Frontier model regulation through government executive orders is premature and dangerous because government power is a one-way ratchet that never gets taken back.

(00:00:00) All right, everybody. (00:00:01) Welcome back to the number one podcast in the world. (00:00:04) It's the All In Podcast, episode 274. (00:00:07) Sachs is out today, but we're very lucky to have Gavin Baker from Atreides Management joining us. (00:00:15) The spicy takes must flow. (00:00:18) Welcome back to the program. (00:00:19) Bestie Gavin. (00:00:20) Thanks for having me. (00:00:21) Always love it. (00:00:22) It's been a huge week in tech. (00:00:24) We can start with the SpaceX and OpenAI IPOs. (00:00:27) We've got Andre Karpathy joining Anthropic, Nvidia crushing it. (00:00:31) So many different places to go, but I think we'll start with Andre Karpathy joining Anthropic. (00:00:36) Karpathy is only 39 years old. (00:00:38) He's already a legend in the tech industry if you don't know him. (00:00:41) I believe he's also coming to Liquidity. (00:00:42) Yeah, Chama? (00:00:43) He's going to keynote on Monday morning. (00:00:45) Oh, fantastic. (00:00:46) No, Tuesday, Tuesday, Tuesday, day two, I think he's. (00:00:48) Day two, OK. (00:00:49) As is Gavin. (00:00:50) Gavin will be there. (00:00:52) Gavin anchoring day two as well. (00:00:54) Excellent. (00:00:55) Yeah, this is Gavin's second appearance. (00:00:57) Look at those two bookmarks, Andre Carpathy and Gavin Baker. (00:01:00) Oh yeah, you know. (00:01:01) Liquidity pulls in the stars. (00:01:04) Obviously, Andre was a founding member of OpenAI. (00:01:07) He led the self-driving team. (00:01:09) Also, hold on. (00:01:10) Gavin is going to help us judge the best ideas section as well. (00:01:13) Excellent. (00:01:13) I don't know if you know that, Gavin, but you're a judge. (00:01:15) You're a judge. (00:01:16) I'm up for anything, man. (00:01:17) I'm easy. (00:01:18) Yes. (00:01:19) Karpathy also coined the term vibe coding. (00:01:21) He recently built auto research. (00:01:23) I think we talked about that here a bit. (00:01:24) That's an open source training tool. (00:01:26) It helps AI models improve themselves by running five-minute experiments. (00:01:30) That got over 82,000 stars on GitHub. (00:01:32) He did that like as a weekend experiment. (00:01:34) And all these civilians started building their own recursive LLMs, really inspiring. (00:01:40) And the Andre Karpathy skills is a tool based on his set of principles for Claude code. (00:01:47) And somebody just released that. (00:01:49) And so that's just pretty crazy when you think about it. (00:01:52) He's going to be in charge of a new pre-training team at Anthropic, the focus obviously being recursive self-improvement. (00:01:58) In other words, (00:02:00) they're going to have Claude improve itself. (00:02:02) And they've already talked a little bit about AI, improving AI over at Anthropic Schmuck. (00:02:07) What's your take on this? (00:02:08) Is this super important in 2026? (00:02:12) Obviously, Karpathy is super well-respected. (00:02:15) He's obviously one of the true talents in the space, but hey, we're in a different inning than we were, say, 10 years ago when he was at Tesla or five years ago when he co-founded OpenAI. (00:02:26) You know what's interesting? (00:02:27) If you go back to Google, (00:02:29) The culture of Google, which they got right, was the singular technical talents there. (00:02:35) They were singled out and they were called Google Fellows. (00:02:38) I don't know if you guys remember this, like Amit Singhal, Sridhar Ramaswamy, Jeff Dean. (00:02:45) These guys are stars. (00:02:46) And what's interesting is if you track what folks, particularly Jeff Dean, I guess, now, because the other two aren't there anymore, but (00:02:53) What they did inside of Google, it's like wave upon wave, they were at the foot of those waves. (00:02:58) What's interesting about Andres, he's been at the wave upon wave of AI. (00:03:02) He was probably the first person that really commercialized the Richard Sutton Bitter Lesson essay. (00:03:10) when he was leading FSD at Tesla, which was really about the brute force computation. (00:03:17) And I remember him telling me this story, I don't know if he said this publicly or not, but where he spent a portion of his time, I want to say 1/4 of his time labeling data. (00:03:25) Could you imagine like 2016, 17, like hand labeling video data from Teslas? (00:03:31) So he did that, then he's co-founder of OpenAI. (00:03:33) He's a star and he's an exceptional human being and he's super curious. (00:03:37) And (00:03:39) But then what he's done as a kind of a free agent is also quite impressive. (00:03:43) So I think that this is a really important deal. (00:03:46) I think he's one of these really curious people that can be sent off and they'll just go and invent new things. (00:03:54) And I think this idea of recursive self-learning puts these models on a combination of overdrive and autopilot. (00:04:04) And so if you put those two things together, I think that you start to, (00:04:09) You could potentially live out this idea that there's an order of magnitude improvement on a yearly basis. (00:04:14) So like this new form of Moore's law. (00:04:17) So then the model quality just goes absolutely parabolically, just like this, straight up. (00:04:22) I think a bunch of compute at the problem and these things learn really quick, I think is the (00:04:28) I ordered a bit there. (00:04:29) Gavin, what's your take on Anthropic's recent success and their massive hiring binge? (00:04:35) The success is extraordinary. (00:04:37) It's undeniable. (00:04:39) I think the fact that they are now, they were EBIT positive per the Wall Street Journal in the most recent quarter is a really important fact for kind of the whole AI narrative. (00:04:50) Because now there's (00:04:53) you could talk about circular funding, you could talk about ROI, and we could go look at the ROIC of the hyperscalers. (00:05:00) But if OpenAI and Anthropic are at, call it $100 billion of ARR now with 80 percent-ish gross margins on inference, like the returns are there. (00:05:12) And then if we add in, and they're growing really fast, if we add in Gemini, we add in Cursor, we add in XAI, we add in open source, you know, it's (00:05:21) not hard to see 200, 300, $400 billion of ARR at the end of this year at high margins. (00:05:29) Across all of the language models. (00:05:31) And you're talking specifically about the private language model companies, maybe not Google, which is- No, I was including Google. (00:05:37) You're including Google, okay. (00:05:38) But I was excluding, you know, a lot of the returns to this GPU spend have come from, you know, better recommender systems at Facebook and Google. (00:05:48) Amazon, better ad targeting, better ad measurement. (00:05:52) Sure. (00:05:53) So I was excluding that and just narrowing it to LLMs, which I think- Tokens. (00:05:57) strictest possible definition. (00:06:00) And it seems like there's going to be a really strong ROI this year, even excluding what are still some of the most economically important and profitable use cases for GPUs and AI infrastructure. (00:06:12) I do think what Karpathy is working on, recursive, (00:06:15) Self-improvement is really important, and unlocking that in continual learning, maybe the two final frontiers for AI. (00:06:26) And just the idea of recursive self-improvement, that the model, while it is training, during a forward pass, has input into its training, or another model has input into the training. (00:06:42) I think that could be really powerful. (00:06:45) And I think Chamath's statistics of 10Xing every year might seem conservative if that comes to pass. (00:06:55) And then, of course, continual learning is the holy grail, where the model learns from experiences the way humans do. (00:07:03) And that's something we haven't unlocked yet. (00:07:06) And those two combined, I think, would (00:07:11) They might pull the future forward in a very real way. (00:07:14) Yeah, and we have right now, Anthropic has a decent lead on everybody else, whether it's three months or six months. (00:07:19) Obviously, they're probably 6, 12 months ahead of open source. (00:07:22) Maybe they're three, six, nine months ahead of their contemporaries. (00:07:24) But they have a lead. (00:07:25) You put Karpathy in there, Friedberg. (00:07:28) Now you have Karpathy. (00:07:29) He does recursive. (00:07:31) And at some point, and it may have even occurred at Anthropic, the AI is going to be improving (00:07:39) the language model more than the humans in the loop are doing it. (00:07:46) Obviously, they're orchestrating at Friedberg, but at what point do we think this, let's call it super recursiveness, occurs? (00:07:53) When will we cross the recursive valley and AI is doing more to build a language model than humans are? (00:08:02) I'm not sure when this idea that you feed the whole model into a context window to train itself and build a new model is going to happen, but I think there's probably a lot of different architectural paths that could be walked here. (00:08:16) One of which is this idea that you could make much smaller models and then create networks of smaller models that work together where you ultimately have less energy or less cost per token produced out of an aggregation of models. (00:08:29) than you did with one single large model. (00:08:32) I've said this probably three or four times now. (00:08:34) There's a lot of work and a lot of opportunity ahead in kind of re-architecting models and re-architecting how models work together to solve problems. (00:08:43) My guess is a lot of leadership that he can bring to exploring those paths. (00:08:49) And all it takes is a minor breakthrough and your cost per token drops in half. (00:08:53) That's a tremendous efficiency gain. (00:08:55) that seems very much on the horizon, because some of the early papers, I think I shared one from MIT a few weeks ago, indicate that there's a lot of room to run here in terms of re-architecting models and deployment of models. (00:09:07) These very small models, small language models, and then verticalized ones are the future. (00:09:14) We've got a company, Abacus, that's doing it for corporations, crushing it. (00:09:18) Everybody's got an interest in doing this. (00:09:20) And I don't know if you saw the news this week. (00:09:22) Chamath, it happened about two weeks ago, very quietly. (00:09:25) Chrome included Gemini, or Google included in their Chrome browser, the Gemini Nano model, without telling anybody, 4 gigabytes on your computer. (00:09:36) And that's the one that does like proofreading, spelling, autocomplete, all that. (00:09:40) So now we have Google. (00:09:43) Covertly installing this on everybody's operating system. (00:09:46) Oh, hold on, hold on. (00:09:47) Covert's a strong word, so let's not use that word. (00:09:50) Without telling people, without giving people a heads up. (00:09:53) Let's say it in a way that we can both agree. (00:09:56) We're in the phase now where I think breathlessly talking about every model improvement is a waste of time. (00:10:02) There's no ROI in it. (00:10:05) We are on a path of accelerated learning, and we're going to start to see end user achievements that were heretofore impossible. (00:10:14) That should be the focus. (00:10:16) So for example, we were able to solve, I'm just collectively saying we, in this case, it was specifically OpenAI. (00:10:24) By the use of a human, and this is important, a math problem that stood outstanding, not been solved for decades and decades. (00:10:34) I can tell you in a different example, there are drug candidates that are about to enter clinical trials and INDs that were sitting on the shelf and people didn't think were very viable at all. (00:10:45) So we're at the phase now where these things are front and center. (00:10:49) They're useful to people. (00:10:50) They're increasingly valuable. (00:10:53) I think what we should do now is focus on these end user use cases. (00:10:58) Because the way that you say it, in my opinion, is part of the problem because it starts to create this boogeyman, us versus them thing. (00:11:06) And I'm not saying you're doing it on purpose, but I'm saying this is exactly why I think. (00:11:11) So many people are becoming sort of like, it's a four-letter word now when you mention AI, because it's presented as this thing. (00:11:18) And I think we have to present the other side of it, at least so that people have the data. (00:11:21) So I don't think Google is in the business of doing or shady things. (00:11:26) They're not that company. (00:11:27) There are other companies that would. (00:11:28) Meta, you're referring to your alma mater? (00:11:33) I'm not going to say which ones, okay? (00:11:35) But Google is not that company. (00:11:37) So I think that the reason they did it was probably because there's user utility. (00:11:40) And (00:11:41) My point is we should focus on the user utility, because I think that's the story worth telling from now on, because I think we collectively, the four of us, can responsibly tell both sides of the story in a well-balanced way, because I think nobody wins if we become Luddites and go back in time. (00:11:56) And I think that if we don't, if we're not careful with our words, that's what will happen. (00:12:00) Yeah, and by the way, obviously not a Luddite, but this is what has been reported by a lot of folks that people were (00:12:10) surprised, shocked when they saw the size of the model being done in the background. (00:12:15) And it has triggered some people looking at it around privacy. (00:12:20) And I do agree that Google is not a bad actor in the space. (00:12:24) So probably a speed error more than anything, I would say. (00:12:28) Maybe just add two things. (00:12:31) There's attorney maxing and then there's attorney maxed. (00:12:35) And Google is probably attorney maxed. (00:12:39) It happened for a long time. (00:12:40) And the second thing I would say is I do think it's incumbent on all of us as Americans who are involved in the technology industry in one way or another to be advocates for the positive, optimistic possibilities that AI introduces to everyone in this world, because it is starting to feel or seem (00:13:03) like there may be a CCP funded campaign against AI and data centers in America. (00:13:11) And that's very logical for China, but it is not good for America. (00:13:16) And so I just, I think it's, we all have responsibility is what I would say. (00:13:22) Yeah. (00:13:23) Who do you think's doing a poor job at that and responsible for this? (00:13:26) Is it Dario with his constant, hey, everybody's going to lose their job? (00:13:32) Who's responsible for this? (00:13:33) Is it the CEOs blaming AI for their layoffs? (00:13:37) What's your take on this, Gavin? (00:13:39) Look, hold on a second. (00:13:39) Everybody is trading their own book. (00:13:43) It makes enormous sense for Dario to try to create the boundary conditions for a regulatory moat, because he will be inside of the tent pissing up. (00:13:53) He's big enough now. (00:13:54) And if you notice that a lot of the breathlessness has ramped up, (00:14:00) and Jason, we've talked about this, you can annotate successive rounds of fundraising and successive scale with the volume. (00:14:07) So I think that it's a reasonable business strategy. (00:14:10) And I think that he's quite clever. (00:14:12) And I think that, look, if you actually, and I do this, if you actually just have a Nash bot, a Nash agent inside of Claude and you ask it what it would do, it would come up with this strategy. (00:14:22) And meanwhile, there are other versions of other counter strategies and counter exploitative strategies. (00:14:27) The point is that each CEO has a (00:14:30) a clear incentive, they're operating at such a level of scale that they're just reading their own book. (00:14:37) So it's up to us to take a step back and actually see the forest from the trees. (00:14:42) I think, Nick, can you find this? (00:14:43) There was a clip of Sham Sankar, friend of the pod, fabulous guy, the CTO of Palantir. (00:14:48) And he was, I think he was on Fox News. (00:14:51) And he said, stop breathlessly asking these model makers what they think. (00:14:57) Go to the end user and ask the person in the factory that's using the model and ask him what he or she thinks. (00:15:02) Ask what the doctor thinks, ask what the scientists think, and start to tell those stories. (00:15:07) That's what we should be talking about. (00:15:09) Yeah, and Gavin, you were, I was sort of asking you your opinion on what's. (00:15:14) Who's causing this? (00:15:16) And then what's the solution? (00:15:17) Do you have folks you think in the industry who are representing it particularly well? (00:15:22) We can point out, hey, Elon has said we're going to move to a world of incredible abundance and working will be optional. (00:15:29) I think that's on the margin, a little scary for people to hear because they hear no job. (00:15:34) But he does say, hey, universal basic income is probably going to have to come into place. (00:15:39) And he said that multiple times. (00:15:41) You have Dario, according to Chamath, (00:15:43) talking his own book, scaring the bejesus out of people in order to get regulatory capture. (00:15:48) What do you think is going on here? (00:15:51) And how do we do better as an industry? (00:15:53) I think Jamath outlined a very viable and positive path forward, where just real people who are not at the tip of the spear, these are the positive impacts AI's had on my life. (00:16:07) I was at an event maybe 10 days ago, (00:16:12) And someone who runs a hedge fund, his daughter was born with a very rare genetic mutation that effectively would have normally condemned her to a life devoid of joy, meaning everything. (00:16:26) The neurons in her brain were not firing. (00:16:29) So she wouldn't, you know, who knows what her life expectancy would have been or what the quality of her life would have been. (00:16:36) And it's a tragic disease. (00:16:39) He said he didn't accept that as an answer. (00:16:41) He found he did an enormous amount of research with LLMs and found an existing safe drug on the market that they thought would have a meaningful impact on his daughter's condition. (00:16:54) And it did. (00:16:56) It took, I think the percentage of times the neurons were firing was 30 or 40%. (00:16:59) And it took it up to 80 or 90%. (00:17:01) And that means that she can live a normal life. (00:17:05) She may not be as smart as she would have been, but she can live a normal life. (00:17:10) And he's now figured out how to use AI, how to further tailor that drug. (00:17:16) And there've been all sorts of advances in protein design, et cetera, et cetera. (00:17:22) And he's reasonably confident he's going to have a drug in months that is a complete cure. (00:17:28) And that's just one person, one dad who was unwilling to accept defeat for his daughter. (00:17:35) and who changed her life and the life of everyone else with that disease. (00:17:38) And we tell those stories. (00:17:40) So I think Elon's doing a good job, a future where work is optional. (00:17:43) I think that sounds great to some people, scary to others. (00:17:47) Four day work week, I think is probably something that sounds good to a lot of people. (00:17:52) I think Jensen is doing a good job of being an effective advocate. (00:17:56) And I do think anyone who is trying to drive, I just, (00:18:02) We need to stay focused on the positives as well. (00:18:05) Yeah. (00:18:05) Freeberg, what's your take here on the AI PR crisis, if we'll call it that? (00:18:13) We had three different commencement speeches that were booed. (00:18:19) Eric Schmidt being one of them, two other ones by maybe less notable folks. (00:18:24) When you hear young people booing AI vociferously, why are they doing that, Freeberg? (00:18:29) And what's your take on the (00:18:31) overall PR problem and how to turn it around. (00:18:36) That's a, there's a long answer to that question. (00:18:43) It relates in some ways to your concerns about socialism and polarization. (00:18:47) What's the long answer? (00:18:48) I mean, that's like, why do people hate technology? (00:18:57) I mean, this technology, they love their phones, they love the internet, this technology they hate. (00:19:04) I think that there's like an underlying view that technology creates leverage for a small group of people, which creates power imbalances. (00:19:14) And nothing represents that more than AI, that a small number of people that control and (00:19:23) profit from and benefit from AI are going to end up getting outsized returns relative to the broader population, that the time to diffusion of the technology, because ultimately all technologies commoditize and diffuse, but the time to diffusion here is such that it's going to be extremely asymmetric for society. (00:19:47) And I think that there is something fundamental about that. (00:19:50) It's like, nuclear bombs, I think, really created this moment in people's minds in the mid-20th century that by the back half of the 20th century gave everyone a high degree of skepticism about (00:20:00) technology and science generally, that those who have the knowledge and those who engineer solutions with the knowledge can create outsized advantages for themselves. (00:20:10) And it puts the rest of us at risk, the rest of the world, the rest of the population at risk. (00:20:14) And because those questions about when does this benefit me, how does it benefit me, can't be answered today. (00:20:21) The economic benefit that's accruing to the few today becomes the narrative. (00:20:26) It becomes the story and it becomes this like power system (00:20:30) that a few people take from the many. (00:20:33) And so there's something deeply disturbing for the average person about that. (00:20:37) They don't understand how it works, why it works, what it'll do for them, when it will do it. (00:20:41) And all that they're being told is that some people are making trillions of dollars. (00:20:45) So I think that it's pretty obvious why this has got such a backlash. (00:20:48) Secondly, I think that there's a deep amount of external energy that's fueling this anti-technology sentiment in the United States and has been for decades. (00:20:57) I think to Gavin's point, I don't think it's just China with NGOs today. (00:21:01) I think that there is a long history of state actors intervening in media activities in foreign nations to try and create the sentiment and fuel a sentiment that reduces progress in that competitive state. (00:21:18) I think this goes all the way back to KGB design during the Cold War and (00:21:24) It's been refined and honed and improved over time. (00:21:27) This is not just some conspiracy theory. (00:21:28) There are plenty of great books about this. (00:21:30) The techniques of what's going on specifically today, I don't know enough. (00:21:34) I don't have any great details on that. (00:21:36) But I don't think that there's no foreign interest in seeing technology advancement slow in competitive nations. (00:21:43) The United States probably does similar things to other nations. (00:21:47) And I think that that's probably a key part of this. (00:21:50) And then I think this like third piece is like when the Copernican revolution happened, it was a mind, like heliocentricity was a totally new way of thinking for humans. (00:22:02) And it was deeply disruptive to the church. (00:22:05) And it was deeply disruptive to the power centers, which were the centers that could tell people Earth is at the center of the universe, we're in control, we're the direct channel to God. (00:22:16) And the idea that the sun is at the center of the solar system and we spin around it and we're a tiny speck in the universe was very hard for people to grasp. (00:22:23) There's something about AI that's very like not human-centric and it kind of shifts and fucks with the ego of the human. (00:22:31) It's almost anti-humanist. (00:22:33) And I think that that's like a deep psychological current for a lot of people and their disdain for this technology. (00:22:40) It fuels it. (00:22:41) It's not the cause, but I think it fuels it. (00:22:43) So I think there's a lot of complicated aspects to this, J. (00:22:45) Cal. (00:22:46) I don't think there's like a simple put Shyam on a podcast and he'll solve the problems with AI right now. (00:22:51) I think that there's a real set of shifts happening and there's a real set of global competition underway where, you know, various state actors and interests are competing with each other. (00:23:02) Trooper, do you think that we should slow down? (00:23:04) I don't think you can. (00:23:08) No, Do you think we should slow down? (00:23:10) No, I think I was just talking to some people on Zoom right before this, but I think after the Manhattan Project, the research labs were stood up to maintain our scientists that worked on the Manhattan Project from effectively leaching back or leaking back to Russia and Germany and other places that were adversary to the United States. (00:23:30) And they all were against the nuclear bomb. (00:23:33) They worked on it because it was necessary for the United States security. (00:23:37) But then when Russia got a hold of the secrets, they were leaked because people were worried that if the US had all the power, there would be no counterbalance to the power. (00:23:46) And so the nuclear secrets were leaked to Russia for that purpose. (00:23:49) Then when Russia had the nuclear secrets and they began developing hydrogen fusion bombs, and it was clear that they were going to race ahead, the United States raced ahead with developing nuclear bombs as a counterbalance to Russia. (00:24:02) When the proliferation began, there was no stopping it. (00:24:06) began, and you had to have this balance in the world. (00:24:08) Otherwise, you have effectively an asymmetric power that can do whatever it wants globally. (00:24:14) I think there's that moment in the world right now where if the United States does not advance its AI technology, the availability of it, TBD, industry, taxation, all these things that we're talking about doing, there will be someone else that will. (00:24:29) And if someone else does, (00:24:32) We can go through what would happen. (00:24:34) There's a complicated game theory on this, but what would happen if China had sufficiently advanced models and sufficiently advanced scaled deployment of those models relative to the United States? (00:24:44) As you do that analysis, you realize, wait a second, that's probably not a healthy place for the world to be. (00:24:49) It's also probably not a healthy place for the United States to be the only one with AI. (00:24:54) And so I think what we end up seeing is if we do try and slow down AI, (00:25:00) we kind of lose this moment of balance that's necessary when you have a technology proliferation, like we saw with the arms race after World War II, that, you know, we're going to see again here. (00:25:12) I think bring up a good point. (00:25:15) should we slow it down or could we slow it down? (00:25:17) There actually have been some discussions about ways to do this. (00:25:21) One of them would be, hey, with self-driving, people are scared that all these cab drivers are going to lose their jobs, Uber drivers, cab drivers, bus drivers, truck drivers. (00:25:31) This is, you know, over 10 million people in the United States driving things for a living. (00:25:35) Would you be in favor of some of the announcements that will be (00:25:41) a paste rollout, it won't happen all at once. (00:25:43) In other words, those people will be giving some amount of job security to stay behind the wheel with it. (00:25:48) Another example that's been given is if you put Optimus into Amazon factories or the Figure robot just did like a week of just sorting packages, I'm sure everybody saw that video, we'll insert it here. (00:26:00) that figure robot sorting thinks, hey, if Amazon deploys those, there'll be a tax on those per hour and we'll tax humanoid robots in some ways and then use that for, say, retraining people. (00:26:11) Those are two very specific conditions and approaches that people have been promoting. (00:26:16) Do either of those resonate with you in any way? (00:26:17) I think it's interesting that in all of those discussions, I've yet to see an actual survey of only the truck drivers and only the package sorters. (00:26:30) The question that I would have is, do the people that do these jobs want these jobs? (00:26:35) And if they do, then there's a reasonable claim to make to keep those jobs the way that they are. (00:26:40) If you're saying, this is the job that I do, I love it, I'm able to provide for my family, great. (00:26:44) That's a very different argument than, well, you know what, Amazon has 35 or 40% churn inside of their warehouses. (00:26:52) And we should probably ask the question, why is that? (00:26:55) Because if it was such a great job, I suspect the churn would be 3 or 4%. (00:26:59) So what exactly is it that we want to protect? (00:27:04) And have you asked them? (00:27:05) And I think that this is just, again, a bunch of people in the peanut gallery who want to take a moral high ground and try to make some other group of people feel guilty or feel bad. (00:27:17) At no point are we actually asking the conversation that we should be having, which is, it's interesting to me that there was supposed to be an EO, a presidential executive order that was announced today. (00:27:29) And then it was pulled, it was scrubbed at the last minute. (00:27:32) Did you guys notice that? (00:27:33) And yesterday, what was leaked was everybody that was attending, it was all the big NeoLabs CEOs, and it was all the big hyperscaler CEOs, including friend of the pod, Nikesh Arora, shout out to Nikesh. (00:27:44) And then it was scrubbed an hour ago. (00:27:47) Why was it scrubbed? (00:27:49) And the president said that there were aspects of the bill that he didn't agree with. (00:27:53) And as far as we can tell, the aspects would have required some amount of (00:27:58) supervision, insight, review from the federal government. (00:28:02) Of language models, specifically, of these frontier models, is what I read. (00:28:06) Not language models, because I think just of AI, because there's going to be many different kinds. (00:28:09) They're not always going to be language models, but of AI. (00:28:12) So look, I think Friedberg is right. (00:28:14) We are in a proliferation with China. (00:28:17) I think it's actually good that China is less than nine months behind us. (00:28:21) I think it allows us to find a detente where we have a certain magnitude of capability that they also have, and that allows all of us to then seek peace and abundance. (00:28:33) And the fact that we are orthogonal societies, we are organized differently, increases the probability of finding peace using the Rene Girard kind of framework of mimetic theory than if it was like us in another country that was exactly similar to us. (00:28:48) So I think what we need to do (00:28:51) We probably need KYC. (00:28:53) I think that should be something that us in China get together and say, you don't want it to get into the hands of people you can't control. (00:29:00) You probably already KYC those models anyway inside of China. (00:29:04) You already review those training runs before you allow these models to get released. (00:29:07) We already know that that's happening. (00:29:08) Yes. (00:29:08) So we should probably do some sort of KYC so some crazy person doesn't create some biological weapon. (00:29:14) I think that those are some reasonable ground rules, but otherwise Friedberg is right. (00:29:18) You have to take a little bit of a deterministic view here, which is that we are in this existential race and we need to get to the place where each of us, meaning us in China, can look each other in the eye and say, all right, weapons down, so to speak. (00:29:32) Cavin, I'm going to hold you to answer 2 questions. (00:29:34) One, should we run frontier models? (00:29:39) Because that's specifically what was mentioned in the leak about the EO frontier models, the powerful ones. (00:29:43) Should they be run through some sort of testing before they're released? (00:29:46) And should there be some regulatory framework for that? (00:29:49) That's my first question to you. (00:29:50) Yes or no question. (00:29:51) And then you can explain your answer. (00:29:55) Geez, like I just think it's such a complicated topic. (00:29:59) It feels (00:30:01) We're a little early for that. (00:30:04) I don't love the idea of the United States doing it and no one else doing it. (00:30:11) I like I think in a world where we hold hands with China, look, I think that's much more palatable and we are aligned and we trust each other and have kind of verification capabilities. (00:30:27) I do think... (00:30:27) Great. (00:30:28) Yeah. (00:30:28) Let me then rephrase it. (00:30:29) Should China and the U.S. (00:30:31) come up with a simple battery of things that have to be tested before these go out, including bioweapons, terrorism, and that genre of, in that vertical of just really known dangerous things, just like the FDA might test for poisons or contaminants in a food or a drug. (00:30:49) Would you be in favor of that? (00:30:50) I'm curious. (00:30:52) So a few things, like the one thing that's great about America is there is, or one thing that's just (00:31:00) We have other forms of regulation. (00:31:03) Self-regulation, sure. (00:31:04) We have self, one, we have self-regulation. (00:31:06) Also, we have the courts. (00:31:08) And if an AI model company behaves irresponsibly, they know that there are ways that people who have been harmed can seek recourse. (00:31:19) And so we already have a system that encourages responsible behavior on the part of the model makers. (00:31:28) That's a great point because OpenAI is being sued right now by a kid who killed themselves after talking to OpenAI's model. (00:31:34) So you're correct in that, yes, after the fact. (00:31:36) And we'll see, we'll see more of that. (00:31:38) I just, you know, to me, once you give something, give a power to the government, (00:31:45) It's almost never taken back and it begins to grow and it's kind of a one-way, a one-way path. (00:31:52) One-way ratchet. (00:31:54) Yeah, second question then. (00:31:56) Chamath was saying, hey, nobody listens to the, you know, these cab drivers or maybe the people sorting the packages. (00:32:02) Do they want the jobs or not? (00:32:03) Actually, the UK (00:32:04) There was just a 60 minute special and UK and also Boston and New York are pretty adamant that they want humans to stay and they want to ban self-driving in those locations or severely limit it or maybe limit it in some way to let those people keep their jobs. (00:32:20) How do you feel about that possibility? (00:32:22) Is that something you think society should be open to? (00:32:26) They're going to get sued for wrongful death when somebody runs over somebody else and you could have implemented a solution that has a 0 death rate. (00:32:34) That's very different from package sorting. (00:32:36) Okay. (00:32:37) Go talk to the package sorters is what I say. (00:32:39) Go talk to the people inside the Amazon warehouse. (00:32:41) Ask them what they would rather do at Amazon. (00:32:43) Ask them. (00:32:45) Yeah, sure. (00:32:45) But Gavin, what are your thoughts here on either one of those examples here? (00:32:48) I think going to a city where you can't get in a Waymo or a cyber cab. (00:32:53) is going to feel barbaric and unsafe. (00:32:55) And to you, I agree. (00:32:57) I don't know if you remember, but the early days of Uber, sometimes you go to a city where there was no Uber. (00:33:03) Yeah, it'd be incredibly frustrating. (00:33:04) Well, I'm not going to come back until they have Uber. (00:33:07) It's so inconvenient. (00:33:09) And I think, so whatever individual municipalities decide, I do think, one, Chamath's point is really powerful. (00:33:16) There's 50,000 automotive deaths per year in the United States, if I recall correctly, and a million (00:33:23) globally, that's not tolerable. (00:33:26) And there will for sure be wrongful death lawsuits. (00:33:28) And then just from a convenience and quality of life perspective, I just don't think it's going to persist. (00:33:35) And that's another great thing about America is, you know, you have this patchwork of different states and municipalities and each one doing things in a different way. (00:33:43) And I'm not suggesting that's good for AI, but it does tend to, you know, (00:33:50) has historically, the curly effect aside, led to, I think, more positive outcomes where cities and states compete with the curly effect being. (00:33:59) It is fair. (00:34:01) This is a really important point you're making, Gavin. (00:34:04) With flock safety as but one example, we had an AI, there's an AI tool called flock safety. (00:34:09) It's cameras that use AI, monitor people who are committing crimes. (00:34:13) There's a privacy issue around it. (00:34:14) is bottom up. (00:34:15) You just do it by town. (00:34:17) It's not top down and states can regulate it. (00:34:19) Same thing with (00:34:20) probably happen with self-driving. (00:34:22) And states will probably have some say in how AI is deployed, even if maybe some centralized governments don't want to do that. (00:34:29) I really think this only comes down to... (00:34:31) The block safety thing, I think, is so good, Jason. (00:34:33) Crime is now a choice. (00:34:35) Yeah. (00:34:35) You know, I think that the Cambridge state has voted to turn off gunshot detectors 2 days ago. (00:34:43) Wait, which city did that? (00:34:44) Cambridge. (00:34:45) Cambridge. (00:34:46) Cambridge, Mass. (00:34:47) That's the place where Harvard is. (00:34:48) That's the place where Harvard is. (00:34:49) So the geniuses coming out of Harvard in that town decided gunshot detection shouldn't occur. (00:34:58) We don't want gunshot detection. (00:35:00) It's wild because, you know, there's a theory that it disadvantages, you know, that it might lead to an illegal migrant. (00:35:07) who's shooting a gun, being apprehended, and we don't want that. (00:35:10) Got it. (00:35:11) And A16Z had a great essay on flock. (00:35:15) We can really, really solve crime, and it's just a choice. (00:35:20) And different states and municipalities will make different choices to be pro-crime or anti-crime. (00:35:26) And I'm sure they don't cast it as pro-crime. (00:35:28) There's an, you know, (00:35:30) some sort of moral or ethical reason. (00:35:33) They're making that choice. (00:35:34) But people will vote with their feet over time, and then voters will vote with their votes. (00:35:39) And we'll see what works. (00:35:41) Have you guys been to Vegas recently? (00:35:43) My wife and I went to visit Vegas, and we spent the afternoon with Ben Horowitz and his wife, Felicia. (00:35:49) She has done this incredible job with the Las Vegas Police Department. (00:35:53) It is one of the most impressive things I've ever seen. (00:35:55) And to your point, (00:35:57) crime is an option and they've said no. (00:35:59) So what happens there is they have gunshot detection, they have drones that get deployed off the roof of the police building. (00:36:05) We were sitting inside of mission control where you see it happening, Jason. (00:36:08) If something happens, (00:36:10) They have eyes on site within minutes. (00:36:12) They can track offenders and bad guys all the way to wherever they're hiding. (00:36:18) And you walk out of it and you feel incredibly safe, like they're really on top of it. (00:36:22) And when you understand the level of investment, it's not, it doesn't take billions of dollars. (00:36:27) It's de minimis compared to the cost of the crime. (00:36:29) It's de minimis. (00:36:30) It's de minimis. (00:36:30) Especially when compared to the cost of the crime occurring. (00:36:33) Exactly. (00:36:34) If you gave the Las Vegas Police Department (00:36:37) 30, $40 million a year, it would be the safest city in America. (00:36:41) And that's all it would take. (00:36:42) Yeah, exactly. (00:36:43) And for the privacy concerns, Freeburg, there are very simple solutions to this. (00:36:47) I am a privacy advocate myself. (00:36:49) Of course, we all want some level of privacy. (00:36:51) I had the Flock CEO on This Week in Startups twice in the past 10 years. (00:36:54) He's very considered. (00:36:56) And the way they do it with Flock is (00:36:59) They allow you to have a rolling database. (00:37:01) And I think there's a maximum you can save the license plates for. (00:37:05) And they don't do facial recognition. (00:37:07) I don't see why not, but let's put that aside. (00:37:09) You can only keep it for two or three years. (00:37:11) And then they insist on having an audit trail in it. (00:37:13) So there are all little things you can do on the back end to protect privacy with audit trails, et cetera. (00:37:17) We got a lot more to get to in the docket. (00:37:18) I just want to give my final thoughts on what we're talking about here in terms of the AI problem and the PR problem. (00:37:26) I think we have to recognize that the layoffs that are occurring in big tech and in a lot of these places are not just the bloating issue anymore. (00:37:36) And I'm just going to point to two factors that I think are scaring the bejeezus out of people. (00:37:40) And we just have to admit that this is occurring as opposed to we've been debating it here. (00:37:45) Is it occurring? (00:37:45) Is this just cover? (00:37:47) And are we AI washing? (00:37:48) The first one I want to give you an example of is Matthew Prince. (00:37:51) who's the CEO of Cloudflare, incredible company, public company. (00:37:55) Two weeks ago, I laid off more than 20% of my workforce. (00:37:57) I didn't do it because Cloudflare is struggling. (00:37:59) We posted record revenue growth, have strong free cash flow, and are adding an unprecedented number of customers, yada, yada, yada. (00:38:06) And he says basically, he's getting rid of measurers. (00:38:09) Measurers (00:38:10) are the people who manage people and who measure data. (00:38:14) And he just says, we're getting rid of all those people. (00:38:16) They're unnecessary because of AI, and we'll be adding people in other positions. (00:38:21) At the same time, Zuckerberg did another round of layoffs, and they were done in a way that people felt was not considered and a bit, what's the word? (00:38:34) Dystopian? (00:38:35) Dystopian, thank you, sir. (00:38:37) He did them in a pretty dystopian way. (00:38:40) Here's Zuckerberg for 36. (00:38:42) In general, the average intelligence of the people who are at this company is significantly higher than the average set of people that you can get to do tasks if you're working through the contract, through these contractors. (00:38:55) So if we're trying to teach the models coding, for example, then having people internally (00:39:02) build tools or solve tasks that help teach the model how to code, we think is going to dramatically increase our model's coding ability faster than what others in the industry have the capability to do who don't have thousands and thousands of extremely strong engineers at their company. (00:39:18) Okay, so what Zuckerberg did at the same time, concurrently, he told everybody, we're laying off these 8,000 people. (00:39:25) A lot of those people are incredibly talented. (00:39:28) Some of them are on H-1B visas, creates all kinds of chaos for them in their personal lives. (00:39:31) And obviously they're having record profits there as well. (00:39:35) At the same time, he was laying off those 8,000 people. (00:39:38) This is after 10s of thousands of layoffs before, which were obviously because of bloating. (00:39:42) He said, we're putting recording software on every single person in the company's computers to study and train our model. (00:39:48) And people were like, oh, and previous people said, I built (00:39:51) During the AI hackathons they had months ago, I built all this AI tools to make my job more efficient, and then Zucker.