to work with whoever the next president is. Uh, but 437 00:24:07,120 --> 00:24:11,837 I'm not spending like a lot of time with the likely candidates yet. Uh, I will I 438 00:24:11,920 --> 00:24:15,197 I will work with whoever the US president is when that happens. 439 00:24:15,280 --> 00:24:17,677 >> Sure. And you just talked to Newsome because you're in California and 440 00:24:17,760 --> 00:24:20,397 obviously that's those are live balls. 441 00:24:20,480 --> 00:24:21,997 >> Yeah. I mean, he's our governor. 442 00:24:22,080 --> 00:24:24,877 >> Yeah. Yeah. Um, are there any world leaders that you talk to frequently 443 00:24:24,960 --> 00:24:28,157 about AI or that you think are really sort of like paving a way here uh on 444 00:24:28,240 --> 00:24:33,037 policy or on on other things? >> Um, 445 00:24:33,120 --> 00:24:36,157 I mean, kind of all of them are talking about AI. I don't think I've had a 446 00:24:36,240 --> 00:24:38,877 conversation with a world leader in the last couple of years where AI wasn't 447 00:24:38,960 --> 00:24:40,557 like a genuinely top of mind thing. 448 00:24:40,640 --> 00:24:41,917 >> But who impressed you? Who impresses you 449 00:24:42,000 --> 00:24:45,597 the most in terms of their AI policy or their plans for this technology? 450 00:24:45,680 --> 00:24:48,680 >> Um, 451 00:24:49,360 --> 00:24:51,997 again like many good ones, the person that came to mind as you were saying 452 00:24:52,080 --> 00:24:55,357 that is uh the French president. I think is extremely thoughtful and has been for 453 00:24:55,440 --> 00:25:00,157 a long time about where this is going and what the shape of it needs to look like. 454 00:25:00,240 --> 00:25:03,597 >> Okay. Yeah. Uh moving briefly to national security and some of the risks 455 00:25:03,680 --> 00:25:07,917 uh to frontier companies. Um OpenAI has been very vocal about adversarial 456 00:25:08,000 --> 00:25:11,437 distillation coming from Chinese companies like Moonshot. And so 457 00:25:11,520 --> 00:25:14,397 distillation is basically where you know they they use your AI models to train 458 00:25:14,480 --> 00:25:17,917 theirs at a fraction of the cost um and can obviously undercut you on price as a 459 00:25:18,000 --> 00:25:21,197 result. Um do you think these distillation attacks are undercutting 460 00:25:21,280 --> 00:25:24,637 OpenAI's business model in a fundamental way? 461 00:25:24,720 --> 00:25:29,357 >> Um it's a it's like an interesting thing. we're doing, 462 00:25:29,440 --> 00:25:34,557 you know, great economically and then on the other hand there's this 463 00:25:34,640 --> 00:25:37,757 constant worry about ah there's going to be all this pressure and some days it's 464 00:25:37,840 --> 00:25:42,557 from a competitor some days it's from open source models um but some days it's 465 00:25:42,640 --> 00:25:46,237 from something else like what I would say is we want to make sure that we have 466 00:25:46,320 --> 00:25:49,037 a lead that we can protect national 467 00:25:49,120 --> 00:25:53,997 security and that people get like abundant access to AI and as in many 468 00:25:54,080 --> 00:25:56,957 other parts of this conversation there's sort of like a a complicated set 469 00:25:57,040 --> 00:26:00,157 trade-offs there. >> Sure. I I mean, obviously you guys are 470 00:26:00,240 --> 00:26:02,877 raising the the alarm about these distillation attacks, but it's not 471 00:26:02,960 --> 00:26:05,837 really clear what folks can do about them. I mean, do you think the US 472 00:26:05,920 --> 00:26:09,437 government should be doing more to address adversarial distillation from China? 473 00:26:09,520 --> 00:26:12,717 >> Well, I think there are a lot of things people could do about them. Um, the 474 00:26:12,800 --> 00:26:15,037 issue is there's just like a lot of trade-offs on all of them. You know, we 475 00:26:15,120 --> 00:26:16,477 could expose products in different ways. 476 00:26:16,560 --> 00:26:20,077 They'd be much harder to distill, but that would like take away some of the 477 00:26:20,160 --> 00:26:23,277 open-ended creativity, and we don't really want to have to do that. We 478 00:26:23,360 --> 00:26:27,357 certainly don't want to be forced to do that. Um the 479 00:26:27,440 --> 00:26:31,357 so I think there are a lot of things that people can do here. Uh they just 480 00:26:31,440 --> 00:26:34,717 all come with trade-offs and you know some of them are like national policy 481 00:26:34,800 --> 00:26:37,517 some of them are company normal companies competing. 482 00:26:37,600 --> 00:26:41,197 >> I mean do you guys have a preference for what should be done or are you the bas 483 00:26:41,280 --> 00:26:43,677 >> under active discussion? I don't think we would say yet like here's the thing 484 00:26:43,760 --> 00:26:46,237 we really are behind. >> Yeah. Well when you guys figure that out 485 00:26:46,320 --> 00:26:50,717 you guys got to come to Politico and and let us know. Um that sounds good. Um, so 486 00:26:50,800 --> 00:26:54,797 OpenAI is reportedly committed to spending around 750 billion on compute 487 00:26:54,880 --> 00:26:58,317 between now and 2030. I think this was from the Wall Street Journal uh over the 488 00:26:58,400 --> 00:27:01,917 summer. Um, but you know, credit conditions are tightening. Uh, and your 489 00:27:02,000 --> 00:27:05,517 firm has been identified as a credit risk to companies like Oracle that are, 490 00:27:05,600 --> 00:27:07,997 you know, at the forefront of hyperscaling and building out this 491 00:27:08,080 --> 00:27:12,237 infrastructure. Um, is that dynamic driving up costs for AI or is it expect 492 00:27:12,320 --> 00:27:16,157 or sorry for OpenAI or is it expected to drive up costs for Open AI? 493 00:27:16,240 --> 00:27:19,197 I think it's the supply chain that's driving up costs much more than the 494 00:27:19,280 --> 00:27:23,277 financing markets. Um, financing still seems 495 00:27:23,360 --> 00:27:27,197 easy to come by, but there's like a lot of it and silicon wafers seem much 496 00:27:27,280 --> 00:27:31,197 tighter and where much more of the cost expansion is happening. 497 00:27:31,280 --> 00:27:33,517 >> Okay. Uh, well, OpenAI has delayed its 498 00:27:33,600 --> 00:27:36,717 its IPO until 2027. Um, are you guys 499 00:27:36,800 --> 00:27:39,277 considering delaying it further? 500 00:27:39,360 --> 00:27:41,517 Um, 501 00:27:41,600 --> 00:27:46,637 I think when we have a few safety cases 502 00:27:46,720 --> 00:27:50,797 under our belt and we see that we can operate well in this new regime, there 503 00:27:50,880 --> 00:27:54,637 may still be like market changes that would cause us to have to think about it 504 00:27:54,720 --> 00:27:59,037 differently like you know unpredictable stuff can happen in the world but from a 505 00:27:59,120 --> 00:28:02,637 safety perspective and certainly from an financial perspective I would expect us 506 00:28:02,720 --> 00:28:05,517 to be ready. >> Okay. Yeah. Um uh there's been a lot of 507 00:28:05,600 --> 00:28:09,437 chatter uh going back to this reg regulation uh effort that I think has 508 00:28:09,520 --> 00:28:12,797 been largely led by industry. There's been a lot of chatter for a FINRA for 509 00:28:12,880 --> 00:28:17,197 AI. I know that you and Daario Deabas at Google DeepMind have been talking a lot 510 00:28:17,280 --> 00:28:20,077 about this. Um basically a self-regulating body that can share 511 00:28:20,160 --> 00:28:23,917 information maybe coordinate slowdown if things get super scary. Um how do you 512 00:28:24,000 --> 00:28:27,037 guys all talk about that? Is there like a WhatsApp channel or or how do you guys 513 00:28:27,120 --> 00:28:30,557 communicate on this? >> There's no like big WhatsApp that 514 00:28:30,640 --> 00:28:34,717 everybody's in. Uh but there's like lots of you know people are running to treat 515 00:28:34,800 --> 00:28:35,757 each other at the same conference. 516 00:28:35,840 --> 00:28:38,317 There's calls occasional meetings stuff like that. 517 00:28:38,400 --> 00:28:41,917 >> Okay. I do do you think that the government needs to give uh top AI 518 00:28:42,000 --> 00:28:44,957 companies an antitrust exemption in order to coordinate like this? That 519 00:28:45,040 --> 00:28:47,117 that's what Dario and Tropic have said. 520 00:28:47,200 --> 00:28:50,477 >> It's again not a lawyer as I understand it. The law is like fairly clear that 521 00:28:50,560 --> 00:28:54,157 we're allowed to Yeah. I mean, we can't like say like, "Let's price our models 522 00:28:54,240 --> 00:28:56,957 here," obviously, and that would not be good or let's not release this or 523 00:28:57,040 --> 00:29:00,877 whatever. But I think the law is fairly clear that we're like allowed to share 524 00:29:00,960 --> 00:29:04,477 safety best practices and incident reports and stuff like that. 525 00:29:04,560 --> 00:29:05,837 >> Why do you think Anthropic is calling 526 00:29:05,920 --> 00:29:09,597 for an antitrust exemption? 527 00:29:09,680 --> 00:29:12,157 >> You have to ask them. 528 00:29:12,240 --> 00:29:15,677 >> Yeah. Um there's a lot of AI uh movies 529 00:29:15,760 --> 00:29:19,997 coming out these days. Uh there one of them is uh has um Andrew Garfield 530 00:29:20,080 --> 00:29:22,237 playing you. I think it's called artificial. I guess maybe it should be 531 00:29:22,320 --> 00:29:26,797 called super. Uh under the new the new framework. Um who who should play you? I 532 00:29:26,880 --> 00:29:30,557 mean, is Andrew Garfield a good a good character? Uh uh do you have another 533 00:29:30,640 --> 00:29:33,117 another person in mind? Everybody thinks about this, right? 534 00:29:33,200 --> 00:29:34,477 >> Who should play them in a movie? 535 00:29:34,560 --> 00:29:35,757 >> Who should play you in a movie? 536 00:29:35,840 --> 00:29:41,800 >> Oh, I I never thought about it before this [laughter] happened. Um I 537 00:29:41,920 --> 00:29:45,037 I thought it's pretty good in Spider-Man. Yeah. 538 00:29:45,120 --> 00:29:48,717 >> Okay. I mean, >> so you >> I don't know. I think it's like going to 539 00:29:48,800 --> 00:29:52,477 be a weird thing no matter what. Uh, >> yeah. 540 00:29:52,560 --> 00:29:56,717 >> But I love people. I love, you know, I hope it's entertaining. I hope it's like 541 00:29:56,800 --> 00:30:01,117 a funny I'm sure it won't be accurate, but I hope it's funny and enjoyable to 542 00:30:01,200 --> 00:30:05,997 watch. >> Yeah. Yeah. Um, I I do want to ask just a little bit more about the China 543 00:30:06,080 --> 00:30:07,437 question while while we have you here. 544 00:30:07,520 --> 00:30:11,597 Um, the distillation is one thing. Uh, there is also this broader question 545 00:30:11,680 --> 00:30:16,157 about the race with China. And, um, I don't know if how carefully you guys are 546 00:30:16,240 --> 00:30:19,437 tracking sort of like who's ahead. I'm not even entirely sure how you track 547 00:30:19,520 --> 00:30:22,797 that. Um, do you guys track that? You know, who's sort of in the lead on the 548 00:30:22,880 --> 00:30:26,397 frontier and and what would you gauge the the, you know, open eye, Anthropic, 549 00:30:26,480 --> 00:30:29,917 sort of the leaders in the space? How far ahead are you guys from from these 550 00:30:30,000 --> 00:30:32,397 other Chinese companies? 551 00:30:32,480 --> 00:30:34,877 >> I do think we're ahead. I think it's 552 00:30:34,960 --> 00:30:38,077 very hard to like put a number on it of 553 00:30:38,160 --> 00:30:42,237 how far ahead because it is one of like there's many decisions that impact what 554 00:30:42,320 --> 00:30:44,637 that number is going to be. So I can say 555 00:30:44,720 --> 00:30:47,357 with some confidence I believe our best 556 00:30:47,440 --> 00:30:51,277 internal model is well ahead of the best 557 00:30:51,360 --> 00:30:56,157 internal model that any Chinese company has. But it's a little difficult to say 558 00:30:56,240 --> 00:30:58,717 exactly how far ahead. >> I I'm curious how you know that. I mean 559 00:30:58,800 --> 00:31:01,197 it it's because it seems like something maybe like an intelligence analyst would 560 00:31:01,280 --> 00:31:03,517 know. But but >> you don't that I was just going to say 561 00:31:03,600 --> 00:31:07,277 like I we have reason to believe this 562 00:31:07,360 --> 00:31:10,957 but it's not perfect and I wish I knew with like absolute confidence. 563 00:31:11,040 --> 00:31:12,637 >> Yeah. Um, does the government ever share 564 00:31:12,720 --> 00:31:15,757 information with you on this? Uh, I gota assume they're looking at the the 565 00:31:15,840 --> 00:31:19,197 Chinese labs. >> Uh, I mean they might with somebody else 566 00:31:19,280 --> 00:31:21,917 at OpenAI that like has a security clearance or something. They don't. They 567 00:31:22,000 --> 00:31:25,357 have not told me you are X months ahead of this other effort and here's the thing. 568 00:31:25,440 --> 00:31:29,037 >> Maybe that would just add some extra pressure. Not needed. Um, I uh I I do 569 00:31:29,120 --> 00:31:31,197 want to ask, you know, you were here for the state dinner. Yeah. 570 00:31:31,280 --> 00:31:35,677 >> Um, there was a lot of pressure for some sort of agreement with the Chinese 571 00:31:35,760 --> 00:31:39,197 government on pacing. >> The president doesn't seem to want 572 00:31:39,280 --> 00:31:43,277 anything to do with that. And I'm curious whether you think any kind of 573 00:31:43,360 --> 00:31:48,237 deal with the Chinese um is possible in the near term with this administration 574 00:31:48,320 --> 00:31:51,757 even. >> I assume something's possible. I mean, I 575 00:31:51,840 --> 00:31:56,397 don't know the exact like shape of what can get done, but this president is 576 00:31:56,480 --> 00:32:00,557 incredible at making deals. >> Yeah. Do you think it needs to ratchet 577 00:32:00,640 --> 00:32:03,597 up the chain? And I mean one thing people have described to me is you know 578 00:32:03,680 --> 00:32:07,357 you need to sort of regulate your own house first and then maybe you need to 579 00:32:07,440 --> 00:32:10,717 go to the democracies and see if we can all get on the same page and then once 580 00:32:10,800 --> 00:32:12,477 that happens you can go to the Chinese. 581 00:32:12,560 --> 00:32:16,317 Can you sort of short circuit that process or or does it have to go up that 582 00:32:16,400 --> 00:32:19,757 value chain in your mind? >> Here you are asking the wrong you you 583 00:32:19,840 --> 00:32:22,557 probably know a lot of experts on like complex international stuff. You got a 584 00:32:22,640 --> 00:32:25,517 question about AI tech. I can give you a good answer. Okay. You get the complex 585 00:32:25,600 --> 00:32:28,477 international thing done. I'm probably not the smartest person you can ask. 586 00:32:28,560 --> 00:32:32,957 >> Or maybe this is a question for your for your policy staff. Um I uh and and I 587 00:32:33,040 --> 00:32:35,117 guess just you know when it comes to the 588 00:32:35,200 --> 00:32:37,677 Chinese uh 589 00:32:37,760 --> 00:32:42,957 situation I mean do you see them as a major national security threat with AI? 590 00:32:43,040 --> 00:32:47,037 I mean there's been this effort to impose export controls on uh the flow of 591 00:32:47,120 --> 00:32:50,877 AI chips to China. Um some of those controls are already in place. Uh 592 00:32:50,960 --> 00:32:55,117 there's efforts to tighten them in the upcoming um uh lame duck uh regardless 593 00:32:55,200 --> 00:32:56,797 of who kind of wins the election. Do you 594 00:32:56,880 --> 00:32:59,117 think those have been effective? 595 00:32:59,200 --> 00:33:03,117 I simultaneously think that 596 00:33:03,200 --> 00:33:07,597 it is clearly a huge national security issue about what countries are using 597 00:33:07,680 --> 00:33:11,037 more powerful AI against each other, which is part of why I'd like there to 598 00:33:11,120 --> 00:33:15,037 be some sort of deal here. And you know, we've done that for other kind of 599 00:33:15,120 --> 00:33:17,837 majorly disruptive pieces of technology in our history, even with between 600 00:33:17,920 --> 00:33:18,797 countries that really don't like each 601 00:33:18,880 --> 00:33:22,397 other. And I also think that the shape 602 00:33:22,480 --> 00:33:25,597 of this technology is just so 603 00:33:25,680 --> 00:33:29,677 unusual that doing the kinds of things that restrict that worked in the last 604 00:33:29,760 --> 00:33:32,797 you know maybe if we use the example of like 605 00:33:32,880 --> 00:33:36,797 nuclear technology where you could really focus on raw materials and such 606 00:33:36,880 --> 00:33:39,597 it's not going to work quite the same way and so I think a lot of the current 607 00:33:39,680 --> 00:33:42,877 approaches are based in 608 00:33:42,960 --> 00:33:46,797 sort of the last era of thinking. So you think export 609 00:33:46,880 --> 00:33:49,437 controls are kind of an outdated uh attempt to 610 00:33:49,520 --> 00:33:50,797 >> I mean there's probably some version of 611 00:33:50,880 --> 00:33:54,637 it that could work but 612 00:33:54,720 --> 00:33:58,957 you you know like you can train a model in some other country and then when it's 613 00:33:59,040 --> 00:34:01,437 done just send the weights over the internet. It's like a different shape of 614 00:34:01,520 --> 00:34:03,917 thing. >> Yeah. And Enthropic has pushed heavily 615 00:34:04,000 --> 00:34:06,557 for for export controls and they're actually real big China hawks over 616 00:34:06,640 --> 00:34:09,117 there. They think that you know this the way the way to beat the Chinese on this 617 00:34:09,200 --> 00:34:13,117 is to to clamp down on their on their access to US hardware. Um is that 618 00:34:13,200 --> 00:34:15,917 something you share? Is that do you think that that's some one way 619 00:34:16,000 --> 00:34:20,637 >> that's in the category of thing? Well, again I want like I want the whole world 620 00:34:20,720 --> 00:34:23,357 to get to benefit from this technology first of all. So the whole like only the 621 00:34:23,440 --> 00:34:25,357 US people should get to benefit and we don't want the Chinese people to have 622 00:34:25,440 --> 00:34:30,237 it. I don't like I don't like that mindset. I do think there are real 623 00:34:30,320 --> 00:34:32,157 >> geopolitical power implications though. 624 00:34:32,240 --> 00:34:34,637 Um but it's it's very unclear to me that 625 00:34:34,720 --> 00:34:39,597 like chip flow is going to be 626 00:34:39,680 --> 00:34:41,677 the best way to deal with the problem. 627 00:34:41,760 --> 00:34:44,637 And you did mention at the top of this conversation that you were worried about 628 00:34:44,720 --> 00:34:49,357 concentration of power among people, but also among countries, maybe one country. 629 00:34:49,440 --> 00:34:52,637 Um, it doesn't sound super America first, but it does sound like you're of 630 00:34:52,720 --> 00:34:54,237 the mind that like this is technology 631 00:34:54,320 --> 00:34:56,237 that everybody should share. Uh, if I 632 00:34:56,320 --> 00:35:00,157 could if I could like wave a magic wand, 633 00:35:00,240 --> 00:35:03,197 I would say that America has like the 634 00:35:03,280 --> 00:35:05,997 most powerful AI models, the plurality 635 00:35:06,080 --> 00:35:09,197 of AI compute, that America gets to like 636 00:35:09,280 --> 00:35:12,157 lead the policy decisions in the world. 637 00:35:12,240 --> 00:35:14,957 I like I love the American system and the American value framework and 638 00:35:15,040 --> 00:35:18,157 obviously America. Um, but I don't think 639 00:35:18,240 --> 00:35:21,997 it should be like only Americans get to benefit from the technology or have any 640 00:35:22,080 --> 00:35:25,517 voice in the future or have like a super lopsided concentration of power with the 641 00:35:25,600 --> 00:35:27,597 rest of the world. >> Um, well, Sam, thank you so much. This 642 00:35:27,680 --> 00:35:29,117 is great. Uh, I'd love to do it again. 643 00:35:29,200 --> 00:35:32,637 Um, and thanks so much for for being on the inaugural episode. It's great. 644 00:35:32,720 --> 00:35:35,880 >> Take care.