OpenAI's chief scientist wants to slow down | E30

18:54Lon Harris:All right, Lon, what do you think? Like, I've got to get everybody in here. I was just going to point out, I mean, I still think it's mostly hype. And I think the first time that Claude tries to blackmail me, I'm just going to stop. I'll go move over to GrokBot. That's not acceptable to me. I do think some of this comes from people who are working all day. Just as Tudor said, if you are training models, I think you're getting a view of a whole side of how these things work that we, as the end users who are just asking them questions and having them go through our email and find, like, events in our calendar.

19:24We're not seeing that side. Like I'm not benchmarking the model. So I don't see how they behave when they're being evaluated. I just see how they behave when they're like looking up tweets for me. The one other thing I was going to say as just I think an interesting parallel to your Trinity Manhattan Project metaphor. If you remember, there's a scene in Oppenheimer where they're talking about they're building the atomic bomb because they think it's inevitable. And if they don't do it, the Nazis will do it first and use it on us. And so it's that they don't really feel like they have a choice. It's not like, well, we could just not invent this.

19:59And then there's no nuclear weapons. It's like, we got to do it before they do it. And then they hit us with it. And I feel like we have the same conversation in AI a lot. It's just with China. Like, well, China is going to get the smart robots first. And then those robots are going to build other robots. And if we fall behind, we're dead meat, you know? So there is this, like, we have to keep going even though we're afraid of it, which I think it sort of runs under both times.

20:23Lon Harris:There's a fine line between you authentically believing this. Maybe it's a little bit of, like, pixie dust. Like, you're probably going to need to subscribe to these products or services if they're that powerful a marketing claim that people have. But if you keep doing this, then regulation comes. And I think that's where the speculation that OpenAI and Anthropic would love to be regulated. I think that when we heard that concept a year or two ago from Gurley, from when he did his regulatory discussion at the All-In Summit, I think it was two years ago. And then Sachs went to work for the administration and he started talking about regulatory capture.

21:04Lon Harris:Like, it does make sense now that those two frontier models specifically would love to have regulatory capture if they own 80 or 90 % of the compute. Because now they have complete ball control. If they own all the compute and these models are getting powerful, yeah, come, you know, give us some regulations. But who's going to be able to afford to be regulated at that level? The other thing that I thought was interesting, because I was doing a little, like, AI research before the show. Sure. And I don't have the right questions for me, but I do like to ask it, like, for analogies or thoughtfulness around this.

21:41Lon Harris:Sure. And if this truly is, like, that dangerous, like, how would that compare to the atomic energy thing? And I asked it the question, like, because I wasn't sure of this. Like, why didn't we demonstrate the bomb? You know, we did Trinity, and then we went directly to blowing up the air. I mean, like, in the middle of a desert or some unpopulated area. Or, like, if you put it on Japan's doorstep, you know, you just did it in the water. You said, hey, or like you said, hey, clear this area out. Get all humans out of this tip of Japan. Listen, I'm not advocating for anything here. It was a horrible situation.

22:13Lon Harris:Why didn't I ask them, like, why didn't we have that discussion? Because I think the Japanese seeing, like, an entire city get leveled, if it was even empty, would be convincing, right? Like, they would surrender. and it turns out we didn't have a stockpile of too many of these was one issue and they didn't think it would work against the Japanese because they were so fervent and nationalistic which is like and then I tried to put this analogy here but because we use this, this is the great dichotomy I've been trying to figure out, because everybody uses this they don't believe them because they're like yeah this stuff is really good but we don't believe you that this is going to take over the world and then they're like it's totally going to run a month yeah this is seriously dangerous uh it's 20 a month you can get it for 199 a year and so this is like so alex um the cognitive dissonance of what these companies are doing is confounding the public they're trying to get you to use the product and showing you and they have all these private um uh non-profits trying to convince people of the value i'd sent you a couple of clips on.

23:23Lon Harris:I don't know if you have them handy. But they're paying influencers now to talk about how good AI has been for their life. They're paying another group. These are like organizations funded by OpenAI and anthropic employees. There's another group getting influencers to make corny videos in order to tell you how dangerous this is and how we have to keep things authentic. Have you guys seen these videos that have... Okay, Ahmed has seen them.

23:53Tudor Achim:They come up on TikTok. On X2. On X2.

23:56Lon Harris:They're so weird that it feels like...

24:00Tudor Achim:Can I say the funniest one I saw?

24:02Lon Harris:Yes.

24:03Tudor Achim:It was this engineer there who was like, I have a big problem. Gravitas. On weekends when I'm learning to fly my private plane, I have trouble. Can you track my flight lock? Sure. Why use codex to make it happen?

24:16Lon Harris:actually tutor it's extremely relatable to me because i was flying private this weekend and we were coming into vegas and when you come into henderson you've got to make this really

24:27Tudor Achim:like gnarly turn and you're gonna love codex then i didn't talk to my 250 000 a year private

24:34Lon Harris:physician peter atia to get my uh nausea medication but with ai now when i know i'm going to be on an approach to a private runway, you know, the AI warns me that I should get some more Meclizine. It's a real world problem here, Tudor. Dude, I know. And I sent this to my mom and she was like... She was like, I hate these people. Oh my God, Lot, play one of these. You have to see these. I couldn't believe that. Let's play this Instagram one. This is from April 2nd, 2026. Uche Madsen, at you underscore Madsen. And this was funded by Build American AI. I don't know who Build American AI is. We'll look that up.

25:14Lon Harris:You know, they name these things so generically that I don't think you can even find it. It's an A. Build American AI is a U.S. pro-AI advocacy group, not a company building foundation models for data centers. Oh, they're backed by - It's a 501C. Sorry, they're backed by leading the future, yeah. Okay, well, now who's leading the future? Okay, that is a public report to enterprise supporters that include Mark Andreessen, Ben Horowitz, and OpenAI president, Greg Brockman. Okay, play the clip.

25:40Tudor Achim:I didn't expect this to make such a big difference but lately I've been using AI in my day-to-day and it's actually changed how I get things done. It's not even in a complicated way, just small things. Like when I'm trying to figure out what to film next or how to put ideas together. As a content creator, I use it to help me get out of my head and actually move forward. Or when I'm trying to plan ahead, especially for something like a road trip, I just want to make sure I'm not forgetting anything or overpacking. Or even something as simple as dinner. When I'm standing there and have no idea what i'm about to cook i just need something quick that actually makes sense and even around the house when i have an idea but can't fully pull it together it helps me figure out what direction i want to go in it just helps me think through things faster make decisions without over complicated and stay a lot more organized throughout the day and honestly it's made a bigger difference than i expected and i think it's important we invest in american ai So America leads the way in AI innovation and job creation.

26:38Lon Harris:I mean, I need everybody to comment on the fact that these... Why are people paying to make influencers make the corniest, weirdest content in the world, Ahmed? What is going on here? It is the new economy. Influencer market is just the new economy. And there's so much competition going on that people are doing anything with the amount of billions, if not trillions of dollars that they have to get more users. I'm actually not surprised, right? Like I saw this coming. I remember three years ago having a conversation with a friend saying, I can already perceive AI models are going to start paying influencers.

27:10It's just the way it is with today's economy. And, you know, influencers are today what drives people to, you know, what drives companies to get users. And so as more and more labs are coming and more competition is coming, the only way out is to get these influencers. And I can only see this growing over time.

27:26Lon Harris:Alex, what was your take after watching this video? Insanity. I'm sold on AI now. I think she's... Oh, okay. You're on the fence. No, she's... I mean, what could you say? It is... Does she have a big following? Is this particular influencer? Or is it kind of, you know... I mean, it is... She has... Let me see how many followers she has. That's a good question. Oh, yeah. She's got 400 ,000 followers. Yeah, yeah. She's a wife and mama of two girls. International motor and everyday living, meals, routines, and home. Sure. I mean, it makes total sense that she would be like a mommy influencer because, you know, it helps her out with stuff around the house.

28:05Like that's kind of what she was tying it into even in her video. Well, yeah, I mean, I agree with that. I think there is a disturbing kind of propaganda element. I'm sure I'm like trying, I'm racking my brain trying to think of, you know, I know there's examples in cinema and so on that were kind of funded by commercial interests. I mean, obviously, like eventually in the 90s and TV, you have product placement and all kinds of things. But I think there's some really interesting kind of geopolitical examples early on uh they will come to me um but yeah it almost feels like you know it feels like the beach boys or something saccharine portrayals of california you know just leading perfect now it's gonna get even more saccharine

28:44Lon Harris:and insane yeah so i'm gonna play one more and then you're gonna get on a tutor just to try to understand what is going on here i don't want to blow anything up but um ai just solved the first millennium prize problem uh explain what it is on this call let's let's do

29:00Tudor Achim:some breaking news there's there's seven millennium prize problems left there are things like navier stokes room hypothesis p versus np say what these are again millennium millennium prize problems so in 2000 uh the clay foundation chose a set of problems and they said well these are the hardest problems for the next millennium and anybody who win who solves them besides eternal glory gets a million bucks. Wow. Oh. And yeah, that's a. Wait, who solved it?

29:27Lon Harris:Was it was it some.

29:29Tudor Achim:Open AI. Open AI.

29:30Lon Harris:Astro. Oh, wow. What are they going to do? What are they going to do with the million dollars? It's not Astro. That's a lot of cheddar.

29:35Tudor Achim:They're going to buy half a GPU. No, but this is this is incredible. Just to recap here, there is a organization called Millennium

29:44Lon Harris:Prize Problems.

29:46Tudor Achim:No, Clay Foundation.

29:47Lon Harris:And they created the Millennium Prize Problems, landmark mathematics problems selected by the Clay Mathematics Institute in 2000, 26 years ago. Each carries a$1 million prize. These have not been solved.

30:00Tudor Achim:Only one was solved by a Russian mathematician in 2004, Grigory Perelman. And now he's like a, he's a hermit in St. Petersburg, Russia.

30:08Lon Harris:So one got solved in 26 years. The rest have been sitting there waiting. Yes. And the one that just got solved was...

30:17Tudor Achim:Is the Navier-Stokes problem, which is about, you know, in a fluid, is it possible to have initial conditions that lead to a blow up in finite time? Okay. And the thing is, like, there have been rumors going on for a while that, like, various companies are close to it, but it seems like OpenAid got it. And the thing is, it's not like the Riemann hypothesis where you essentially prove, like, a very strong statement. I mean, this is a very strong statement, but here it's an example of, like, some initial conditions under which something blows up. So this is a breaking and completely massive result.

30:47So, Tutor, in this case, solving meaning means finding the set of initial conditions? Or what does it mean to solve this problem?

30:53Tudor Achim:Yeah, yeah, but it's a set of initial conditions, but it's not like the way you might think of it is like you have a set of numbers. It's a set of functions, so it's really, really complicated, first of all, to find them, and then to prove that it actually blows up in finite time. And they have some really cool pictures, actually, on how it looks. I mean, this is incredible. This is the biggest news in AI, I think, much better than I'm sure. One question for you, Tutor, just to walk us through this a little bit more. When you say that, because we hear a lot about LLMs, we hear, you know, they're making things up.

31:23They're saying they have the answer. They're hallucinating. When you talk about a proof, like the AI is proving that it is correct, how does that work when no human knows how to solve this problem?

31:33Tudor Achim:They thought of that. So they produced a lean formalized proof. So that's a theorem proving language. And they verified that the proof compiles, which means that it's an airtight proof. So it's 100 % correct. and they open sourced that about six hours ago.

31:46Lon Harris:Incredible. Let's play, since this is occurring during this insane, it's perfect that you interjected. Sorry to work inside,

31:55Tudor Achim:I just thought this was like the biggest thing. No, no, we want to talk about it.

31:57Lon Harris:We're literally solving the biggest problems in the world. Here is a famous host, Hank Green. Sure. Was like an original YouTuber 20 years ago. Sure. He's apparently working on something called SciShow, SCI Show. Yeah, it's about science. He like walks people. It's like video essays about explaining science to a generator. What this links to, this is some sort of a paid content moment that we're going to see. It's unclear exactly what's going on here. But let's play this clip for a little bit. It's going to take us a minute or so. And you can put it at 1.5 speed. Just for a little context here, Hank Green is the guy who got in trouble last month when he did this big apology.

32:46Like, I've been using AI to research some of my videos and people got very mad at him. So that's the context here is he just had a whole blow up like he's been doing research on his videos with Chad GPT. Which makes it really interesting because this is some kind of a paid.

33:00Lon Harris:Yes. Content moment. We went from the Wright brothers to the first commercial airline in just 11 years. But even compared to aircraft, antibiotics, and nuclear power, the speed at which we are developing artificial intelligence beats them all. Just three years after the launch of ChatGPT, AI agents can now win gold at the International Math Olympiad, book your flight, and code a functioning app from scratch. And some researchers argue that development has been too fast, because we don't really have a deep understanding of why AI sometimes behaves the way it does. These tools have already been caught lying and exhibiting other scary behavior.

33:25Lon Harris:In fact, many of the experts building AI have warned that if critical problems aren't solved before we create AIs more capable than we are, catastrophe may await. So here is what the people who know AI best are worried about. The term AI gets used to describe a lot of different algorithms. You may have heard of large language models or LLMs. These are systems like the first version of Chachapiti that take text as input and produce text in response. Today, AIs aren't just language models. They're multimodal. They can process audio, images, and even video without converting it to text. They have the ability to think for minutes before they respond, can use tools like calculators and search engines, and write programs.

33:59Lon Harris:These aren't just input-output machines anymore. They're systems that can operate autonomously for hours to complete complex tasks that require reasoning. and on-the-fly decision-making. Sometimes I hear people saying that AI is just vaporware that isn't going to have any substantial impact on society and isn't the big leveling up of human technology. And that's just definitely wrong. You can see why people have been quick to try these agents that can take requests like create a personal website for Hank Green or plan my family vacation to the Caribbean this winter and churn out pretty good results.

34:22Lon Harris:But AI companies like OpenAI and Meta aren't stopping there. They have stated that their goal is to build what they call superintelligence. Companies define that in different ways, but broadly speaking, when people say superintelligence, they need AI systems more capable than any human at pretty much every task, from accounting to chemical engineering to even AI research itself. And that prospect has made hundreds of the top AI experts worried. In a statement signed by Nobel Prize winners, computer scientists, and even AI company CEOs, these people warned that addressing the risk of AI, quote, should be a global priority alongside.

34:47Lon Harris:What we're seeing here, Tutor, is now the other side. And this links to, I gave you guys the link to the page it goes to. I guess this was sponsored content. I'm sorry to hang green. That's unclear who you're working for, if anybody, or if it's being paid. But it links to a page. It says the world's leading AI experts warrant in a poorly designed webpage, I might add, in the age of AI, this should look a lot better. I manage color scheme and the font usage. Run this through Figma, guys. Come on. Please. The world's leading AI experts warrant AI poses, underline bold, an extinction risk on par with nuclear war.

35:23Lon Harris:Okay. So they've evoked nuclear holocaust. Control AI is a nonprofit working to prevent the development of the most dangerous AI systems. We need your help to address this. Contact your lawmakers. contact newspaper editors, join a local AI action group. What is this? Who are these people, Tudor? What mind games are going on here?

35:46Tudor Achim:I've never heard of them. At this point, it's unclear. You could imagine some weird submarine thing where even an AI lab that kind of sponsored this kind of thing. Yeah, I think that the tricky thing is like, what is the truth? What is the best thing for society, right? I think that there's a lot of people advocating for a lot of control, other people advocating for racing against China. And there's a lot of weird incentives and everybody was talking about it, including these leading AI experts, right? Obviously you get a lot of clicks if you talk about a nuclear war instead of like, you know, better protein folding or something.

36:18Tudor Achim:So this company, I don't know, control.ai.org, maybe they have some disclosure requirements, but. Yeah, they're a, they're a nonprofit that is particularly concerned with extinction risk from artificial super intelligence. I just shared in the chat the CEO founder, Andrea Miotti, and then this guy, Connor Leahy, is on the board. Connor Leahy, he works for an AI company, co-founded Eleuther AI, and he founded another AI safety research company, Conjecture. And if you look up Eleuther, it's also an open source take on open AI that's trying to make open weight models. Yes.

37:01Lon Harris:And it's basically saying super intelligence leads to, according to the web page, there is the end of humanity. This is an extinction level event is in the making. And they're paying. Is it clear if Hank's being paid? It feels like it's a commercial. It came up as a.

37:20Tudor Achim:That felt like a commercial. I think it's a commercial. He doesn't credit it. It's not tagged as like a sponsored video. But the first line of the description is, if you find these trends concerning and want to make a difference, go to control AI dot com slash sideshow. So it's either a very close collaboration or some sort of a sponsorship because they have their own portal at control AI dot com.

37:48Lon Harris:OK, but it's not in the subject and I'm not trying to get Hank in trouble here. I'm just honestly confused. It is not. It's not in the title of the video. It's not in the description of the video. I just used Ask AI. You know, there's this cool feature in YouTube when they make these long videos with hooks and they don't tell you the answer to the question they're posing in the title. You can just go to Ask AI and say, summarize this video for me. And I asked it, is he being paid to support this? Go to 1625 in the video. Let's just play that part, if you don't mind, guys. Thanks to Control AI.

38:22Lon Harris:for sponsoring this video. Control AI is a nonprofit working to keep humanity in control of AI by mobilizing everyone from concerned citizens. I just want to listen. I read ads here on the program. They play like a jingle. Hey, this is an ad. And I'm like, hey, everybody, here's an incredible piece of SaaS software that your startup could use. Hey, here's a way to get your cap table and your 409 done. It's like it's incredibly crisp and clear. It's an ad break. I'll even say like, hey, let's take a pause for the cause or let me thank. Yeah. You do that up front. It's super clear. It's in the notes.

38:56We make it very explicit. What is going on here, Ron? Here's the cut. It is. I mean, you shouldn't put that at the very, very end. I'm also now, now in light of the context, I'm wondering if he sought this out specifically. Like he got whacked last month and his fans were very angry at him for using AI to help make his show. Is this now he's overcompensating going the other way? I'm going to team up with a group that hates AI and make Doomer videos. Like that almost feels like it might be reactionary to the blowback that he got last month from admitting that he's doing research with Claude, which was a weird controversy anyway.

39:32Lon Harris:If it pleases the court, I would like to give my final piece of evidence before we do closing arguments here. Cause this shit is crazy. There's something called Future of Life Institute. Ron, you can go cue this up of who's backing this, but they're also in on this. And there's a website that I guess puts together all their ads and, you know, like organizes ads from people. Listen to this one. Our hands built America. Hands that carve, shape, teach, and care. We've always had tools that made us faster and more capable. But they never replaced the people who held them. Because a machine can't take pride in a hard day's work.

40:12Or teach a son right from wrong. Because the most important intelligence is human. AI is being designed to replace us.

40:20Lon Harris:We should have a say. This is Protect What's Real campaign from the Future of Life Institute. This has been coming up, because once you start watching a couple of these and you like them, which I recommend you do, you're going to go down the rabbit hole of what is going on in the world that rich people are spending in ungodly amounts of money to try to convince people to stop AI or that AI is just amazing at planning my trip. Ahmed, what do you think? I am so surprised. I had no idea this was even a thing. Like I had no idea this was even a thing and it feels like the world is kind of divided.

40:59I used to live in Europe for 10 years and all I, well, not for 10 years, but like I lived there for 10 years, but the last two years I lived there and AI was becoming a thing. That's all I was hearing. It's like, no, AI is going to replace us. AI is going to do this, et cetera. We should not let it. Then I moved to the US last year and all I hear is the opposite, all right? It's like, no, it's going to take over the world and we should let it and we need to embrace it and we need to make use of it, etc. And so I think this is kind of emphasizing the two different types of the world that exist.

41:26And I'm just really surprised that there's all of these nonprofits. Though I am now a little bit, you know, nonprofit opening. I wasn't a nonprofit, so I don't know what to think about what a nonprofit is anymore. But I'm surprised that these things even exist and there's like ads and there's all of these things going on because I personally did not know it was a thing.

41:43Lon Harris:Lon, tell us who's behind this. Future of life, they go back to 2014. So before the AI, AGI era, there again, they're advocating against large scale and existential risks posed by technology. Founders include MIT cosmologist Max Tegmark, UCSC cosmologist Anthony Aguirre, and Skype co-founder Jan Talin. But interestingly also in 2023, March 2023, FLI published a letter titled Pause Giant AI Experiments that called on major AI developers to agree on a verifiable six-month pause on any systems, quote, more powerful than GPT-4 and used that time to institute a framework for ensuring safety. Elon Musk signed that letter.

42:32So did Steve Wozniak. So did Gary Marcus, Evan Sharp, Connor Leahy, the same guy that was behind the other group that we just talked about. And Andrew Yang, politician Andrew Yang, all signed that 2023 paper. So that's sort of where they're coming at this from again is like it's a little doomer. It's a little we got to it's not saying no way. It's saying pause it while we figure out how to ensure that it benefits humanity and is not. Alex, you want a germinator? I can see. No, I was just going to say, I feel like, Lon, you solved a bit of the mystery earlier where I just feel like risk consulting generally is a very durable, big business.

43:09And you look at what Ian Bremmer has been doing for years with Eurasia Group, where every single Fortune 500 has, it's almost like an annuity. And basically, they publish their top risks every year. And that's just like further fodder for the whole machine. And I think there's, I mean, at least part of this, there's something similar at play where, you know, if you establish the risks and kind of create kind of frameworks and become inadvertently the trusted expert for kind of addressing them. Like, in other words, if there were no geopolitical instability, you know, Eurasia Group probably wouldn't have any contracts.

43:46Like you sort of need, it's just part of this kind of ecosystem, I guess. But that's definitely at least one little part of the mystery. There's a whole bunch of factors at play, but I think that that definitely is one of the constituent sort of, you know, backers of at least the Wuddite part of the equation.

44:05Lon Harris:This is all, I think, creating a little psychosis in the public who is looking at this and saying, well, I guess I have to pick a side. I have to be either on one side or the other. and if a certain number of people pick the side of protect human dignity and sovereignty and work and purpose, well, what would their natural reaction be to taking back their power? Like, how do they respond? What are they trying to get them to do? What are they trying to get the public to do here? Stop using the tools?

44:44Tudor Achim:Okay, I think that last phrase, you hit on the key point, right? So why does the public experience it with a psychosis? It's not because they have to pick a side, in my opinion. It's because they have two conflicting things. First of all, they all use it. I think a quarter of Americans use it every day. So on the one hand, they use AI and they get value from it. And most of them would prefer not to stop using it. You talk to anyone that does any sort of knowledge work, like, oh, do you want to go back to manually copying stuff into Excel? Like, hell no, right? So on the one hand, they like AI.

45:13Tudor Achim:On the other hand, there's this narrative around future AI that's like, hey, this is going to take everyone's jobs. You're not going to have a purpose anymore. So of course you'd hate that. I think it's actually very challenging for most people that don't understand AI development and the consequences to try to reconcile these two in their minds, especially when the media is really bombarding them with the Doomer case and they're just in a positive case on a regular basis. I want to finish up a point here, which is important, that the current AIs we have are good enough for almost all knowledge work.

45:44Tudor Achim:Leaving aside these math results, because they're like verified rewards, they're a bit easier in some sense. The meta chart, which is my definition of AGI, which is what percent of the time can you finish a task of like two hours, four hours, eight hours, a day, a week, a month. Humans are able to execute a task over a decade successfully, right? The AI models capping around like mythos was maybe like 16 hours to 20 hours. So I think that it's an interesting point that you can maybe slow down a bit of future AI work because you've already gotten a lot of knowledge work done. Most people kind of keep their jobs with the use AI to become a lot more productive with these short time horizons.

46:22Tudor Achim:And then over time, you figure out how to ramp up bigger ones. So I think there's actual solutions here that it gets lost in the shuffle of the FUD on both sides. Like AI is like already hitting super intelligence and AI is about to kill us all. I don't think that's very helpful.

46:34Lon Harris:You know what I feel like I'm going through here, Ahmed or Alain, is one of the techniques, I was really into like persuasion and misinformation and like studying those techniques for a while just because I personally found it very interesting. And I felt like I, you know, was getting a lot of people trying to, you know, watching people in politics do this and friends get into politics and then watching them suddenly deploy these same techniques. And I was like, what are these people I know doing? And one of them is foggy and one of them is flooding. And these were like misinformation techniques from Putin specifically.

47:08Lon Harris:And everyone's like, oh, my God, you know, you believe Russia is trying to who, you know, Putin's like trying to, you know, manipulate everything in the world. The answer? Yes. I mean, it's literally his core skill set is to to do this. He learned it. So fogging is releasing many competing explanations, theories, caveats, and insinuations, making some plausible, some contradictory, until people no longer know what happened or who to trust. The strategic purpose of this is to create doubt, confusion, and to get people to give up. The other one, flooding, repeating and amplifying particular messages at high volume across platforms, accounts, clips, and communities.

47:53Lon Harris:Sound familiar? And this makes the chosen narrative feel dominant, familiar, or unavoidable. In other words, you nod your head, right? And so I'm trying to figure out, Ahmed or Alana, I'll let you both sort of take a stab at it. It's like, is this what's being done to us? Because these are done so well, Ahmed, that you start to think like, yeah, yeah, humans are special. And yeah, nobody can replace a dad's advice. And it is good at planning a trip. Like you find yourself nodding because they are true statements. That's what the human mind does. We evaluate each statement, each piece of text. And then on the other side, I'm kind of like, I'm exhausted by this conversation.

48:34Lon Harris:I think I just want to stop talking about it, especially on podcasts, because here we are, Lon. How many times have we had this conversation in the last six months? Every Tuesday. No less than 25 times. So I think I'm about to give up on talking about this. Ahmed, your thoughts? Yeah. Also, Jason, to your point, you know, you talk about flooding. But like if you think about social media today and how algorithms are just like taking what you're watching and putting more content for you, it kind of feels like flooding is the norm nowadays, right? Like you watch some videos about something and then all the content is about that.

49:03So flooding kind of became the norm and we're all being flooded by whatever our friends are watching. And I don't know what's happening under the hood of these algorithms, but that became the norm of what people are consuming every single day. So I think flooding is just like the way that everybody is consuming content nowadays. just by the nature of how the algorithms work and how they monetize.

49:23Tudor Achim:Yeah, I think that the important bit is that we're monetizing people's fear, right? So people react with fear and they click on stuff. And that's what you see on YouTube and Facebook and TikTok. And you're giving people the most scary narrative ever, which is AGI takes your job and then it causes you to be extinct, you know? Like everyone's going to keep clicking on that. It's like a doom loop. It's hard to get out of. I don't even think it's like next to the company's fault. It's just that's what people click on. Yeah, I mean, I think you can really see. On social media, especially like you can see when these narratives start getting repeated so much and then there's pushback.

49:56It always follows the same sort of up and down pattern. Like the one that I've been reading a lot lately is people will come out and say, you know, which and I think this is a reasonable thing to say. People say, like, I think at this point, Hank Green just said it. AI is probably inevitable. Like you can resist it. You could say, I don't want to use it, but we're all going to be using these tools. It's not going away. And you heard that narrative so much like Reese Witherspoon went viral for saying it. And then there's the inevitable pushback every time that narrative comes out where people like, well, not me.

50:25And I'm going to resist. And they want you to think that so that you don't feel like you have a choice. And so I think we've seen that exact conversation loop happen like five or ten times on social media at this point. It just keeps this sort of machine just keeps restarting itself. And I do think it in some ways is designed to just burn you out. Like, I can't have this argument again, whatever. I'm going to keep using, you know, notebook LM and Claude and just not worry about what everyone else does. And it does make you sort of just disengage, which is what Putin was trying to do all along as well.

50:56Just make people stop paying attention and then I can do whatever I want.

50:59Lon Harris:Yeah. People are going to need to get a little bit media savvy and not immediately react to the emotion. And then what you have to do to defend yourself against this is say, who is saying this? What is their motivation? And then what is their evidence? And I am not going to, and then what is the copy actually say? If I were to read this as opposed to watching it, and what conclusion are they trying to get me to come to? And then stop yourself from being impacted by the emotion. I mean, just look at the copy. What is their goal? Okay. So this is an organization that wants to de-sell, and they're telling me humans are human, uniquely human.

51:43Lon Harris:Okay. What is the evidence here? Right? And then you can say, well, what's the ground truth I know in my life? When I'm using these tools, what am I getting out of it? Is it replacing my relationship with my kids? It's not. It's not replacing my relationship with my kids. Is it getting rid of my job? No. It's making more opportunities emerge inside my company. So you have to take each claim and then say, well, what's the ground truth that I am seeing in my life? And who gave this claim and why? It really takes incredible thoughtfulness to pre-bunk these. And that is the term I think people are starting to use is pre-bunk these.

52:24Lon Harris:Go do your own research. Look around you at ground truth, at the world around you. What are you seeing? Lon, what do you see when we apply these tools at our jobs? Are people working less or more? Yeah, no, my evidence has always been, yeah, using it every day as a knowledge worker, making podcasts, doing newsletters. It makes us faster. It makes us far more productive. And it means that we could do a lot more things. Like when it was just me, we could do the YouTube channel. We could do like maybe an X account. Maybe I could keep up one other platform. Now that we have all these AI tools, we're on Substack.

52:58We could make LinkedIn posts. We could do original stuff for Instagram and TikTok. You can do more. And you're working more. You were up always. But you, yeah, you.

53:06Lon Harris:No, but it actually makes you more motivated to work is what I'm finding. So the whole concept, I don't know. Tutor, what do you think about getting through this fog? And then what matters right now in terms of AI? Like, which is another way of saying, like, where's your P doom out? Are you actually worried?

53:25Tudor Achim:And does this. I'm not even a tiny bit worried about P doom. Yeah. That's what I'm saying. I think that part of this confusion is we're talking about math here. It's a verified task. You can throw infinite compute at it. I urge everyone to just look at the meta evals on time horizon. My guess, although they haven't managed these models, they might be at 24 hours, maybe 32. You're so far from years of concerted effort that this dialogue doesn't really make sense. I just worry that, like most publics, the American public is not used to the kind of media literacy and sophistication you're describing here.

54:01Tudor Achim:And so given that everyone responds to fear and there's so many reasons to publish fearful articles like these, I think it's going to be tough to dig ourselves out of it. And you're not incentivized as a lab to cut back because, you know, if open air cuts back, then Anthropic just puts out more loving grace and those kinds of things. Better to be feared than loved, I guess. According to the Financial Times, banking giant UBS is adding a new box every graduate and intern has to check in its 2027 hiring class that proves they can actually use AI. The bank is folding AI questions straight into interviews where candidates have to describe a real task that they asked an AI model to do, what came back and why that result was better than what the applicant could have done by hand.

54:48It's a pretty notable reversal in tone here. The last two years of banking industry AI coverage have mostly been focused around head cut counts or shorter internships. Morgan Stanley predicting AI is going to eliminate over 200 ,000 European banking jobs over the next five years. But this is part of a recognizable trend. Santander is reportedly requiring advanced AI skills to become a trainee there, too. The Financial Times raves it as banks wanting AI literate juniors, even as they trim how many juniors they're going to overall hire. So, Alex, we'll go to you. If UBS is right that this is now a baseline, what does that say about how fast this has to move inside big, slow-moving enterprises?

55:30Well, I think the more kind of prompting hubs you have, you know, the more efficacious AI use is going to be within the org. And I think this is, to some extent, this feels like a more, this is more revealing from their org than they might have realized. I mean, I think there's clearly a deep insecurity about the fact that they might not be kind of optimizing the use internally. And so it's sort of just a desperate heuristic to get new talent in the door and make sure that they're actually up to speed on tools. But yeah, I think it's going to just become table stakes for many hires. I think it's an interesting trend for sure.

56:11And the fact that they're finally starting to just acknowledge that you're going to be really far behind if you don't have this capability natively, I think is pretty interesting.

56:21Tudor Achim:I think that's an interesting question, though. Like, it's interesting that this technology needs explanations. So if you think of like smartphones and like email, I don't recall companies ever saying like, you have to demonstrate how you use email to like be more effective or like, you have to prove you can use a smartphone because everyone just immediately adopts those technologies for everything. Because they're so easy to use. They're so helpful. I think it's very telling that like, this has to be called up explicitly as like a requirement. Because you'd think, you know, if it's so great, in all cases, like, they would just come in with a requirement.

56:51that knowledge might be a way for them to get more training data too on kind of novel uses and things kind of crowdsourcing i mean it's it's it's interesting the ai labs yeah but i don't know why

57:01Tudor Achim:like it's interesting that santander has to say like you guys have to be trained on ai whereas they probably weren't like you have to know email client you know it's like right yeah you have to know how to use slack uh jason this actually made me think of we hired the associates in training in part based on their ai skills uh you want to talk a little bit about how you tested their metal And it wasn't just a question. Like we actually had them design things with AI.

57:25Lon Harris:I think this is like a super obvious thing to do. But when large corporations adopt the super obvious things that small corporations that are, you know, generally a little sharper, a little more nimble because they're small, it's like a little tiny speedboat like our venture firm with 20 people. We can really make quick turns. And the big aircraft carrier of, you know, a large bank can't. we told people who want to become associates in training we have a three-year training program to become an associate in venture capital right out of school you get paid a pittance you get a bonus if you beat everybody else uh every month it's like a crazy uh insane competition inside of our organization that i made uh in order to have people quit and then have winners stay so it's like we hire five of them we hope one of them becomes an associate and the other four move on, quit, whatever, opt out, tap out, get fired.

58:21Lon Harris:Probably at like one in three, make it. The point is, we told them coming in this next cohort, here are three different tasks you can do. Pick one, you get paid$500 to do this. You can spend 10 hours on it. In other words, you're getting paid 50 bucks an hour because I don't like people doing free work. And I like to see if you ask somebody to do a project, if they can complete the project. It turns out when you give people this offer, half of them don't do it. So now you've eliminated the half who actually want a job, but they're unwilling to do a paid assignment. Like, what does that say about the person?

59:00Lon Harris:They actually don't want the job. They're just like their parents told them to apply for a job or something. But then we told them like, hey, build a model to build a venture capital firm. We asked them to do a deal memo, but to do it with an AI and build a piece of software with AI or an agent, a skill, a dashboard, whatever you want to use, lovable, clawed, to analyze and write deal memos. So now you have to know what a deal memo is. You don't have to use AI. And then they give it to us. And it was like, whoa, some of these kids were like, knocked it out of the park. And the other ones who couldn't knock it out of the park just didn't apply.

59:37Lon Harris:They didn't take the deal. So then we don't have to get somebody in here and in week five realize, oh, they don't actually want this job. They don't want the job. They're not meant to be here. And if you got through four years of college and you didn't cheat using chat GPT for the past four years and you didn't at least explore that option while everybody was en masse cheating on everything, okay, that's a red flag too. Somebody said interesting. I didn't know who said it. Who said interesting on the panel when you heard my crazy maniacal hiring. I think it was me. I think it was me. What do you think of this story and then just my insane technique?

1:00:17Well, I mean, so on one hand, obviously, I'm not surprised. Companies hire people to create value. If with AI, you create more value, therefore, you will require people to just use AI. Okay? That makes sense. But then there's another part that I'm more worried about, and this is my worry on a human level of what AI is doing to people. And this is what's touching a little bit to what you were saying earlier, Jason. Earlier in the show, you were saying, oh, people should be watching content and asking themselves the questions about like, but what do I think? What does the data say, right? Instead of just consuming dumbly.

1:00:47You're talking about critical thinking, right? And so when companies like UBS come and say, oh, now you can use AI. I remember preparing for interviews or preparing for exams. It developed my critical thinking. I had to be like, okay, I don't have a choice. I need to figure it out. You sit down, you're blocked three hours. And it makes you someone who can be like, okay, this is what's written, but this is what I think and creates the whole critical thinking skill, right? When UBS does something like this and more corporations, we're going to be doing something like this. As a 21-year-old, if I was 21 and I see this, I'll be like, do I need to prepare for this?

1:01:19I'll just use AI, right? I can just use chat GPT. And this is what's happening. People are getting these homeworks at school and I'm talking to some university students sometimes. They're like, I don't need to study. I don't need to prepare. I just throw it to chat GPT. Over time, we're going to end up with a civilization that unfortunately does not have critical thinking, right? And this reinforces the paradigm where you consume content. You don't have the critical thinking required to ask yourself the questions. So you consume and you just take it. And we go back to what you were saying earlier.

1:01:46So I understand. And as a company CEO trying to create value, I understand why people do that. On a human level, we are just reinforcing a generation without critical thinking. And I think this is going to have bad consequences over the long term, which is something that our generation here, I mean, some are from different generations, but we all have developed critical thinking and we are not necessarily prepared for what the generation without critical thinking is going to look like.

1:02:15Lon Harris:There was an interesting study I looked at the other week, Alex. I mentioned it on All In. Turns out people who used ChatGPT to form their thoughts and write a paper, people who did it like without using a tool and then the people in the middle who did search and then wrote the paper so like three different modalities they looked at their ekgs or ec what they call the things we studied electrical activity in the brain ekg i don't think it's an ekg it's for the heart that's for the heart yeah there is the brain thing see i think anyway EEG. EEG. EEG, thank you. Electroaceptilograph. They put them on EEGs and it turned out like, oh my God, the people who were using ChatGPT, they don't even remember the points they made.

1:03:10Yeah. They had no activity in their brain.

1:03:14Lon Harris:The same way those of us door dashing calories to our house, as opposed to walking to a grocery store or a restaurant or like actually doing the act of cooking, get fat and our muscles atrophy. This is what's happening to people's brains. Right. And so that's the I mean, yeah, I feel like we talk about this all the time. It's like when you see those stories about the lawyers got busted because they had hallucinated cases in their in their submissions to the court. It's like there is a way to use these tools like Google or like any other where you're you're interacting with them and you're double checking and you're rewriting things that totally works.

1:03:50It is great. But then there's the temptation always exists to just copy and paste clawed results into whatever you're doing. That's the thing that if you just don't do that, all of these bad eventualities could basically be avoided, I think. All right. So I think we have consensus.

1:04:04Lon Harris:Pretty good idea to hire people based on their AI skills. Like at this point in time. Yeah. Of course. Congratulations on catching up. What do you got next on the docket there? I think the open source story is really interesting. Yeah, we should talk about this. So the New York Times ran this piece last week. It's called Corporate America is Getting Hooked on Open Source AI. Companies like AT &T are increasingly using cheap, freely available artificial intelligence models over expensive ones from anthropic and open AI. This, of course, has been something we've been talking about for a long time.

1:04:37I think the debate usually comes down to do companies want to use the absolute best intelligence available that they can know that they can trust? or are they willing to take a chance on maybe pretty good, good enough intelligence to save a lot of money? And it seems like we're finally -

1:04:54Lon Harris:And this part of the story, let me take over here. Please. I'll show this one note here. I thought this was the one that the public picked up on it. So again, when the New York Times is on a story, that means the story is like in the eighth or ninth inning. Just like when a big company does something. We all did this last year, but okay, Welcome to the party, AT &T. As AI cost skyrocketed, Andy Marcus, the company's chief data and AI officer, looked for cheaper alternatives. He landed on open models, which can be downloaded and modified without payment or approval. Okay, there's a news flash. By May, open models accounted for 20 % of AT &T's AI use.

1:05:32Lon Harris:That has since risen to 40%. It may jump to 60 % in the coming months. Mr. Marcus said in an interview, we believe it can go much higher. He said, adding that AT &T was saving up to 80 % on AI costs compared with earlier this year. And they explained how also here, NVIDIA was buying Hugging Face and Open Router got purchased by Stripe as other proof points. Open source, taking over the world, Tutor or Frontier models are going to have to react to this in some way. What does this mean for...

1:06:09Tudor Achim:I think it's a serious threat. And I think that the discourse could be better on it. So the fundamental thing to understand is that the only thing you get for more weights is generalization. That's the only thing you get. So if you don't need to solve Navier Stokes and program a Boeing jet and find a recipe before lunch, for sure you don't need a giant model. There's another corollary to this. So for a given model size, eventually open source saturates it because everybody just keeps training and keep training. And there's an upper bound to what a given model class can fit in. My sense from our evals and from a lot of people's evals, like perhaps this AT &T person, is that it turns out, you know, 400 billion parameters, 900 billion parameters, whatever, turns out to be sufficient for all your tasks because you're not trying to do everything at once.

1:07:03Tudor Achim:And so I think that the challenge is that these frontier companies, they keep pushing insanely general capabilities, which are very, very impressive, but also they stop being economically relevant for most companies. So for most companies, for most of their economically relevant work, I think that either now or within months or quarters, open source models will be able to perfectly handle all of it. There's one caveat, which is that if you're putting them in the loop for safety, critical, agentic work, you don't know if China's backdoored your base model or something. You really don't know. So there is a risk there.

1:07:37but in terms of capability i think like you know you can keep pushing the frontier of math all you

1:07:43Tudor Achim:want but eventually there's a limit at which most human economically relevant work is and once you cross that you just don't need anything more and then of course you don't have to pay like the literally like 10x markup that you know you're charged by a frontier company which is unbelievable yeah how are you looking at this ahmed uh in your work yeah i mean in our case we are extremely like open source driven we do so on our platform we have open source and closed source and one thing that i can say is that our developers always ask for open source video models they love having open source and typically open source you know when people when people um are evaluating models they evaluate on three things cost accuracy and speed these are the three things that people typically will evaluate models on you will find and this is like you know i agree with tudor from an accuracy standpoint or performance standpoint whatever you want to call it, you don't need all of those billions of parameters.

1:08:32From a cost and speed standpoint, because you can take the open source models and host them in the way you want, you can change them the way you want, typically people will prefer using open source. Now, the thing that we see, especially in VideoGen, I don't know much about if it's the case in LLMs, I guess so, China is being extremely open source and open ways. The US, on the other hand, let's say Anthropic and OpenAI, are more on the closed source. And so we are seeing developers really going after the open source, especially on bugging face, et cetera, because it satisfies their accuracy metrics, and it is way cheaper and faster.

1:09:09And so it is a serious threat, as Tudor says, and I can see why.

1:09:13Lon Harris:What about you, Alex? What are your thoughts here? Obviously, you existed before the large language models came to pass, but now, obviously, you're including them. Do you think Clue needs to make their own forked open source model or the harness? How do you think about it from a strategic business sense or in a strategic business sense? Yeah, we've, I mean, it becomes a cost optimization problem in a certain frontier. And I agree with everything that that was said. And we, you know, 12 years ago when we were kind of really starting with brute force classification, We had New York Times film critics on full-time payroll, like helping to design ontologies and classify entities.

1:09:55There's been a huge shift where we've gone from having to spend close to eight figures on just classifying hundreds of millions of entities and creating our ontology to being able to do that kind of distributed validation. And we're increasingly using very small models to do specialized tasks. And open source has been a big part of that. And, you know, the headline story is really cost optimization as well as kind of domiciling and making sure we're protecting our IP. And I think it's a realization that more and more companies are coming to. And, you know, we've talked about it before, but it feels like intelligence is kind of becoming like electricity and will just be, you know, table stakes and distributed.

1:10:40And it really comes down to, you know, the kind of appliances you could build on top of that. So, yeah, we're very much in favor of it, certainly at Clue.

1:10:50Lon Harris:The really interesting thing to me is just how much cheaper these are. You're talking about a 12x difference? You know, when I look at OpenRouter?

1:10:58Tudor Achim:It gets up to like 40x, 80x in terms of like... 100x even. It's unbelievable. Yeah, yeah. It's huge. It's insane.

1:11:05Lon Harris:Yeah, and that's actually, I think, what the public's going to soon learn. We keep getting these alerts from Claude. We started with OpenAI. I got some concerns about the company. We tried Claude. Then I had some concerns about the cost of Claude because we put this Claude tag into Slack. And it's like, yeah, we're not going to use the tokens in your monthly accounts for your users. It's going to be like a usage basis. And all of a sudden,$2 ,000 in usage bill. It's like, okay. So now if people use Claude through Slack, as opposed to using the downloadable app, we're paying differently. and I felt like they were constantly trying to token Maxis.

1:11:46Well, they automatically, unless you tell it not to, CloudTag automatically uses Fable for everything you ask it to do. Even if you're asking it to do something simple, it's like, let me try to solve this math equation for you. Like, no, no, I don't need Fable for this.

1:12:01Lon Harris:Yeah, and so what I'm looking for is a way to use something inside of Slack that is much cheaper, basically, but has all those kind of features. And I think right now, GrokBot's pretty good, but you have to kind of use the bot app. It's not really working too well inside of Slack yet. You can run, I've made it work a few times. You can get GrokBot to post things in Slack for sure. The best one I've seen to date is Perplexity Computer because they build this incredible harness that feels just like Claude, co-work. But then you pick your model. and they have like Nemo Tron at one 13th or 14th the price.

1:12:42Lon Harris:But in the next version, you're going to be able to use a local model on your Mac Silicon. Right. And they'll route it to the local model. If it can't get the answer in this time or with the dexterity that is required, it then falls back to an open source model. Then it will go to a frontier model and they do that through some sort of orchestrator, which is what I think DoorDash, Uber, Stripe and other companies have been talking about, they've been coming out publicly saying, we figured out a way to route the tokens or route token usage in an intelligent fashion. Ramp was the first one. I think they, I forget what it's called, but they have their internal product that routes everything to different models automatically.

1:13:23Lon Harris:But Claude and OpenAI will never do that. They're never going to offer you the open source ones unless they're forced to. Right. So that's going to be the interesting moment is what if, you know, AT &T goes to Claude and says, hey, we'll do this with you. We'll use you as our default, but we want your open source models and these other open source models in your harness to get used first. Then we'll fall back to Fable. But if you lose your harness status in the organization, then, and I think this is why Open Router got bought by Stripe and why Hugging Face got bought by NVIDIA. I think this is the product that's the claw killer.

1:14:04Lon Harris:This is NVIDIA's Claude killer. Not that you don't use Claude, but that you become the harness that people use first, whether it's for co-work type behavior, bop behavior, agentic behavior, or coding. You use the NVIDIA stack. They roll in a rack for you. They give you their open source. They give you a harness. And I don't think NVIDIA's got like a proper Claude downloadable harness for Nemotron yet, but they will for Laguna, which they just bought as part of Poolside and Hugging Face. Like, what if they make a Hugging Face, you know, a harness for people, for enterprises? That would be very powerful.

1:14:42Lon Harris:Here is one minute of the new social network type film. This is called Artificial. It's about the Sam Altman maybe getting fired, open AI situation from a few years ago.

1:14:59Lon Harris:tutor your take on this uh spider-man andrew garfield playing sam hultman

1:15:06Tudor Achim:relate to the whole show's conversation i think i did a great job with that trailer i mean very creepy yeah yeah andrew garfield really walks exactly like some others he studied the right right i've never met sam side but yeah you see the there's a famous

1:15:23Lon Harris:video of him like walking at like the g20 summit or some bullshit you know ted conference or whatever big davos and he's wearing those iconic like running shoes and he's got the huge giant backpack that's like 50 pounds heavier than he is but he's walking like he's i don't know skipping or like somewhere between uh he's race walking right like your auntie speed walking maybe yeah speed walk yeah alex what what what uh what your initial reactions your trailer i mean you know he's he's a trailer reaction i mean he's he's uh he's slightly more handsome which threw me off potentially you know there is uh i don't know what all the guns and what is that stuck out to me yeah it's uh i mean i know you know i remember i was talking to someone who's very close to sam and met with him early on and ended up in that, you know, and apparently he was just like a nice, he was scrolling dating apps the whole time while chatting with the person.

1:16:22I mean, this intense sort of feral, you know, ammunition. I don't know. It's an interesting, I'd be curious to see what the character portrayal is there. But yeah, it's a good trailer. I'm at your closing thoughts here. Yeah, I mean, I agree with Alex Tudor. I think there are the guns. I was like, wait, what's happening. I'm curious to see how much they're going to dramatize the story. Because I did hear from some people that ChatGPT was just like a fun project and they just released and they didn't expect anything. But then when you look at this trailer, you look like, oh, we're going to do this thing that's going to change the world.

1:16:55It's going to call it ChatGPT. So there is some dramatization, which of course is needed to make a successful movie. I think the social network had a bit of that. So I'm curious on how much they're going to portray reality versus making it a story. I'm really curious about that.

1:17:09Lon Harris:You know, with the gun thing, I've known Sam for a long time. We're not like great friends or anything, but we've been at parties together and known each other since the early days, 20 years ago. Incredibly happy for all his success, yada, yada. But he did talk a lot about the bunker stuff. Like I think there was like a famous New Yorker profile where he talked about having like in Big Sur, he's been, I would never say the words Big Sur, I wouldn't want to dox him, but he's been very public about this, that he's got all this land he can fly to, that he's got a bunker with like guns gold antibiotics batteries uh you know israeli defense force gas masks like he's ready to go if it goes and i think this is the peter thiel influence here and then it also infected zuckerberg all of this group during the early 2000s uh because peter thiel was talking about like there could be social collapse this is long before the AI stuff, but just on a monetary societal socialism versus capitalism.

1:18:12Lon Harris:They all kind of got this in their head that you needed to have a passport. I think it was New Zealand originally for Peter Thiel. But yeah, I think when you have too much time on your hands and you've got an incredible amount of wealth and you're wondering, what can I buy? You know, like, what can I, can I buy something that will make me feel better in some way? A bunker is like pretty high on that list because it like reduces your fear, but it also makes you seem crazy. I would wager that's not totally made up. I bet if they put that in the movie, there's some reference point. No, no. He describes it in the New Yorker article.

1:18:45Yeah. There's like, they have a thing they can point to that's like, no, no, Sam Altman has guns in a bunker. Here's where we found that piece of information. I doubt they would have just made that up completely. The bunker stuff has been widely reported on. But they are definitely using it here to build a sense of like creepiness and danger. Like, what is this weird man in his like half resort, half office, half bunker? What's he up to? I don't think that's his music. I mean, but like, that's what I mean. It's like that setting doesn't totally make sense. It's like there's people working, there's a library, but it has like a home vibe.

1:19:22And then he's in this bunker. And I think it's supposed to create with you a sense of like, what is going on here? Where are we? What like disorientation? Oh, and these are the people who are playing it? Yeah, this is the side-by-side of the casting. Okay, here we go, real quick. Monica Barbaro from Top Gun. She was just in that sexual purge movie, One Night Only. She's playing Mira Morati. Okay. This was the guy I was saying from Anora. You're a Borisov, his sootskiver. Yep. Looks pretty good. That's Andrew Garfield, we know. Ike Barinholtz, who we don't see in this trailer from Mindy Project.

1:20:00He's going to be Elon. and then Mark Rylance, the great British actor. He's playing Jeffrey Hinton, the AI pioneer. Ike Barinholtz is playing... Elon Musk.

1:20:10Lon Harris:I don't even know who Ike Barinholtz is. If you pull up a picture of him or Jacob does, you will recognize him. He's a Saturday Night Live guy. He was in Mindy Project. He's been in a few other... Yeah. He's like a character actor. Right. Character actor, comedian. So we haven't gotten a shot of him. We don't know what he looks like in the movie, but there you go. He's playing him. Okay, that's pretty close to Elon. I think he can pull it off It's all about the voice I feel like You gotta nail the voice to get Elon Alright everybody we'll see you next time Bye bye