Altman’s End Game | Eric Schmidt’s 1% Stake | Mistral Founder Joins | Onion Futures

Whoa! Let's do this. Big day. Big day. What day is it today? Big day. What a time to be live. Everybody is hyped. Everyone's pumped. Everyone is hyped today. What's the reason? Ronin, tell them why I hear you cry. Why is everyone hyped?

Only because we've got the CEO of Mistral, the co-creator of AlphaGo, now founder of Metis Reasoning. Let's go. We've got the CEO of Napster. Uh, we've got Ian coming back in. Ian the goat. Ian over there at Ledger, chief goat officer. I've just given him that title. Um, and we've got, uh, we've got John over at Scaffold Ventures coming in as well. It's a great lineup. It's a great show. Let's pull up the Olex agenda. Let's go. Let's get this thing up. Let's, uh, let's see what is happening.

So, big show. We have Thor coming in, founder at Metis Reasoning. Uh, he has released or come out of stealth essentially this morning, I believe, uh, at 1 minute past 100 a.m., uh, to release Metis Reasoning. Uh, Thor was a co-creator of AlphaGo, previously at DeepMind, and he's come out with an absolute star-studded cast to launch Metis Reasoning. Their bet is simply that LLMs, uh, aren't the way forward. They've just released a paper called The Age of Reasoning. It argues that today's LM chatbots confidently give wrong or misleading answers to queries because they do not reason or check their answers, and the authors have set out a different architectural approach that solves AI hallucinations and, uh, reliability problems, which is, uh, which is great to hear. We've got Arthur Mensch coming in, co-founder and CEO at Mistral.

Oh Lord, oh Lord, he coming. Um, he's going to be coming in and breaking down his build or surrender article, giving us an update of where Mistral at today. What does it mean to be sovereign in 2026? Are we in the singularity? Does he believe? Does he subscribe to those views? And, uh, what is European ambition? Why do we, why do we, uh, why do we need to be bolder, be braver, think bigger? Uh, we've got Ian Rogers coming in, chief human agency officer at Ledger. Uh, I think third time on the show, second time on the show, pretty sure maybe third time on the show. Um, uh, always a good chat to catch up with, with, with Ian. That's going to be fun. Ledger have just launched Ledger Atom and it's a phone-shaped security device essentially. So going to be digging into what is the, what is the release? What does that look like? Then we've got John Casten coming in, partner, co-founder at Skyfall Ventures. A lot to talk about. He's going to be calling in from Oslo, um, from Oslo Innovation Week more specifically, but the conversation is going to be around why, where Norway can win AI, energy, logistics, robotics, maritime, physical AI, defense, bunch of different stuff to get into with him. And then John Akunto, CEO at Napster. And yes, I know what you're thinking. It is that Napster. Um, they're doing AI avatars, going to be talking about enterprise, uh, AI, enterprise ROI, where is the gap, how are AI avatars going to play a part in the future and what does that gap look like today, um.

That's right. And, uh, it's a huge moment here at ETN, and maybe just a brief sort of, um, second, you know, we started this show, um, about 11 and a half months ago properly, but, you know, we really, we, we, we took the dive to go five days a week only about three months ago, and, uh, you're incredibly thankful for our ecosystem, uh, everyone that supports us, you know, to sit down with, with CEOs of, of Mistral, the co-creator of AlphaGo. It's a, it's a personal dream come true. It's a dream for everyone here at ETN. So, we're incredibly excited to, to get into that and get into the conversations that we've got lined up today. It's going to be a good show.

It's also my dream to work with companies like Airwick. Leave our business banking behind with Airwick. The only one financial platform built for the pace of modern business, multicurrency accounts, corporate cards, expense management, online payments, all-in-one platform. Where do you go? You get it. Wallex.com to get all that good stuff. Um, we love them over there. Let's, uh, talk about a little surprise that we have. We kind of hinted on it a little bit. Big reveal. Another big guest, but, um, Sam, do you want to go to the intern cam, please? Intern. Here he is. Who's that handsome guy? Look at this guy. What about the other one? Let's get the other angle. Can we do that? Get, get him on the mic. Where's it coming? Where's it coming from? Who knows? Who knows? Oh, look at this. Look at all those guest announcements on the screen as well. Let's go. Let him cook. Um, Archie, please talk to the mic. Tell us about yourself. What's up, everyone? How's it going? Let's go. He's here. He's here. Quick round of applause. Let's go. He's a natural. He's a natural. How we feeling? Welcome to ETN. Thank you. Thank you. Very, very excited. It's a big, uh, big second day in the studio for me. No, it is. So, yeah. Cannot wait to, uh, to get started. Amazing. Amazing. Well, thank you for joining us. Thank you for being here with us on such a, a monumentous day, a monumentous occasion. Uh, there will be a pull-up test. This is, um, this is the, the shareholder value that he's already created. Yeah. Like Arthur, Arthur said, "Hold on, Archie's in the studio. No coincidence. Let's get in and do an interview now." No coincidence. There is zero coincidence there. Uh, no word of a lie. Obviously lying. Uh, Arthur Mensch is coming on because we hired Archie. Uh, it just makes sense. You know, he started yesterday. Arthur's coming on today. Look, I, I'm not a data guy, but the data shows that. Yeah, there's a trend though. Archie brings the, Archie brings the swag. Uh, he does. He does. He brings the hutba. He does. Um, you love to see it. So much to discuss today. Yeah, there's so much to discuss. Um.

Altman, uh, was in The Information, uh, this morning. Altman describes three end games, uh, for AI apparently. One including, um, one where a single firm melts everyone else's GPUs. Um, that one doesn't sound that... What does that mean? We don't know. We're going to go put it up 'cause it sounds like an error. So he, he did this in an all hands. He did his famous all hands speech and the last, the last time we got an inside look at a company all hands was the, uh, at OpenAI from Sam Altman was the code red. Uh, yeah. Oh no, no, no, no. Am I wrong? Error. Um, code red. Then we've also had the, uh, when what happened, it was, it was about the, the IPOs and basically saying, you know, when Anthropic released the fact they were going to IPO, it was like, hey, we don't need to look at our competitor, we're kind of building on our own race. Do you remember that one? I don't. That was in an all hands, all hands as well. Okay. I'm completely wrong. I'm live on air. Um.

That's okay. But that's the, that's the, yeah, the chronology of, um, the all hands from OpenAI historically and like using the code red as an example, I think we, we deliberated back and forth on that for quite a while where we discoursed, as they say. We discoursed, uh, you know, surely there's a better way to go about, uh, you know, maybe steadying the ship or just sort of, you know, maybe a quick walk around the desks. Hey guys. Um, a quick one from me. Focus is just we're going to cut some of the side stuff we're doing and let's just focus on the models for now instead of a better way to do it. Code red. Call the press. Uh, let everyone know there's a, there's a, there's a code red. Even the name code red, like danger. They didn't create the name. They, they... No, I think he, I think he issued a code red. A, yeah. Yeah. But like as in like a code red is quit everything you're doing and focus on the most important thing. That is a thing within the tech industry. Um, but other companies have done that before. Google has done that before. There has been code reds before. So it's like, but why is it a code red? Why should it be like code? Because their main might get to AGI before them. Yeah. I just think code red... That's not a go to the desk and say, "Hey, by the way guys, we're going to have to..." I think, I think all is fair, but code, the term code red I feel has negative connotations where it's like danger, bad, warning. Yeah, I think it was a, it was a warning signal. It was like, hey, uh, we are struggling. We're going to have to sunset a bunch of product projects that we've been working on and a bunch of products as well. Um, we're going to go and focus only on the main mission. Code red. Go. So, um, I think it's warranted. I think it's... that apparently very quickly, uh, Sora might be coming back. Yeah. And I'm going to, we're going to talk about that as well.

Let's talk about the three different scenarios. So, OpenAI CEO Sam Altman last week told employees about three proposals the industry has discussed for how to safely develop AI, stark framing that comes just as the Trump administration set up its first system for policing, um, superintelligence developers like OpenAI. So one path involved developing an international organization similar to the International Atomic, uh, Energy Agency that would enforce safety standards, a scenario Altman appeared to support. Um, another option... What do you think about that option? Um, I mean... I like it personally. I think I, yeah. So lower down in the article I have more context on like, um, uh, on how they're going about it. He doesn't really, uh, like, doesn't actually, um, like that I don't think from what the article then ensues later on despite the fact him saying he appeared to support it. Um, I think it probably is needed but then again, uh, adding another, just another regulatory body just... Yeah, like... Well, we don't really have any to be honest. Is that going to help? What's the problem? Like people say like is, uh, selling a sandwich is more regulated than releasing a frontier AI model in the US right now. Is that what people say? Yeah. Yeah. Yeah. Apparently. I mean I need to look into that. But... No, I feel like the food industry in the US is pretty unregulated to be honest. It's probably more dangerous to be fair eating a sandwich in the US.

Um, another option would involve merging all of the AI labs into a single body to develop new AI safely. Um, that I think is just not, uh, the right way to go about it and I think wouldn't even happen anyway. Altman described a third scenario which he didn't prefer in which one AI developer would obtain enough economic power that it could deprive all other developers, um, including those that could be dangerous, of computing resources, ensuring that they couldn't catch up. Uh, these people said obviously I think that one's terrible, man said, adding that he was including it for completeness. Um, according to a person with, uh, knowledge of the comment, man has said internally that there are too few viable proposals for how to introduce a so-called speed limit on developing, um, on the development of the technology. He also said, uh, he wanted humans to maintain control of the technology and to avoid scenarios in which power in the industry is too concentrated.

This, that's his recurring line by the way. That's his recurring line. Always bring the humans back into the... Uh, concentrated power seems to be the one thing that everyone's using right now as their like go-to tick. Um, which, you know, is, is a great one. I'm obviously in favor of that. I don't want one person to own everything. Yeah. Because that's probably bad. Um, but that seems to be the recurring, yeah, line that keep, when, when we're going in through, whether it's mainly Altman to be fair who's bringing that out, like the guy, you know, when they unlocked the new character podcaster guy, I forget his name now. We interviewed Sam. We went over the interview. What was that with? New character unlocked. Yeah. Unlocked. Unlocked a new interviewer. Um, he was saying the same thing then. He was like one of his main things was like I'm incredibly worried about, um, centralization of power, centralization of like AI in one firm. Yeah. One firm ruling them all. You know, very Lord of the Ringses. Now there's been a lot of, uh, there's been a lot of discourse over the past few weeks at least, uh, whether another regulatory body, another kind of, uh, you know, I guess institution is even needed. Um, the, uh, The Information kind of went a little bit into, into whether that is actually needed in like an actual group. Is that even like the solution? It seems like, you know, for Altman at the top of the article it's like, hey, like he's really supporting this. Um, but, uh, but yeah it's an interesting one. Uh, they go on to say Altman's comments highlight the dilemma AI forces, AI labs face in coordinating with rivals. For now OpenAI, Google and Anthropic are working on new US AI safety group, the Standards Authority for Frontier AI that, uh, they hope to launch by the end of the year or in early 2027. The group would support independent evaluators embedded within the AI developer who would, uh, who would, uh, ensure, uh, safe practices and determine how developers should report safely, um, and safety and security incidents. However, the companies, uh, have disagreed on a number of points including whether the new organization or an outside body should conduct safety and capability testing of new models. Anthropic has also begun to push its own safety agenda to US policymakers. The company recently pitched lawmakers on a draft bill that, uh, that, um, which proposed allowing the commerce department to stop an AI lab from training or using a model if the AI could cause more than a billion dollars in damages or the death or serious injury of more than 50 people. Uh, that's pretty scary. The proposed bill, hold on, would also require AI labs to hire independent, uh, independent safety evaluators who would assess the models at least twice a year. Just final point, some AI leaders have argued that new regulations or testing-focused agencies would only further entrench the incumbent developers, namely, uh, Anthropic, uh, and OpenAI. Um, executives such as Meta CEO Mark Zuckerberg and Nvidia CEO Huang said such regulations would impose, uh, like overly, uh, burdensome, uh, rules on up-and-coming developers, a concept known as regulatory capture. Um, and so like that's the main argument within all of these groups, um, at the moment, adding another body just doesn't really get there. They don't like it. Yeah. Um, the, yes. So and then Anthropic has kind of gone and gone off and gone to go and do their, you know, their own thing.

So that point on, I think it's interesting for two reasons. One, uh, it's a billion in damages or the death of a serious injury of more than 50 people. So they're kind of putting a monetary value on the death of 50 people which you could equate to a billion if they're saying it's like, oh, you know, like both of these things would constitute for, um, you know, further action or further investigation or, you know, looking into it further. Um, I want to know how, so like a billion dollars in damages. Yeah, cool whatever. Um, but the death or serious injury of more than 50 people, like how do you gauge like... How do you directly... If one of those was Trump for example, then like it would be one. Yeah. You know, but like why 50 people? Why is that the benchmark? How are those, if is it like, you know, how do you associate who out of those 50 is more valuable, uh, in, in theory? You know. Yeah. That's why I think it's a bit like... You have to assume, well you have to assume that I think they're putting everyone on like a life is a life. So could be Trump, could be someone else, could be someone on the general population. But if it was one and it was Trump, they, they wouldn't be like, oh 49 more to go. Like don't worry about that one. Uh, like phew, that was a close one, like didn't you know, or if it was 49 and they wouldn't. So I do feel that's a bit of a weird, it's quite weird. I think... They're not going to be counting. They're not going to be counting the number. If they've put a number, they have to, right? Like they've said, they specifically said serious injury of more than 50 people. So like I find it weird how they're committing to... I think, I think the fact that they're even committing to a, a unit. Um, especially like not even just one life. I think it's like, hey, like yeah we can get to a 50 person threshold before things get serious and we, you know... That's pretty mad. I think that's... I just wouldn't have included a number. I would have just said like death or serious injury of, of people. Yes. Yes. I wouldn't give it a specific number 'cause... 25 are absolutely fine. Yeah. Exactly right. It makes no sense in my mind when I, when I read that and I don't think, I mean this did go viral, that, that specific passage, uh, did the rounds on X. Yeah, we had Andrew Curran, uh, in... Mr. Curran was, was, was getting stuck in with it. Um, but I just can't imagine how they got to that number or how they worked that out or... Honestly I don't think they, I don't think they have even thought about it. I don't think so. It was hidden quite, uh, deep in the report as well. I remember Andrew saying it was kind of like, huh, just found this. Yeah. Both people. Yeah. Um, that was weird.

They also explained like, uh, on a, a... CERN and other leaders at OpenAI have been publicly pushing, publicly for a body modeled on the International Atomic Energy Agency for years. I think I remember Musk as well has asked for some, uh, or called for something similar to this in, in years gone by now. Obviously a little bit less, but they say in May 2023, Altman along with OpenAI co-founder Greg Brockman, um, published a blog post arguing that the industry would likely need something like an AEA for superintelligence efforts that would be able to inspect systems, require audits, and track compute energy usage. Um, the body would require international cooperation from all countries that are developing state-of-the-art AI models, including China. Uh, last month, Chinese President Xi Jinping acknowledged the dangers of AI, saying that China and the US have the capability and responsibility to develop and manage AI for good and ensure that the development of AI is always under human control and serves the well-being of the people. And it's another example of how much we're like comparing... The, the race, right? Race, nuclear arms race, like they're looking at it and saying, "Okay, that was potentially world-ending technology." Um, it basically resulted in a number of individual world leaders coming around and saying, "Yeah, we're cool with this. Let's all shake hands and let's all agree on stuff." Um, but for some reason, I just don't feel it's... It has to be centralized somewhere, right? Yeah. Are they going to have access to, uh, you know, are China going to give access to, uh, you know, the leading test, like testing facility or regulator that's based in the US? No, probably not. Um, I think that's the same thing that happened with the nuclear thing. Everyone... And I think look, like that's the argument from, from, uh, Zach and, and, and Jensen, right? Where it's like all you're going to do by adding another regulatory body regardless of where it is in the world, regardless of who the stakeholders are because there have to be, like it has to be a centralized system. Um, because there has, you know, you can say yeah, like they're, they're third-party evaluators, uh, but realistically like, you know, that there's going to be some form of, um, I don't want to say espionage, but there's going to be, you know, the stakes are so high that you have to kind of play, uh, you know, play the game. Um, and so they can't necessarily be third-party, uh, who knows where they're from, right? Um, and, and what the agenda is for them. And so, uh, you know, I back what Jensen and, and Zach are saying where it's like, look, like adding another regulatory body isn't going to help. It's just going to entrench the problem deeper. Um, where you're going to have to go and evaluate the evaluators and you have to go and figure out, you know, it just adds another, you know, problem to the problem. Yeah. Yeah. I think if you all act independently and just create some form of, um, you know, I guess create, uh, some further negative consequences on certain, uh, on certain behaviors within models. And you know, if a lab releases said model and these certain thresholds are hit within that, whether that's like, you know, the death of said people or, uh, you know, whatever creates a billion dollars in, in, in, uh, in damages, that's then, you know, uh, you're then liable to some form of, uh, you know, some, some form of repercussion. Yeah. Yeah. Okay. I think, I think that's like a better way to go about it. Uh, obviously, you know, Altman and, and, and Anthropic are, are kind of pushing for some form of regulatory, um, you know, uh, body or whatever it is. And, uh, if they think that's great, then great. But then again, you open up the, the conversation of regulatory capture. Uh, they've obviously said that they are trying to, to minimize the, the potential for regulatory capture. Um, yeah, but I think it's just that Jensen and Zach are, are clearly seeing straight through this, but then again, it's an, it's an agenda thing, right? Um, you know, the two top lab leaders, um, obviously want more regulation. The two, uh, well, Zuck, who isn't necessarily a lab leader, um, he, he isn't, you know, owning the two, one of the two best, um, is obviously going to not want, uh, a regulatory body. So, we have to take it with a pinch of salt, right? Um, because he had a, he, he did an interview a couple of weeks back, I think two weeks ago, um, and they touch on a lot of this stuff. Yeah. And like his opinion was, I think he quoted Charlie Munger in the interview, um, basically saying like, um, uh, show me the incentives, I'll show you the results basically, which I thought was really interesting. Exactly.

Um, they go on to the second path. Um, uh, Altman has described that second path as, and I love the name, uh, the Demis plan. Uh, a reference to a prior proposal from Demis about a decade ago. He proposed creating an organization similar to CERN, um, part, the particle physics lab, uh, man's version of the Habis proposal. Major AI firms would effectively merge into or produce work through a single entity that would provide new models to countries that work with it. Um, cannot see this happening just purely like, I just can't see them, let alone like, you know, actually cooperating and working together and even coming to a sit-down, but I just can't imagine Musk, Altman, Dario, uh, you know, in the same room working together. It just wouldn't work. No, their egos, their personalities, like just too big of a clash. Like I just can't never see that actually doing anything fruitful. Then you have the, the, you have the, um, uh, the level of the state as well, uh, impacted within that. So, you know, you have the potential of the, the, the Chinese, you know, wanting to get into a conversation like that or wanting to make sure they're not in that conversation. So, like, you, you just can't win. We got to be there. We're, we're, we're frontier right now. So, we got some good AI tools internally.

Um, you also talk about the Eliza plan. This is the third option, the most scary option, and it's a reference to Eliezer Yudkowsky, a well-known AI safety theorist and advocate with ties to key members of Anthropic. Uh, Yudkowsky has long argued that advanced AI systems pose an existential risk to humanity. In a blog post in 2022, um, that whoever develops superintelligence first will need to use it to perform a pivotal act, an intervention that would prevent other people and organizations from developing dangerous superintelligence. One intervention could involve destroying the AI chips that other developers use. Altman said this moment could occur when an AI developer attains, uh, RSI, uh, that's able to improve itself. Um, uh, and he could get the resulting economic might, uh, uh, the resulting economic might to melt everyone else's GPUs. Um, which obviously is not a path that, uh, anyone wants to go down. And they kind of reiterate Altman's message, characterized this path as an attempt to centralize and control advanced AI under a single company, uh, a future that Altman has explicitly said he did not want. Um, so, uh, yeah, and given Yudkowsky's connections to Anthropic, the comment was viewed as a subtle dig at Anthropic. So, one of the people with knowledge of it, I do like how whenever these guys are in interviews, they will never use the names of other labs. It would always be like some others or like other labs or like other leaders think that they will never, um, because as soon as they do it, it's like headline news, Sam Altman, Dario is, uh, it's just, it's just kind of funny. Um, and I watched the, oh, we never talk, we never spoke about this. I watched the AI doc, uh, the other day, was on Netflix. It was all right. It was pretty good. So the guy... You said to me you didn't use, thought it was a bit pointless. Right. It was, it was, it was one I kind of feel like it's already outdated. Two, I thought it was more of a... Will the general, what will the general population think of it, do you think? One, I think it's done very badly in the box office anyway or whatever, like the, the, the release, I think it's done pretty badly. Yeah. Okay. Um, I mean look, if you're in the general population, like why are you going to care? Yeah. I think their, their reaction would be negative towards AI from it. You're going to have a different, you're going to have a different lens to them. Yeah. Yeah. Yeah. Yeah. And I think the whole reason isn't to like cater to you, it's to cater to a... To the, to the, yeah, to the masses. I think, um, the director of it does a good... So he, he enters it as a, um, as a bit of a doomer basically. And then towards the end he feels that like he has been pulled out of the doomer. He's very positive about it. Sure. Excited about it, but he is, he is still scared and does a good work of like, um, like chronicling his journey through, um, because it, it takes him like multiple years to film it basically. Um, and at the start he's like, has a kid on the way and that's a core premise. He's worried about bringing a kid into the world and then by the end of it he's actually quite excited when he sits down with some, with some like accelerationists and, uh, they kind of educate him on, on, on the benefits of AI and all the good it can bring. Um, so they do a good job of getting a nice balance of like it's not just like a doomer film. Yeah. They get like a lot of like good people who like Bezos in it, which is hilarious. I had sat on the couch. I was watching my mate and I was like, that's Bezos. Who the hell is Bezos? I was like, Bezos, Jesus. He, he asked me, do I mean Jeff Bezos? And I said no. Um, as you would, as you would, as you would. Yeah. It was, it was, it was a fair question. It's a fair question. Uh, he does really well in it. He like really, he does, he does incredibly well. Um, but they have... Shout out to Bezos. They have Altman. They have Dario. Nice. Uh, Zuck declined to comment and then Musk said he would and then didn't turn up. Uh, basically, which is quite a joke. Yeah. Yeah. Yeah. I thought it was quite funny. Okay. Well, I'm going to have to watch it. Uh, it's super interesting. It's like worth a watch. It's not like... I wouldn't like... Well, I texted you. You said, "I'm watching the new AI doc." And I was like, "Oh, how is it?" And I think you just said... I was, yeah, it was okay. It was likely all right. I, I was pretty... I think I then just said, "Yeah, fair. All right." And then we never spoke about it again. Um, anyway, let me tell you about AdQuick quickly. The intelligence platform for out-of-home advertising. Plan smarter, buy faster, measure everything. One platform connecting you to 1700 media partners across 30 countries. Go to adquick.com. Um, let's move on to... Can we just, we don't have this on the thing, but I real, I should have put this in today, but yeah, they're bringing Sora back. Oh yes. Okay. Plans to revive Sora. So on the code red, Sora was the first to go when they released this, like why the hell is this a thing? And I realized today, I kind of feel like we passed the AI image/video Turing test and no one really cared. I feel like, I feel like we're there, like I now can't, um, like decanter whether or not a video is purely AI or not. Uh, yeah. Um, yeah. Uh, it's, um, it's, uh, and the reference I'll give, like, so did you see in the riots there was the Pink Panther rioter? It was dressed up as Pink Panther. AI mate. No way. Yeah. It's come out in the BBC. A lot of, I'm not sure if like there, I think there was one BBC. I think there's one still that was legit, but all the other ones are full AI. And there's some ones that I saw that I was like, I fully believe this. Um, which is, which is crazy. And that's when I kind of realized, I was like, "Oh, I have been fooled. I am the fool." Uh, which I thought was mad 'cause they did look all, you know, pretty real. Um, but check. Yeah, the BBC have a thing on it. Um, but it's in like the, there was, there was someone, um... Yeah, there was someone dressed up as a Pink Panther in one of the images, but basically all the ones that have come out are fake. Okay. Yeah, fine. I can see the video. I can understand the video being fake, but the, the photos look real. They're fake. They are. Wow. I know. They are just, just on that one. So, the reason they look so real is because a lot of them are taken from actual pictures of protesters or in some cases PSG fans from like a year ago or something like that. And then the only thing that's been changed is the person being grabbed by the police, for example, into a Pink Panther. Those PSG fans, they're up to no good again. Their usual no good tricks and they're deep fakes. I see. I see.

Um, but let's talk about Sora. Um, so it's, it's, there's plans to revive it. Um, uh, they say remember OpenAI's video model Sora and all the great Sam Altman memes that came from it earlier this year. Catchb Maker canned both the model and the TikTok live app. Uh, it powered as part of a broader culling of server-intensive side projects. We saw they were like, "Hey, we're not doing these anymore. We're going full speed ahead on the models." They had fallen behind a little bit and then kind of regained their lead. Um, but there is a chance for all those Sora fans out there who think I need more AI videos in my life. There is a chance that version of Sora is coming back from the dead. Um, they say a stealthy startup developing AI video for filmmakers, which we revealed last month, is in talks with OpenAI to access Sora, tweak it, and sell the software or other services using it. According to a person who's familiar with the talks, Hollywood mogul Jeffrey Katzenberg, former Sora research lead Bill Peebles and former Dropbox CFO Sujit Jasua co-founded the startup in recent months. Um, so they're looking to basically acquire everything, remodel it, and run it. So, not an OpenAI thing. Okay. Uh, not clear whether the startup will get access to Sora, awaits the settings that determine the model and how it responds to queries. But if the deal were to go through, the startup would get deep enough access to the model to heavily customize it. Um, interesting. Okay. So, so this isn't going to be under the OpenAI umbrella tool. No, no, no. But OpenAI will receive equity in the new startup if the deal goes ahead. Sure. Um, and, uh... How early are the talks? Um, not looking for a time scale. I'm just saying, did they say early, in early talks or it is like... No mention, but it sounds like there's a team formed around the new company, uh, and it's been co-founded in recent months. I mean look, I, if, unless, where are they going to get the compute from? I don't know. Probably from OpenAI, wasn't that the main reason why they sunset, obviously focus as well, but... Yeah, they were like compute... The team, the team focus, but then again, yeah, it's the compute and I think that was, that was a DeepMind thing as well where it's like, look, like, uh, all of the frontier research which is building on, you know, science and everything else. That was like their push and then obviously, uh, you know, the, the Google team were like, hey, like we also need, we need the compute to go and generate cash basically, to go and generate some, some form of revenue from this, uh, technology and, uh, they were, each one was compute unless the other one stopped, um, and so... And so that, that was, you know, obviously very high level, that's kind of what happened. I assume this is going to be the same thing. Know, um, they must have figured out how to, how to go and find more compute. Yeah. Like, well, if it's a separate entity, I don't think they have to like worry about, like they won't be focused too much on, um... But where are they going to get, they're going to go and, uh, get it like from somewhere else. Yeah. Yeah. I think they'll, you know, they'll just go and probably raise money and spend it all on getting compute and rolling it out. Um, fine. But, uh, yeah, for just a quick one, thought it was cool if, uh, for all you Super Sora fans out there, don't worry, you can still watch AI slop. Um, that's good. I know you guys have been gagging for it, so it's coming back. I personally, uh, don't care. Super pumped about, uh, you know, getting more AI slop on my timeline in my brain. I don't think I'm slopped out enough right now. I could do with some more slop. And, uh, yeah, boy golly, we're getting it, which is, which is a great, great sign. Exactly what we need right now. Indeed. Indeed. So, I'm super happy. Uh, let me tell you about Chargebee. Chargebee is the billing and monetization infrastructure built for the AI era. Scale prepaid credits, complex enterprise contracts, and all other modern commercial contracts all in one system. Go to chargebee.com. Um, let's, uh, let's talk about Eric Schmidt.

Um, and here's 1% of Anthropic. Um, this is, this obviously did the rounds on the timeline. Um, it was first kind of picked up by Bloomberg, I believe. Um, it goes, the former Google chief executive, uh, officer obviously Eric Schmidt owns around 1% of Anthropic thanks to his early investment in the startup, according to people familiar with the matter, setting him up for one of the largest individual windfalls from the company's imminent public market debut. Now, um, this sparked some huge interest in, uh, the way that Anthropic went through their funding journey in terms of not, uh, it was like basically all super angels until, until later, later down the line. I'm going to say Series A. Uh, there's a bunch of, bunch of tweets around this. Um, and Eric Schmidt is one of them. Um, Schmidt invested tens of millions of dollars in Anthropic's Series A funding round, according to people from, uh, according to people who asked not to be identified because they weren't authorized to discuss private matters. It's unclear if he has sold any part of his stake since the initial investment. Um, Anthropic's highly anticipated IPO is set to generate massive windfalls for its employees and investors. The startup's co-founders are already worth $8 billion each based on its last funding round. Um, when Schmidt, who's 71, made his first bet on Anthropic in 2021, it was valued at $461 million. Uh, um, investors indicated last week that a fair valuation for the company is in the 1.8 to 2 trillion ballpark. Um, the end of that range would value Anthropic at more than $4,300 times its Series A funding round. Pretty good. Pretty good markup. At the value of the Anthropic stake would boost Schmidt's, uh, net worth by 1/3 to nearly $80 billion, according to Bloomberg's wealth index. Even using Anthropic's most recent valuation, um, of, uh, as we talked about yesterday, $965 billion, Schmidt adds nearly $10 billion to his existing 57.5 billion. Um, so the dude's set to make an absolute, uh, yeah, an absolute wedge of cheese, which is always good. Um, I want to pull up some tweets about, uh, about what's happened here and, and the kind of story behind it. So Natasha on the timeline, um, kind of broke this, uh, and it obviously did the rounds. Um, Natasha said one of the more interesting things about Anthropic's cap table, um, is until its Series C, its funding rounds were dominated by super angels and nonprofits. Um, so here's the scoop, uh, with, uh, uh, with, with Tom, um, obviously announced it and then, um, Trung Fan then came on the timeline and said Eric Schmidt owns 1% of Anthropic. Dario's wife, who Schmidt used to date, uh, introduced him to Anthropic. Schmidt's peak ownership of Google was 5%, which is now 1%. He got that Google CEO job after going to Burning Man with Larry Page and... That's an all-time right there. That deserve quit. It's not the golf course. How cool is that? Right. Um, then someone came in and said, uh, well over on Wall Street said you're one Burning Man away from changing your life. So where do you need to go? Go to Burning Man and you could potentially go...

and start Google and get into a company like Anthropic. Early days, pretty good returns. I don't know how much that cost, but maybe it's worth... someone could do the maths. Maybe I can't do public math, but maybe someone can break that down. I do... so her saying that... sorry, I forget her name... saying that, yeah, until Series C funding rounds dominated by super brain nonprofits. I'm surprised it's that long, like I would assume pre-seed... but think about it, they're like a three, four year old company.

Yeah, yeah, which is wild. Yeah, yeah, which is wild. But I think it's the same for a bunch of... well, DeepMind in particular, and also OpenAI was dominated just by super angels and like huge... not nonprofits, but like philanthropists. I think it was like Reid Hoffman put like a bill into, I think, DeepMind really early on, which at the time was like one third of his net worth.

Yeah, yeah, which is like, you know, for any like investment that's insane, right? Imagine putting one third of your net worth into a single investment. It's pretty risk on, but obviously, you know, great, great bet from the fellow, but like that's just a wild, a wild amount of cash. So I'm not like... yeah, I'm surprised it went all the way to Series C, but I guess, you know, success of OpenAI, they're like, people are already looking for this type of thing. Okay, they've maybe proven out that you can monetize this. Let's, let's get into the next one as fast as possible.

Yeah. Um, got intro, you know, by his ex, went... yeah, met the ex for a coffee. [sighs] Um, went to like, hey, I'm actually... my husband now is starting this company, to, you know, is this interesting? Oh, yeah, actually, maybe, yeah. Yeah, would love to do that. Like, maybe that's a great investment strat. I should call up all my exes, [laughter] see what their current husbands are, see what they're doing, and maybe they're starting a new AI lab. [laughter] Look, that's... maybe there's alpha there. It's slightly contrarian, but that's where you make the money. Maybe I should do that on a massive email or something.

I can definitely back that, not that many people will be doing that strategy. Uh, I don't... yeah, well, apart from all the LPs, all the LPs be like, tell us about your thesis. What's the thesis? So, I have a few ex-girlfriends and I know their boyfriends and their husbands very well. I want to, I want to, I want to reach out to them. I want to start investing. Um, that's it. I think it'll be good. Um, uh, we can go on to the next, uh, segment. I know we have our first guest coming in slightly early.

Uh, let's go check out what the timeline's doing. Let's just see what everyone's doing. This one is a personal favorite of mine. It's doing the rounds at the moment. Um, uh, the University of Chicago, some of the students are protesting at the CME demanding an end to the federal ban on onion futures. Uh, this is a student-led group. For Our Futures is rallying to call for a repeal of the 1958 Onion Futures Act. Good news. We're fighting for the trading spell. Which side are you on? Let's go.

Doesn't look quite as crazy as Paris. No, but a good cause here. Are you for the onions or will the ban live on? Great work over there. So, there was... kind of, kind of reminds me of... [laughter] so I can figure... like I know that... Oh, hold on. Oh, sorry, I've switched. I'm just getting up the future... the ban. Um, the, uh, the ban. So in 1955, traders Vincent Kosuga and Sam Gold accumulated about 98% of Chicago's onion supply and they basically, uh, like created a monopoly of onions, uh, in Chicago. And then after building the monopoly, they basically shorted the market and then just dumped their whole inventory all at once, completely crashing the price, and a 50-pound bag of onions went from $2.75 down to 10 cents. So the bag that the onions were held in were worth more than the onions at the time. Insane.

Um, and, uh, and yeah, and then obviously like, um, this was not cool. This was just genuine market manipulation. So they rolled out a ban on trading onion futures, but it's never been revisited. Like, it's never... for some reason they just haven't repealed it. So, um, obviously great cause. These guys are on the front lines trying to repeal a ban on onion futures. Um, and, uh, and I don't think there's any reason why people shouldn't listen to them, to be quite honest. I think... Yeah, I think the, uh, that costume kind of says it all. I think they should. Well, they... yeah, they care. They care, which is good to see. You can tell by the costumes.

Um, do you know... do you know the story of Chocolate Finger? No. Anthony Ward? No. Anthony Ward, he was nicknamed Chocolate Finger or Chocfinger. Um, he's a British commodities trader, or was the co-founder of London-based hedge fund and trading firm, um, Armajaro. They were in, um, the FT a while ago. Um, I can't remember why, um, but anyway... So basically the story goes, this is quite a while ago, but in 2010, um, July 2010, Ward made international headlines by purchasing 240,000 to 241,000 tons of cocoa beans through Armajaro for roughly 658 million pounds, which is about a billion dollars. Um, [snorts] that single purchase represented 7% of annual global cacao production, enough to make roughly 5.3 billion chocolate bars, and gave Armajaro control over the vast proportion of exchange-monitored stocks.

The price surged. The massive trade pushed cacao prices to a 33-year high. Industry groups, including the German Cacao Trade Association, accused him of artificially manipulating the market and creating a supply squeeze. Critics noted that West African cacao farmers saw no financial benefit from the price spike. Um, then he was nicknamed Chocolate Finger, um, which obviously is the play on the James Bond, uh, Goldfinger. Um, [laughter] Armajaro's core trading and asset operation shifted over the years, and the firm ultimately closed its doors around 2017, changing its industry landscape and the rise of automated trading. Um, so yeah, that's, uh, that's the tale. Um, there's a really good Medium article about the genuine story of Chocolate Finger.

What a legend. Super high level. I mean, yeah, he's been in the news previously, uh, a little bit later on in his life for not some very good things. So less of a legend. Okay, okay. [laughter] I retract my statement. That story is quite legendary, let's say. Um, and... is that... but go and read the Medium article. Super interesting. Uh, I think it... it was... I want to say it was a gray area. I mean, look, like, uh, I'm not a fund manager. I don't know this stuff, so, uh, I don't want to say whether it was illegal or not, but like, I assume it was probably a gray area. Um, it was probably like, hey, you can't go and... Well, I guess, look, if the market's small enough and you have the funds to go and buy it and you're willing to put almost 700 million pounds into one trade, I guess that's fair game if it's like within the confines. Because look, you can do things like a hostile takeover. You can go and buy a, you know, a public stock and you can go and rally... like you can go and buy more of the public stock to go and take control of the company if you want to. That's not illegal, I don't think. So if it's in the confines of the game, I'd say like fair play, you know.

Yeah, yeah. Um, we'll do, we'll do some research. We'll do some research. Although there's a moral question, but there's a really good Medium article. Go read the Medium article. Search in... I don't know, uh, Chocolate Finger. Probably coming up with that surely. No, no, because obviously, well, if you're in the UK, you're going to get chocolate fingers. Uh, that's like the name of the confectionery. So maybe search up Chocolate Finger and search up Armajaro Holdings and, uh, you'll get it up. It's very interesting right there. Incredibly interesting. Sorry, that was a bit of a tangent. No, that's a great one. [laughter]

Um, this one here, uh, uh, so Lance Oppenheim says he bribed Robert Pattinson. Pull the screen up. Um, uh, Lance Oppenheim says he bribed Robert Pattinson into starring in Prime Time by sending him 200 canisters of Zyn with a letter saying, 'It's good. Dear Chris Hansen, uh, Rob at the time was grappling with a Zyn addiction.' [laughter] So he's blackmailed him using his own personal addiction to get him to star in the film. And it looks like it was successful.

That's so European style. Good to know that the strategies from, you know, the old days of Hollywood are still alive and well in terms of getting someone to come in. Anyway, we have our first guest coming in. We have coming in the founder of Metis Reasoning. They're just coming off the back, fresh launch and a brand new research paper. Uh, very excited. An absolute honor to have you in the building. Can't wait to have a conversation. Hello, how are you, sir?

Hello. Very well. Calling from Berlin, GTC. Nice. How is it over there? Yeah, it's wonderful. Fantastic. You know, I studied in Berlin, so for me it's extra special to be back here, dude. Really a full circle moment. And what a momentous day for you as well to be back where you studied. I mean, you've got two announcements kind of essentially coming out of stealth, but also the paper has been published as well. Uh, would love to hear from you. Metis Reasoning, what is the company about? And then I'd love to ask some questions about the paper and what you guys are working on.

Yes, fantastic. Yeah, we have identified reasoning as still being the big open problem in AI. And by reasoning, we really mean reasoning according to the scientific method. You know, the scientific method on steroids, if you like. And, uh, you know, you might wonder what's the current problem. But, um, you know, really large language models have no notion of truth. They just have a notion of what plausibly comes next. You know, the next token, the next word, and so on. And we believe that you need to really, um, have a new architecture in order to do real reasoning. An architecture that keeps a record of what it knows, but also what it doesn't know, so that you can really advance, um, your reasoning, um, increase your knowledge, and also have a record of how exactly that came about and why you give the answers that you give. That's what Metis Reasoning is all about.

Okay, really cool. There are multiple parallels or multiple lines of conversation I'd love to go off the back of this. One being like Luke and I were just chatting the other day around, um, like the Millennium Prize Problems and, you know, those getting solved and OpenAI releasing these new set of proofs and they're kind of throwing a lot of resources at it. Um, something that we were struggling with to understand on the show that like, yes, that's a great achievement, um, but on like the reasoning element or like how we actually got to those, um, breakthroughs, like, you know, I'm kind of sensing maybe like that's an example of how the problem of LLMs, if we were to extrapolate it further into other areas of research and understanding, like is that, um, is that one of the problems in itself? Like we don't really... we have the proof, we have the result, but we don't really know how we got there. I can maybe, you know, think about the problems of that in the future that we might come across if we sort of... there's like a gap of knowledge that we're moving into for generations to come.

Yeah, that's exactly right. I mean, that is a problem. I mean, imagine you have a proof of an important theorem but nobody understands it. [laughter] So that would be a bit of a tragedy. Um, but the other problem, of course, is that all of those successes have been had in highly verifiable domains. You know, mathematics, coding, they're verifiable because you can see once you have a solution, you can check if it's correct. And that's super good for, you know, reinforcement learning and methods that rely on that. But if you go out into the real world, the complex world, the open world, things are just not like that, right? You have a lot of uncertainty, you have a lot of complexity, and, um, you have this kind of open world properties. And in those circumstances, there are other things you need to do. You need to make sure that you have a sound reasoning process that also takes into account uncertainty in the right way. And you need to make sure that once you've reasoned through an important problem, uh, people can really follow how the system arrived at that solution. Just imagine, um, medical applications. If there was an AI system that decided about an important surgery to be done on you, wouldn't you want to be able to go back and understand exactly how the system arrived at the decision that that procedure was necessary? Yeah.

I'd, um, I'd love to understand, uh, because at the moment, uh, LM chatbots, you know, if they are, you know, being, I guess, misleading or they've hallucinated in some way, when they give a misleading answer, they're still incredibly confident with that answer. And I'd love to understand what is the, like, um, uh, the architectural approach that you guys are changing to be able to, like, you know, I guess, solve AI hallucinations. Like, what does that even look like?

Yeah, that's a great question. So, um, what we call the, um, original sin of the, um, architecture, of the transformer architecture and LLMs, is that, uh, really the world knowledge that the model has and the procedure by which it manipulates that knowledge are all baked into the weights. So basically, if I gave you, uh, one trillion parameters of a modern LLM, there's no way for you to extract these things. You cannot even find out what's in there other than through prompting. Okay. And so the fundamental architectural decision that we've made is to separate the knowledge that the model operates on and the reasoning process. And the reasoning process now operates on that model, on that knowledge. And this separation unlocks a lot of positive things. For example, if you explicitly represent the knowledge, then you can inspect it and you can edit it and you can, uh, do access controls on it. You know, all the good things that we know are good from databases. Yeah. With large language models, they're lost, and we want to bring them back in that sense. Yeah. If you... and sorry, go on, please, go. Sorry. And then on the other side, you have the reasoning process. When you isolate that and it operates on the knowledge, you can now improve the reasoning process in isolation and really make it better. And so you get these two axes of scaling, one on the knowledge and one on the reasoning capability.

Right. And then in terms of, uh, you know, if we kind of zoom back then and look at, you know, uh, hitting AGI or whatever that might look like in the RSI, um, why is this never... like, why is this architecture not being kind of done before? And, um, what needed to happen for this to kind of be in the right place to actually happen, to kind of, you know, separate those two sides and allow them to kind of develop, you know, I guess, separately? Yeah, I mean, funny that you ask. You know, 10 years ago I was part of the, uh, the AlphaGo team. You are? Yeah. And if you remember, you know, that was the system that played the game of Go and beat the strongest player at the time in this legendary match in Seoul. And AlphaGo had these two aspects of the architecture. You know, it had neural networks that made predictions and judgments, like how good is this Go position or how good is this move, and it had the other side also, namely, um, the game tree search. You know, it had a structure that exactly stored what the system already knew, what it had yet to explore, and so on. And it was really these two systems combined that produced, for example, the legendary move 37 that surprised everyone, right? It wasn't one part of the architecture alone. And you can think of ChatGPT and that development as just the development on the neural network side, but the actual reasoning side of AlphaGo was discarded. It was left behind to a degree. And one way to understand what we're doing at Metis Reasoning is we're bringing that back. You know, we believe that this explicit reasoning on an explicit representation of what the system believes about the world is necessary to do true reasoning.

Wow. I have to ask, uh, what was it like witnessing move 37 for the first time? What was going through your head? Like, how was the response to that? Because it is, you know, a legendary sort of pivotal moment in all of the human versus AI debates. Like, what was that like as a moment? Yeah, you know, I was sitting in the audience, uh, the commentators were up front discussing the moves, and then, you know, uh, Michael Redmond, the nine-dan professional who commented on this, put up the move, and I was like, oh no, that doesn't look good, right? I wouldn't have played that. Not that I would be able to judge really, but then he said, oh no, that must be a mistake. And, you know, panic in my heart, in all of our hearts, I think. But then he put it back and, you know, kind of started nodding and said, wow, you know, this could be something really special. This first looked like a mistake to him, but then it turned out to be such a brilliant move. And yeah, you know, and we had great relief, but also I think we were witnessing the first time that a machine really made a creative contribution in the sense that it was novel and useful. Yeah, yeah. And I believe that it was only able to do that because it had the reasoning component.

If you think about it logically, if you just have an imitation learning process, you know, like our neural networks back then, but also like large language models, then they can never really break out of the hull that has been created by that knowledge. It takes reasoning to get outside of that boundary to genuinely discover new things. And so I think that's the way in which we can really boost scientific creativity as well.

Yeah, really cool. And like talking about that, like novel reasoning, it's something that Luke and I have got incredibly excited about through interviewing guests previously, where we get excited about the fact that if these, um, if frontier AI does continue to improve at the rate that it is, that we will come into a realm of completely new discoveries that are like beyond human comprehension. I guess like microcosmically, that's a good example of what happened with move 37. It can kind of think in a way that humans aren't able to do, and it brings novel discoveries and like a creative element to the way that it works. Within that though, like we're seeing a lot of, um, worry on the timeline, uh, in the news right now around some of the existential threat claims. I was wondering if you subscribe to any of the claims being made at the moment by, uh, lab leaders or general discourse around future threats of RSI or AGI. Does that sit with you? Because I'm sat here now thinking like surely we'll reach a ceiling with LLM architecture right now. So that would maybe negate the existential threat in a way.

Yeah. You know, I think the existential threats are overblown. This is very powerful technology and it opens up so many possible positive avenues for research, medical research, for example, material research, and so on. And because it is so powerful, we need to be very careful about it. That seems clear to me. But I don't subscribe to any of those doom scenarios because there's just no evidence for that. And, you know, I think strong claims require strong evidence. But what I do believe is that we need to make these systems safer. And the first step to making something safer is to make it more transparent. And so if an AI system makes a decision for me that's important, be that in science, engineering, medical or whatever, air traffic control, you know, we need to be able to go back and ask why did it make this decision? What sources of evidence did it tap into? How did it judge those sources of evidence? How did it combine those insights? And only if we can do that will these systems be safe. And that's why I think that this kind of architecture will eventually take over because it's the only safe architecture. We can't allow ourselves for these systems to be completely black box.

Yeah, yeah, yeah. I remember like, uh, I think it was when, uh, o1 was released. I remember reading an article around people were slightly worried at the fact that its reasoning, uh, like transparency became a lot more opaque essentially. Like people weren't really sure why that was even internally at OpenAI. Um, is that traditionally like a cause for concern you believe then? I guess this is why Metis Reasoning is getting so much attention right now because I guess the opaqueness equals worry, right? We're not aware of what's happening, we don't know why, it maybe makes us feel uncomfortable. Exactly. That's right. And it rightly makes us uncomfortable. So if we don't understand how the systems decide, that is a reason for concern. And of course, you know, these systems have what we call the chain of thought, you know, which are these intermediate tokens that are being generated that lead up to the final decision of what the system outputs. But it turns out that these chains of thought can also be very opaque, and that, you know, sometimes the system comes up with the answer and then only uses the trace to do a post hoc justification of what it outputted, and sometimes they downright panic. I don't know if you've seen these, but, you know, sometimes the model gets really aggravated in its chain of thought and so on. And I think we can do better. We don't want AI systems to be like humans and just being driven by some kind of inner monologue. We can make them better. They're machines in service of humanity, and we should make them follow our rules.

Yeah. I think the launch of Metis has kind of made me realize like this feels like almost something that is existential in terms of, you know, building frontier AI and getting to, um, you know, the next frontier, the next one. Um, are the lab leaders, you know, thinking about reasoning as much as they should, or do you feel like this is, uh, you know, we're still very early in the reasoning? You mentioned, you know, after what you guys did, uh, with AlphaGo, reasoning was kind of, you know, almost forgotten. Um, do you feel like, uh, do you know why it was forgotten, and do you feel like these lab leaders are thinking enough about, uh, reasoning? I'd love to understand your insight on that.

Yeah, yeah, it's interesting. So one thing to realize is that the large labs are in this scaling phase, right? It's all about scaling, more data, more compute, larger models. And they've discovered that to be a phenomenal way of making the models stronger. Um, but the problem is that that cannot fundamentally change some aspects of these models. For example, by scaling more, these models will never get more understandable or more auditable in what they do. In fact, it's the opposite. They become less auditable. They produce less of a clear reasoning trace. And while they may be strong on the performance side of what they actually output, uh, it becomes ever harder to understand what they do. But for the large labs, it's very difficult to get off that train, if you see what I mean. You know, they're all on that scaling train. And I think that opens up opportunities, um, you know, for initially smaller players in the ecosystem, um, to do things that are necessary.

Yeah, really cool. I know we're just coming to the end. I'd love to ask one question just around what is it that, uh, you know, when you're going to bed at night, one thing that keeps coming into your brain about something that you're incredibly excited about for the future and, uh, the development of this technology. Like, what are we expecting five years from now? I know that's hard to compress timelines that far out, but, um, what is it that maybe, you know, the public should know about how great this technology can be?

Yeah. No, I, um, you know, at heart I'm a scientist, and I'm really excited about the prospect of, uh, AI helping us advance science. And, uh, when I speak about the reasoning process, it's not just reasoning within the data that you have. Reasoning is this interactive process where the model reasons about what it has, but also suggests new experiments, takes new measurements, and so on, and then processes that. Mhm. And, um, and it's that interactive reasoning that I think will lead to scientific breakthroughs in medicine, in material science, and, you know, in the basic sciences as well. And I think that is super exciting. And then if we can even understand how the system arrived at those insights so that we can really trust them, I think that would be the perfect moment in a few years' time to arrive.

Yeah, amazing. Amazing. Phil, thank you so much for joining us. Congratulations on, uh, Metis Reasoning, the new paper, everything that you're doing. It's, uh, it's incredible for the ecosystem. We're huge fans. Thanks for your time, appreciate it. Enjoy Berlin as well. Thank you. Bye. See you later, sir. Thank you. Uh, wow, awesome dude. Um, uh, really great conversation. Maybe, um, that was, uh, probably one of the best. Yeah, that was so good. [laughter] We got to ask about move 37 as well, which, uh, which is a great off as well, which sick.

Um, let me tell you about Valyrian. Valyrian lets organizations deploy AI. You control your data, your agents, your keys. Stop renting your intelligence. Go and own it. Go to valyrian.com. Um, let's, uh, let's start talking a little bit about what's next on the docket.

Well, we have, uh, next guest is obviously Arthur Mensch, co-founder and CEO at Mistral. Uh, in terms of the news of the day, we have, uh, the bugging situation with OpenAI. There's been developments. Um, the individual was chatting with Saudi royals, um, during this process basically. Um, so this is the former OpenAI worker who was accused of bugging, has been discussing jobs with Saudi royals. I did see his package was leaked online. It was $375 million and $2.1 mil in stock options as well. Um, $375 million... sorry, $375,000. Uh, he was not making $375 million. Uh, $375,000 and 2.1 mil in stock options was his package. Um, but obviously this is getting sort of traversed right now. People are uncovering different bits and bobs. But a judge ruled, uh, that Alzaben should be held without bail after a Saudi consulate had offered to pay for his release. So they are trying to help him. Um, before he was arrested, the Saudi national accused of bugging a restricted OpenAI conference room had allegedly been in touch with a member of the Saudi royal family about job opportunities in the kingdom, federal prosecutors have said. And I guess this obviously causes concern because they're like, okay, now is there a reason in which, uh, the bugging went down? Like, is this a, uh, you know, they're kind of finding like, okay, we have a reason, an agenda to do that, and this is the resulting impact. Yeah. Um,

The new allegation came at a hearing where a federal magistrate judge ruled that Alzaben, a 35-year-old former OpenAI employee, should be held without bail pending trial before he is a flight risk. Uh, there is a real threat here. US magistrate judge Alex Gay said once he leaves, the case is over. Prosecutors repeatedly brought up a document from Saudi Arabia's Los Angeles consulate that said it would cover Alzaben's entire bail bond. Said, 'Even if I set it at 10 million,' asked Judge Gay, 'I find that problematic. Seems like a blank check.' Uh, we're going to pay whatever the court wants to make sure he gets out, essentially. So they just haven't even set a bail because they know whatever number they... as soon as they set a number, it's game over. Remember in the thing, they're probably like 1 billion. See, if they get a bail, you know, they could maybe milk this. They can make a quick buck. [laughter] Make a quick buck. Quick Billy.

Uh, William Perick, which is a lawyer representing, uh, Alzaben, argued his client is a green card holder and wanted to stay in the US. He is not a visitor. He is not a tourist. He's an immigrant who was weeks away from getting his citizenship. Um, Alzaben is accused of wrapping a thumb-sized circuit board and microphone in toilet paper and inserting the device behind a wall-mounted TV inside the conference room, which we covered in depth. Uh, I think last Friday did that break? Uh, yeah, maybe Thursday. Maybe Thursday. They were sitting on it for a little bit, but, you know, we know that they wanted to honeypot him. Um, uh, OpenAI said he was scheduled for two months of paternity leave beginning sometime in October, but, uh, and prosecutors said in a filing ahead of the hearing, um, he had booked a one-way plane ticket from San Francisco to Saudi Arabia via Istanbul for himself, his wife, and his three-month-old son, departing on October 12th. So obviously maybe there was a plan put in place to, um, uh, plant the bug. Yeah. And then bounce, basically. Um, Alzaben's wife and son left, uh, without him shortly after his arrest. Prosecutors said their one-way tickets were initially paid for by the Saudi government, but the flight was cancelled. They were then rebooked on another one-way flight to Qatar, this time in business class, by a member of Saudi military, the filing said. So damn, they went. It was all set. Oh, that would be kind of tough, wouldn't it? You'd be like, ah, he's nearly there. Just got, just right at the end of the tunnel. Um,

What's that, um, uh, what's that movie where right at the end the, um, it's, uh, I think Alex was talking about this movie the other day. It was about... and no, I think it was us three. Uh, we were talking about this movie and right at the end they like get on the plane and it's literally like at the... what it's about. It's about, it's like a fairly political movie. Something happened. It starts with like, um, I think these rioters like, uh, infiltrating the, um, uh, the embassy that they're in. Oh, Argo. Argo. Yeah, yeah, yeah. What a film. Great movie. Brilliant, brilliant. Um, and it's Shark Tale. [laughter] Argo, Argo. Not Shark Tale. Yeah, yeah. Like that. That's obviously, you know, obviously not the same, but it's very much like mentioning that they just get on the plane. They can see the... I think it's literally you can see the army trucks following the plane and then they take off. They were like Canadian actors they pretended to be or something like that. Document, something like that. Um, what, just worth, uh, bringing that up anyway. Such a good film. Um, but, uh, so yeah, so wife, child have gone, um, Alzaben behind, no bail has been set because blank check, basically. He'll just come out and he'll go. Um, they've said that, uh, they've been asked by the judge what the brother-in-law's role was in the Saudi military because there's connections there as well. Um, they believe that he was in the Navy, so a bit of sort of gray area around that. Um, Alzaben's wife has told the FBI, uh, the day she left, that her visa was expiring and that she and her son were permanently moving back to Saudi Arabia. Um, allegedly also told prosecutors that Alzaben planned to come back to the US but eventually move permanently there as well. So, you know, saying we're going to go, we have some things to sort, he's going to come with, but he'll be back. Don't worry. Pre... um, uh, oh, sorry, no, post, post arrest. Um, but, uh, in court on Monday, Alzaben's lawyers said his wife had been planning to return to the US. Yeah. Purchased a one-way ticket because her visa was expiring. It was unclear to her how long the renewal process would take. Um, lawyers for Alzaben also said that his wife would use a visitation visa to fly back. So they're very much adamant that like they were going, but we were coming back. It's hard to defend that really, I think, in court. Like the evidence is kind of... I think the evidence stacks up all against him completely. Not in a good spot. Not in a good [laughter] spot, unfortunately. Um, as he should, as he should be, he shouldn't be in a good... Oh, mention here. Okay. Alzaben had been an employee on OpenAI's product team since July 2025. His base salary when he joined OpenAI... so you can imagine that there's been a refresher since then because I think they usually happen six months. He's got a promotion. Okay. But he did, and obviously for said promotion, no longer huge for him. Um, it was on his side. Yeah. Big air bomb. Uh, 325,000 with a new higher equity grant of 2.2 million, um, in stock options. Wow. What an astonishing number. Um, and his base was then later increased to 385, and typically that comes with, I think, a further refresher on stock options in traditional, uh, large-scale SaaS firms. Um, uh, throughout 2025 and 2026, Alzaben was in frequent WhatsApp communication with an unnamed member of Saudi royal family as well. Um, uh, royal family member told Alzaben an artificial intelligence initiative was being launched in Saudi Arabia and he wanted Alzaben to join it. Um, uh, 'Saudi needs you more than the US,' Saudi royal family member said. He replied 100. Not sure if that was the emoji or just text-based 100. I assume it was the emoji. Um, that's, yeah, that's why the emoji is there. Yeah. So you don't have to type 100, uh, when you do reply 100. Um, Saudi royal repeatedly tried to recruit Alzaben, but at the time he wouldn't leave until his stock vested. Fair enough. Uh, it's unclear whether Alzaben's equity had vested before he was fired on the day of his arrest. I mean, if he's been there since what, July 2025? It's now... yeah, I mean, one year, he's probably got, um, 25% there. I'd assume like one-year cliff, um, like three-year vesting schedule potentially. So he's probably unlocked what, 25% of that. Yeah, but he's going to be a bad leaver, so he's not going to get anything. Uh, well, yeah. No, he's obviously screwed down, but, uh, [laughter] I mean, what constitutes like that? He's like, yeah, you're a good leaver. You can get some of the stock. You know what? You weren't that bad. [laughter] Uh, yeah, thanks for all your hard work, uh, and for bugging our office, which is so unfortunate. It's so tough. Like, what is he playing at here? Obviously, you're going to get done for that. It's like, out of all the companies in the world right now, that is literally one of, I guess... he was just laying a trail of breadcrumbs as well. Like, it's not a good spy. Come on. He's not a good spy at all.

Anyway, help yourself out. Um, let's, uh, let's go to the next story. Uh, let's just keep moving this on. [snorts] Google executives dismissed European fears over US dominance of AI. The Financial Times has reported, um, a top Google executive has dismissed European concerns that US companies are set to dominate AI, apparently arguing fierce competition between big tech and a cadre of fast-growing startups is driving down prices and creating new opportunities for the region's businesses. Um, this is actually quite timely, um, with coming on very soon. Quick clap. Let's clap. Um, Kent Walker, um, Google parent Alphabet's president of global affairs and chief legal officer, told the FT that, look at the space right now, it's remarkable, you couldn't imagine a more competitive environment. Walker, uh, pointed to big tech groups such as Nvidia, Meta, SpaceX, as well as US OpenAI Labs, OpenAI, Anthropic, France's Mistral. Shout out to Mistral. Plus another... god's sake. [applause] Plus another half dozen or a dozen Chinese companies, he said. And what's the effect of that? Prices are dropping at an astonishing rate. Capabilities are increasing at an astonishing rate, too. European Competition Commissioner Teresa Ribera is stepping up scrutiny of the rapidly developing AI industry. Brussels officials are concerned that the dominance of established US technology companies could allow them to gain an unfair advantage in a market that is predicted to underpin the bloc's economy. Uh, Alphabet, the third most valuable company in the world, has faced intense scrutiny in recent years as regulators have investigated its dominance of online search, vast advertising business, and its ubiquitous Android operating system. Uh, despite losing a trio of US antitrust lawsuits, the company was not forced to sell any parts of its empire. Instead, it was required to share more data with rivals and open up certain services to third parties. Um, they then go on to talk about DeepMind and, uh, how, uh, the labs unveiled an upgraded version of the Gemini model that has caught up with the market leaders Anthropic and OpenAI. Plans to integrate its AI technology into everything like Maps, Gmail, Android, YouTube has reignited concerns that the big tech giant can use scale, distribution, and financial firepower to squeeze out emerging AI rivals.

I do forget Android is a thing. Um, yeah, but I think it does run a lot of... Yeah, I think it is big. Well, how big is it? Like, how many people actually have an Android phone? Like, do we know? Yeah, but it's more the operating system around it, I think. Um, that's most important. I mean, like, think about the... Yeah, [clears throat] it's just the operating system. Whenever I see someone with an Android phone, I'm always just like... Then again, Google's phones have been incredibly... um, uh, the performance of Google's phones, um, sales have been great, I think. Really? Yeah, yeah, I want to say they have. Um, I don't know. It's just whenever I see someone who isn't using like a MacBook, I'm like... Yeah, yeah, it's negative. Well, they're either like... they're either, uh, cracked as hell or they suck. It's completely polarized. MacBooks, you can't really tell, to be fair. It's just like you don't know if they suck or... Well, they could be seriously cracked as well. Yeah, yeah, yeah, yeah. Um, but I feel like there's just... they're all... you also have like the middle sort. It's like a bell kind of distribution, but everyone's got one. But I think with, uh, non-users, yeah, it's just like bosh, bosh, like... could be cracked, could not be cracked. Um, yeah, yeah. Um, in Brussels, the company is facing scrutiny under the EU's Digital Markets Act, including, uh, over whether it is providing rival online...

search engines and certain AI chatbots with sufficient access to anonymized search engine data, which is key to developing competing AI-powered services. Now, I'm reading this like the conversation that we just had, you know, about Metis, it's just like, but think about reasoning, guys, like it feels so existential considering what Tour has just released, you know what they've launched, I feel like you can't hit AGI without that happening.

Should I email him and say can we put a check in? [laughter]

Hey, love the chat, please, can we invest? I'm begging you. Please, can we please check in?

I'll do a readoff. Yeah, I'll do a genuine reed hopping right now. $100. [laughter]

Big check. Huge check. Let me lead the round. [laughter]

I don't care about the terms. This is me. This is my qual. Yeah, [laughter] this is my qualitative.

Uh, look at his mustache. So good. Uh, let's keep going. Um, anyway, well, actually, let me tell you about 11 Labs quickly. AI that communicates at human level across every channel and modality. Deploy conversational agents across every interaction and create on-brand content at scale in 70 plus languages. Get it at 11labs.io.

I love AI. You love AI. You'd love to see it. Um, so what is happening right now? We're having a studio-wide discussion.

You just, Sam's just given me a charger and Alex is now also giving me a charger. So, I'm doubled up.

10 minutes later, though. [laughter] Twice the power. Twice the power. Do you want to take this back or no? Let's just leave it there, I think. Um, it's fine. I'm kind of like live on the show right now. Uh, [laughter]

Uh, let's have a look at the timelines. There are a few more things that I want to get into. Um, namely something that I've been discussing recently in the studio. Um, I've been bringing it up quite often because I am British. I think it's what British people do. Uh, but this is a tweet from the Weather Channel. Um, and they've approached it with all lower case. You love to see. They say, 'We are officially weeks away from the sun setting at 4:30 p.m.' Oh, go away, Claude. Sorry. Uh, we're officially weeks away from the sun setting at 4:30 p.m. Please begin mentally preparing to eat your dinner in the pitch black.' That's from the official Weather Channel. And I wanted to ask you, Luke, do you prefer a darker UK or do you prefer a brighter UK?

Uh, this morning I woke up and I thought, wow, this is the first time this year that I have, well no, first time, you know, the back end of the year that I left the office dark. I also got to the office when it was dark.

Double dark day. Double dark day. Double 12 hours time.

Um, yeah, it was pretty rough. I didn't even get to the office that early this morning and it was still dark.

Yeah, it is weird. I got in like maybe 7:45 and it was like dark and I was like this is awful.

That's so the cool thing. I used to getting up early when it was dark. There was something associated with... I do like feeling I've got that dog in me.

Yeah. You kind of feel like you're ahead. You kind of feel like everyone else is obviously asleep but I'm out here doing my thing.

But the sad thing about, and that's great in summer because you know because it's dark that kind of tells you that you're awake earlier than everyone else.

Sure. Sure. The thing that sucks about the winter when this happens, you can't tell. You could leave late. You could be late, but your body's kind of in that reward system where you're like, 'Oh, I'm outside and it's dark in the morning.' Yeah. Yeah. It's like now I've got to do a 40-minute tube journey and get stuck in the crunch. Um, but I will say that I do prefer... I like that, the crunch. Uh, stuck in the crunch. Yeah. Yeah. Um, she came up with it then. Um, I think it's pretty good as well actually. Um, I do think that I prefer, there is a time for dark London. I think... I think dark London, everyone turns into like locked-in mode. It's time to go.

Yeah. Yeah. I agree. I fully agree. It's game time. I think you can use it to your advantage because I feel that more people are, and this is very like bro of me. Um, this was a thought I had when I was more of like a broer to be fair, when I was like a 75 harder. Um, but I do think it's like a strategic advantage in a way that you can use to get ahead of others. I feel most people not going to go to the gym, not going to get up early because it's cold and dark. But if you're like, parts cold and dark, chance to get ahead, I think it can be pretty beneficial.

I think so too, you know. Yeah. Um, I think it's an under-discussed like leverage.

I think there's an unspoken rule between lots of people and say, you know, we're going to get hard.

Yeah. Yeah. Yeah. No, I agree. I agree with you. But I thought it was worth bringing up because I, you know, I've been saying recently in the studio like winter is officially here. Uh, it's... and it now genuinely is officially here.

Yeah, it's brass. I had a scarf on today. I'm wearing my Valyan barber jacket. Um, so... and it's going so well. I've been waiting months to be able to wear that thing.

I've been wearing it in summer. I've been boiling hot walking down the road in Soho. But... to be fair, every time I see Max, he's always wearing it. He was wearing it during the height of summer.

We can ask it. 35, 40 degrees in the, you know, in summer. Get the bomb jacket.

I was swimming in it. I was in the lake swinging, [laughter] swimming in my bomber jacket. It was amazing. Um, let's pull up this tweet from Ryan Sle. Um, he says IKEA manuals, but they come to life. Check this out. Um, I built the real Billy instructions in 3D so you can watch every step and happen exactly as drawn. This is so clever.

How cool is this? Any sound? No, doesn't. Let's go. Let's reload that. Pull that one back. [music] That was a fun song.

Who is this? Who sings this? [music] It's Kanye, isn't it? It's a Kanye beat. Let's go. [music] [bell]

So, you'll go through the video as obviously you're building the thing. You don't need... Oh, wow. Look at that. You can [music] navigate the manual as well. I think this... So, I assume you need to take photos of all the manual. Um, and then you can say, 'Okay, now I have all the manual. I'm on this page.' And now you can click on your laptop what the manual looks like. I assume you can build, you can probably download the manuals off the IKEA website as well.

Well, that's just solved, that's just made so many fathers' lives easier there alone.

Yes, agree. Um, let's bring Arthur onto the studio.

Arthur, CEO, co-founder of Mistral coming onto the European Technology Network. Hello sir, how are you?

It's an absolute pleasure. As you probably know, we've been looking forward to getting something set up for a while. So, it's a pleasure to have you on the show finally.

Yeah. Amazing. Um, Arthur, you're coming off the back of a blockbuster viral article. Um, 'Build or Surrender' was the title and we wanted to start the conversation just asking why was it the right time to push this out? What was the message you were trying to get across? And we have a lot more to discuss as well.

Yeah. I'd love to just quickly touch on the application layer before we go deeper onto Mistral's kind of strategy around this. Um, we've seen, you know, a bunch of value being accrued from the application layer. But I'd love to understand like now, have we kind of squeezed all the value out of the application layer now? Um, are you starting to see kind of like, you know, um, like the application layer being democratized and the value kind of going away and moving elsewhere? I'd love to understand like where are we in that kind of like, you know, value accrual kind of like, you know, position I guess.

Yeah, really cool. I mean you mentioned the infrastructure layer there. Let's talk about Mistral full stack positioning. Where are you guys at today? Would love to get like a temp check especially off the back of Lechon and the rollout of that recent model. Where are things today over at Mistral? [snorts]

Yeah, you've kind of touched on it, but I'd love to get, um, you know, I'd love to kind of ask point blank. Um, in the article that you wrote, you said dependence on others, even on their openness, is still dependence. Um, can you explain why like, you know, building open models is like only half the thing? What is Mistral's like, you know, real strategy here? Um, obviously, you know, you're moving into the infrastructure layer and doing incredibly well and you have, you know, amazing ambitious targets for that. Um, but I'd love to understand like what is your thinking? What does Europe need to do, um, to have this independence that we want?

Yeah. And you stated, you know, I think you mentioned we're already building frontier models. You obviously, you know, Lechon is an example but then you've also got 11 Labs and Black Forest Labs as well and I think you mentioned them in the article. The kind of the question I guess is like the resources and the compute and your guys' ambition. Um, you know, I think you've mentioned previously where it's like we know that we need to go and raise more capital to go and get to the 1 gigawatt of capacity goal. Um, why has Europe never kind of been that ambitious almost? Like, you know, I feel like you're one of those people that are pushing ambition to another level in Europe. There are obviously a number of other founders. Can you talk to us about the ambition that is needed to kind of like get to the frontier in the infrastructure layer that we need to be able to go and build more and more better models?

That's amazing. Okay. Um, you mentioned just quickly, you mentioned at the end of the article kind of the first thing is breaking the pessimism. Uh, you kind of just mentioned there like, uh, you know, we are starting to wake up, the equity markets is starting to wake up about how fundamental this technology is. Um, do you feel like we have broken the pessimism or do you feel like there is still pessimism to kind of break? [snorts]

Yeah. Yeah. Yeah. Um, a question on just like on macro scale of things right now. We're seeing some people on the timeline subscribe to the foothills of the singularity as Dario came out and said a few weeks, a few months back, comments around like AGI and then within that as an extension of that we're seeing concerns around existential humanitarian threats, civilization ending threats. Um, to be more specific, do you subscribe to any of that at all or is that just noise to you? [snorts]

Yeah. Yeah. Yeah. Yeah. No, I think we align with that in general. I just know you have a hard stop. Maybe one more just around the naming of Lechon. I have to ask like why, we're huge fans of the fat cat and we have to know why it was called Lechon.

Yeah. Yeah. They do. They do. Well, we love it over here at ETN.

Thank you sir. We appreciate it. We'll see you soon. All the best. Ciao. [applause]

There you go. [cheering] The team's very happy. We are very happy too. Um, that was awesome. Uh, that was a dream come true. Very cool.

A lot was it larger cats was there. Now have one of the infinity stones of the lab leaders.

One of the infinity stones to come. Who's next? Dennis, we're coming for you, buddy. [laughter] In a friendly way. Of course. In a very friendly way.

Um, let me tell you about Metaview agents for winning teams. The agentic recruiting platform built for clear, efficient hiring end to end. Source, engage, and hire better with agents that truly get your brief. Get it at metview.ai. Um, I am entertained. That was very entertaining. Um, before we have our next guest on, Ian, uh, Ian reunited at last with Ian. He'll be joining us in about four minutes, I think. Um, very, very fun. Oh, let's bring him in. Let's get him in here. More time with... Hell yeah, I get it. Let's go. [laughter and applause] Let's go. He's back. He's back. The big shoe is behind you. How are we? The room is looking amazing as always. Wasn't in view, but you know what can I do? I should have... Oh, I would love to see that. I would love to see that. Uh, the room's looking incredible as always. The big shoe is there. The crypto punk is there. How are things with you, Ian? What's going on?

Oh, you know, not much. Not much happening in France these days. Everything's pretty normal, [laughter] been boring. How about you guys? Luke was in France. Luke was... I was in Paris this weekend. Probably the worst weekend to go. Um, but no, I actually had a great time. Um, yeah. Are you based in Paris? Yeah, I'm based in Paris. Oh, no way. I didn't know that. Um, should I hit you up? Let's go. Hey, you should have come by. Yeah. Yeah. Quite a nice rooftop that you can come have a coffee on. Okay. See other things out there, you know. [laughter] It's a great viewing point. Lovely. Sounds idyllic.

Um, Ian, it's so good to see you again. Uh, let us know what Ledger just released? This is incredibly interesting.

Yeah, we had our big event last week. We have an annual keynote event. We had that last week and we had, you know, two big announcements. One, I talked about all the agentic security stuff that we're doing. Super cool seeing Arthur a second ago and he talked about the fact that, you know, controlling agents is going to be a challenge. So our point of view on that is kind of a division of labor where agents can be agents but then you have this kind of deterministic gate that agents need to ask permission to go through. I think I might have used this example last time, but it's like a 16-year-old with the car keys. You know, they're not in the bedroom with the 16-year-old, they're in the parents' pocket, and when it wants access, it asks, and if it's Monday morning and they're on the way to school, then it's a yes. And if it's Friday night and, you know, they're hammered, it's a no. So, there's a lot of gray area in between. That's the same thing with agents. You know, they're going to do a lot of work on our behalf, but they shouldn't hold the secrets. They should ask you for them. Um, you know, so my OpenClaw doesn't start unless I plug in my Ledger. Um, you know, if my Claw code wants my Jev key, it asks me for it. I say here it is. Um, you know, that's just that kind of a gate. And then the other thing that we announced is a device called Atom. It's a new Ledger device. It is an Android-based device. So it'll run all of your apps. So any Android app can be meaningfully more secure running on Ledger Atom. It basically has a mode where it becomes a Ledger. So it can be an Android device, you know, 90 plus percent of the time and then when you need, um, you know, for Android to go away so that you can have a direct connection to the secure screen with the secure element, boom, it basically like a transformer transforms into a Ledger, a secure Ledger device. You do whatever the secure thing you need to do and then you go back. And then the interesting thing about that is that goes beyond crypto. You know, you can just have secure messaging, private messaging. Um, you know, messaging which is invisible to Android and only available to two people holding two Ledger Atoms.

Sure. Yeah. Um, I think a nice question to kind of properly start off on is, um, you know, the last time we had you on the show we obviously kind of always talk about the importance of your role at Ledger. Um, and we always kind of talk about the adoption side of things. Um, I think, you know, when in 2025 people thought you guys were pretty paranoid I think about agents and AI and now that, you know, is obviously shifting. And I'd love to understand just as a starter point where we are today like from the last time you were on the show like how has that shifted? Have you seen any kind of like real shifts happen since the last time we had you on? Um, yeah.

Yeah. I mean it just keeps going, right? I mean it is true and I said this on stage last week that I stood on the same event last year. I talked about, you know, this agentic future coming and the security challenges with that and sure there wasn't much interest and then now, you know, I mean the reality is I had a conversation about Hugging Face with the guy who served me ice cream in Paris a few weeks ago, right? I just heard something similar, someone was saying, you know, I was listening to a podcast and they were like, yeah, and I had a conversation at my barber yesterday about Hugging Face. So that's sort of like, there's just no one questions if this is something we should worry about or not at the moment. Right. So that's a good thing. I think that, you know, we still have, you know, very few tools to actually deal with it. Uh, one of my colleagues is at a conference in Southern California right now and he said there are actually many hardware devices there which are kind of about human in the loop, which, you know, just even the fact that human in the loop is something that we talk about I think is a positive thing but none of those have the secure screen that the Ledger does and the differentiator there is if you're actually approving a transaction, you know, I mean if you're just proving, hey, I'm here, you know, you can do that with a very simple device but with a Ledger, the screen, you can actually approve what it is you are signing on that secure screen and that's really the difference. So I think where we are now is we know it's a challenge. I mean Arthur just mentioned it in your previous segment and now I think we're all getting up to speed on exactly how to deal with it. We've launched our agent stack which is headless versions of our wallet products. So Ledger wallet, Ledger Enterprise. Now, we have this agent intent that we announced last week, which is a way that your agent can be autonomous and put approvals into a queue for you to come back to later. And then now, and we've also got some proof of human things that we're doing. We announced a partnership with Persona last week. Um, and next we move on to the concept of policies which we already have inside of Ledger Enterprise but we want to widen that to even more agentic builders and operators.

Yeah. Yeah. There's like a word, proof of human that you mentioned there. I find it such a crazy... like I don't think these words were being mentioned maybe a year and a half to two years ago. Maybe by some early adopters they were but like, you know, I imagine if you went onto Google Trends and looked for proof of human mentions there's probably a spike around a year and a half ago, a year ago. Um, what... and I saw the other day I think it was, I can't remember if it was Teams or Google or someone, but they now have like verified human badges on the calls when they're on group calls in some of these enterprise companies because of deep fake attacks and that kind of thing. Um, what does this future look like that we're heading towards, you know, the proof of human tag, like do you feel that will be, you know, affluent in every single company? Will it not just be on devices but sometimes, I mean there's even a crazy sci-fi situation where in person you'd have to have like a proof of human badge with humanoid robots, but how do you see this playing out? Like is everyone going to have to have a proof of human badge? Like...

Well, I think this is such an interesting topic and we're put... this is really fun part, right? Because you can imagine a different internet than the one we have today. I think that's, it's, you know, and it's somewhat obvious and predictable at this point which is kind of fun, you know, now we get to just watch it happen over... well, we'll see what exactly the timeline is, but if you think about it we could have used proof of human go back to like kind of Cambridge Analytica times, right? You know what I mean? Like that, you know, and I think that it just hasn't really been an incentive for the Facebooks and the Xs to make sure that people are unique human beings. You wouldn't have even needed like a passport from them, you could have just said like prove to me you're not on this site with 10 other names, right? You know, or prove to me that you're not a bot farm. But they've never really had the interest in doing that. Um, you know, now it's like becoming mission critical because as you said, you know, you don't even know if like the person showing up to a job interview is a real person. So, you know, I think now it's just something that we need. I really think though that the most interesting thing happening is what Cloudflare has going on here where they are, you know, they're Cloudflare, they basically like they sit in front of the majority of the internet, right? And if you're a bot, they don't let you through. Um, and you know, now they've came out and they said that I think it was May of this year, the majority of internet traffic is bots. 94% of all login attempts are bots. And they thought that was going to be next year. So if the people who should know this better than anyone else had it wrong, then that, you know, is just increasing like way faster than anyone imagined. And then what they're doing if you like look closely into Cloudflare's solution to this is they're allowing anyone who uses Cloudflare to basically charge for entry. Right. So right, my point of view is there will be a VIP line on the internet. Yeah. If you're a bot, get in the queue, right? [laughter] But you know, but if you are a... you can prove I am a human. If I can prove I am a registered user of this site, you know, I am an Amazon Prime holder or whatever, you know, whatever the kind of equivalent of that is, and maybe even if you have a wallet, right? Like I am willing to pay for access. And that's not absurd. I mean, I pay for the X API. Um, you know, because I use it in a lot of my... like my business intelligence agent uses X to figure out what's going on in my sector, right? And that's worth paying for, you know. So, I think there are a lot of things where, you know, you'll be willing to pay, uh, you know, a fraction. Yeah. Sorry. Go on. No, I was just going to say like we've had that in our lives before, right? I mean, if you had cable television, you've paid ESPN whether you watch sports or not, right? So it could be that my, you know, 10 bucks of my Starlink bill goes to paying tolls, you know. Yeah. Yeah. Yeah. That's not absurd.

That's fascinating. We haven't actually had that angle on the show before.

Yeah. And it reminds me of like taking like dead internet theory as an isolated thing where I do think we would all be blown away by the percentage of accounts or bots or, you know, messaging going around that is bot dominated in terms of percentage. Um, is there a situation... I've been like trying to theorize this a little bit. I think like Nikita Bier has been put for the past eight months has been making predictions around dead internet theory and what that looks like. Where is there a case here where, and you kind of allude to it a little bit, right, there'll be like a VIP section of the internet where it's like only humans and you have to pay to get in and it's set up for humans, but is there a sense where we kind of retreat away from the old internet? And you mentioned new internet at the start of the conversation, right? Like is there an element here where we're like, okay, this is so bot dominated that it's almost not unusable, but it's not as pleasurable an experience as it was previously. Do we then retreat to the new internet and kind of leave that, you know, in my mind I can see it's like a burning pile of, [laughter] you know, internet and computers and stuff. We kind of retreat over to...

Go, you're doomed. Look, the internet is already like a burning pile of like [laughter] dirty shoe boxes full of garbage, right? It's all... the internet's been that for 15 years. But I think that I'd actually go one step a little further to go, okay, what does the internet look like from my perspective? I think that, you know, the user interface for humans is agents, right? We talk to agents. I want to get, you know, I want to get this done. Agents then use what we used to use as the internet, right? So I think in that way, like to me, that's kind of what dead internet theory looks like is yeah, a lot of that stuff does just become like either ghost town or just fodder for the Wayback Machine. Sure. And the stuff that we actually use, we're going to be asking our agent, hey man, can you take a look at this for me? And then what the internet really becomes is kind of like agent to, you know, to internet, you know, communication and that will become the best interface. I mean, I had something, you know, today, you know, Revolut's been down today. I tried to buy something, my card was denied. I spent 30 minutes on hold at Revolut and finally just asked my agent, 'Hey, is Revolut down or what?' And they gave me the full rundown and searched the internet, gave me the full rundown. I mean, I should have just done that to begin with, right? I mean, why did I even bother going to the internet to try to solve my problem? My agent's going to solve it for me, you know, while I'm...

doing something else and faster, right? So, I do think that is what the future of the internet looks like. We just have... I also think that, you know, what my lessons in life tell me is we always think that that's going to happen tomorrow. Yeah. And it always ends up taking a lot longer and there are a lot more bumps in the road to get there. So I think we should, you know, that, you know, timing is everything. That's why I think a company like Browserbase is actually incredibly interesting because they give, you know, their point of view is, um, actually the internet is not going to adapt for agents. We're going to make the internet accessible to agents. Yeah. Right. That's what, that's what they do. Now I think that over time the internet does become accessible for agents, but we get these kind of like, um, you know, these way stations in the way, right? On the way there, we get some other solutions. Um, just going on the Revolut, like, problem that you had and like the way that you kind of think about it. I think, you know, for us who kind of sit in this kind of tech bubble, like now I would think about that problem and go, hey, you know, Instacart or whatever it is, can you just like monitor a Revolut situation and if any article comes out or something news comes out that they're now back online, tell me and then I'll give them a call, um, and just allow them to monitor it. Right. But like that's been, um, that's been ingrained in me because I'm in this tech bubble. But then if you extrapolate that out into the kind of general population, that's going to take so long. And you're obviously talking about the timing. Um, then pulling it back to the core Weave stuff. I think you measur... sorry. Um, about, I think you said 94% of searches are now bots. Is that just login attempts? Login attempts. Okay. So is, do you think that's just a, like that's not the volume of, you know, people using these bots, that's just the amount of login attempts that these bots are doing for a very, very small amount of the population. What do you think happens when, you know, that's probably not even 1% of the population? Um, what do you think happens when, you know, 10% of the population or 50% or even 100% of the population are using these bots to the extent where, you know, [laughter] I don't know, like login attempts go from one a day from a human being to thousands a day, I don't know, and everyone's using them, like what does that just not like break the internet?

Sure. And they're not nobodies. They're not some guys in a garage with an idea. They already sit in front of the majority of the internet. So, they actually already have the toll booths, right? Um, and so that maybe this is where, you know, it does move faster than, um, you know, I was listening to, um, to Ben Horowitz on All-In over the weekend, from last week's show, and, you know, he was pointing out that like the problem in the dot-com bubble was that, you know, the audience just wasn't there yet. We had a lot of supply and the demand hadn't caught up. [clears throat] There was a lot of internet services and not a lot of internet users. Well, now there is actually a tremendous amount of demand for the intelligence that AI is creating, you know, so and now we're, you know, there's actually more of a supply-side problem. So I think in that way it probably will accelerate, but that also from a security perspective leaves us in this place that we are right now where, you know, if you grab my phone and it's unlocked, you know, you're not going to get into my Revolut account because Revolut's already solved that one. But if you open up Claude, you, you know, it can walk you through a bunch of secrets that I don't want it to walk you through. That to me is the gap that we have to close relatively quickly and what we're working on at Ledger AI security roadmap. Yeah. Yeah. Yeah. Really interesting. And like you guys are like pretty like frontier in terms of when we look at like security and the interactions between agents and that's, you know, shown in the fact that you guys are releasing products before anyone else and kind of like before it gets incredibly bad, right? Um, like how do you guys philosophize about this internally? Like how do you come to make decisions about security? Do you have any future predictions around security specifically internally that you're like taking some bets on?

Yeah. Do you, um, if we zoom out onto like the macro sense and this is probably the last question, we've only got a minute left. Um, do you feel like people are talking about this enough? Uh, do you feel like, um, you know, this is a big enough topic yet? Um, you know, obviously Ledger, as Ron mentioned, you guys are, you know, at the frontier of this conversation. Um, and, um, you know, we look up to you for that. Um, but I feel like this conversation isn't happening, uh, in a practical sense. Like a lot of people are having it in a, you know, theoretical sense of like agents can go and do this, that, and the other, but it's like practically how do we think about that in relation to human beings? You guys are having that conversation and answering those questions, right? So, like, yeah, I'd love [clears throat] to know that.

Well, hopefully. Yeah, hopefully we're pushing that message for you.