Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails

All-In with Chamath, Jason, Sacks & Friedberg

The episode begins with a retrospective of the All-In Summit, highlighting a spontaneous call from Donald Trump during Jensen Huang's presen

Key takeaways

• Token usage has flipped from 80% closed/20% open to 80% open/20% closed in just 12 weeks.
• Anthropic's IPO may be delayed due to internal concerns regarding human extinction risks and the launch of a new wet lab.
• Meta's Muse launch was intentionally 'slow-rolled' by Zuckerberg to ensure product safety and reliability.
• Open-source models like Alibaba's Qwen and Xiaomi's Mimo Pro are significantly reducing the cost of high-performance AI.
• The debate over whether frontier companies are 'labs' or 'corporations' centers on their desire for liability protection similar to Section 230.

Main topics

• AI Regulation and Product Liability
• Open Source vs. Closed Model Competition
• Anthropic IPO Risks and Corporate Governance

Notable quotes

The adults came. Daddy came. Elon came. Jensen came.

Conclusion

The episode provides a deep dive into the structural shifts in the AI

Transcript preview

Speaker 7 (0:00) All right, everybody, welcome back to episode 290 of the number one podcast in the world. It's the all-in podcast. With me again, it's the core four. David Friedberg is here, Chamath Palihapitiya, and David Sachs calling in. I am. Your host, Jason Calacanis, and we have a full docket today. I'd be remiss, though, if I didn't ask everybody for their fondest memories of the fifth annual All In Summit produced by Mr. David Friedberg. David, do you have any thoughts? How do you rate your performance as executive producer of the All In Summit? You Speaker 5 (0:39) know what we didn't do, what we forgot to do? At the end of day two, we didn't give a big shout out to our production team, to Kimber, to Lisa, to Nick. I think we, okay, I thought we should have brought them out, but I just want to give a big shout out to them. Speaker 7 (0:52) Bring them out for the standing up. The Speaker 5 (0:53) team backstage absolutely crushed it as always this year. And then I had a great time actually. You know, my favorite part is going to Universal Studios. You just get to have a good time. It's great. I'm Speaker 7 (1:03) reliving my childhood, Shabbat, with no trauma. Speaker 5 (1:07) By the way, shout out to our friends at DeLonghi. I got one of those coffee machines that we had at the event. I put it in my kitchen. You Speaker 6 (1:14) know what I got? I don't know if you guys saw, but there was a water machine in the VIP area. I as well. That is a zero plastic water machine. It goes through super filtration, no microplastics in the water. I have one now in our office. Thank you very much to them. I appreciate it. Amazing company. Sax, Speaker 7 (1:32) do you have a favorite moment from the show? Speaker 4 (1:34) Well, I think the highlight of the show, the thing that we'll remember for a long time, was the president calling in during Jensen's talk. And I think Jensen's overall, the content and Demeter really helped calm down this national panic over AI. It felt like we were the center of this national conversation. And I think it happened very organically. You remember there was... basically been the psyop for the last few weeks leading into the summit to convince everyone that humanity was going to go extinct unless we did exactly what the Democrats want and stop all AI development. And then, you know, we had first Satya came out and kind of calmed everyone down. And then Jensen did the same thing. And the president called in and did what he did best, which is seize total control of the narrative that I'm sure these Duma groups have been carefully planning for weeks and kind of thwarted that whole operation. Speaker 7 (2:25) nerds, enough, get back to work. You have a super intelligent, Trump-telligence is now the term. We're going to call it Trump-telligence going forward, but I think you make a good point. A lot of people wanted to know, Freeberg, if this was planned. Was the President Trump call-in planned, Freeberg? Take us behind the scenes. Basically, Speaker 5 (2:44) Jensen was connecting with him right backstage and said, hey, call me back, and then he called him while he was on stage. There was some texting going on. Speaker 4 (2:53) You know, it was completely spontaneous. It was spontaneous. Yes. It Speaker 5 (2:56) was not like a planned thing where Jensen was like, hey, let's do a whole thing. And so that was pretty amazing to have this conversation that Jensen was waiting for from the president to happen live. And clearly this was the topic du jour in D.C. as well as in the news and the media. And it was amazing to see it all happen on stage with us. Speaker 7 (3:15) And yeah, just to be clear, I thought it was a bit and I was like, well, put him on speakerphone. And Jensen says to me, how do you do that? And I just handed him, I always keep an extra mic there in case somebody's mic goes out, put him on speakerphone, put the microphone. And President Trump, the funniest president in history, funniest politician in history. Did you notice, Sachs, that he drops a one liner through a speakerphone perfectly? Perfectly. Jensen makes the greatest chips. Everybody tries to reverse engineer them. Impossible. Impossible. But he can't figure out how to press the speakerphone button. Okay. I was just like, how does he do it? How does he do it? It just drops the perfect line. But Shamath, in all seriousness, the nerds were losing their mind last week. And the adults came. Daddy came. Elon came. Jensen came. All the adults in the room came. And they just said, listen. If you have software that's defective and not ready to put in the world, maybe don't put it in the world and make it a little bit, you know, as you like to say, less brittle. So your thoughts on where we are in this national discussion, which went from slow down to maybe just release non-brittle resilient software? Speaker 6 (4:28) Well, I think that there was this classification that happened last week between Speaker 6 (4:36) up and coming organizations that want to call themselves labs, but are in fact, companies that have P &Ls and shareholders and have to abide by the same rules that everybody else does. And then the sophisticated companies that have lived in scrutiny for many decades and have learned the hard way. the process of making things robust and the cost of making things not robust. So what do I mean? On the one side, you saw, in many ways, it's like curious times creates curious bedfellows, right? You have the cohort is Satya, Jensen, Sachs, Zuck, Elon, the president, Lina Khan. We have product liability language. Strange Speaker 7 (5:23) bedfellows indeed. We Speaker 6 (5:25) have this product liability perspective in the United States. It's... meant to prevent companies from doing things. that they're not supposed to do, especially if they know that those things could be harmful. And on the other side, it's folks making the thing that were being, I think, a little naive and basically saying, we're not a company, we're a lab, protect us. And I think the mature response is you should have internal controls and you should robustly test things. And what Zuck said earlier and what Elon has said in the past is we're just under a lot of scrutiny. And so when things aren't perfect, we just slow things down. Elon has done that with FSD. You said this, Jason. He could have been at FSD three or four years ago, but he has publicly stated, we can't afford it because every accident that Tesla incurs is magnified 1,000x than what any other company would incur, and so we just have to go slower. Zuck last week said, we were about to roll out Muse, which looks like one of the most successful product launches in recent memory, has huge implications, actually, which we should talk about. And he had to slow roll it. Speaker 7 (6:29) He Speaker 6 (6:29) and Speaker 7 (6:29) Alexander, Speaker 6 (6:30) who I was talking to last night, had to slow roll it for a couple of months to get it perfect. So it's all within your control. It's just what do you choose to do? And I think that was the most interesting observation. We have to stop calling these companies labs. They're companies. And we have to be judged on the same standards, including product liability. Speaker 7 (6:50) Yeah. And there was a lot of coverage of this moment, Saks. Interestingly, some of the fake news media. We're like, on a popular podcast, and then they play the clip, but they're just afraid to say the name of the podcast. Pretty Speaker 4 (7:04) hilarious. I think Chamath is getting us into the first topic here already, which is whether we're going to have individual responsibility as our North Star or some sort of collective action. And the thing that always works the best is individual responsibility, whether it's at the level of individuals or companies. The relevant decision... to make in every case is, should we ship this product? Not, can we get the whole world to agree? And the problem with talking about these collectivized approaches is that they defray accountability away from the actual decision maker. And I think people sort of feel this, you know, like this week you had... Dario and Sam go to the United Nations and say we need some sort of global AI governance. And it kind of reminds me of like the billionaires who fly to Davos in their private jets and then rail against climate change. You know, there's something fundamentally sort of unrealistic and kind of out of touch about it. Speaker 7 (8:06) Hypocritical. Hypocritical. Speaker 4 (8:07) And it's like, guys, look, what we need is for you guys to make smart decisions about what to release. If a product is unsafe, which is to say it's unreliable or behaves unpredictably, then don't release it and stop looking to some sort of global governance as the answer. Because if you guys don't get this decision right, it's not going to matter what the United Nations does. And by the time the United Nations agrees on anything, we're going to be much further along in this whole journey. So there's something just very weird about the way that the principal actors in this keep talking. Because it really feels like they're trying to push off the responsibility. Speaker 7 (8:47) Well, there's a theory going around here, and I think it is worth us discussing here, that the Frontier Labs are covertly... They're Speaker 4 (8:56) companies. They're companies, yes. Let me just finish the... No, no, no, but Chamath made a really good point that I want to make sure we just don't miss. Okay. They're corporations. Yes. And when you call them labs, they sound like non-profit entities, and that's not true. Speaker 7 (9:08) Okay, so the corporations making these products... are looking for President Trump to specifically give them what the internet companies had with Section 230, which is shield them from liability. And that there is a consideration inside the Trump administration, the 47th administration, that maybe there could be some leverage put on these companies that they would give 10 % of their equity to the sovereign wealth front of the United States, et cetera. And then that... would result in them being given some pass when it comes to their liability. That's an insane thing to do. It's like giving some plane company that gives 10 % of their equity to the U.S. government the ability to not be responsible for their planes falling out of the sky. But what is your thought on that conspiracy theory? I'm sure you've heard it, Sachs. that they're looking for. You're Speaker 4 (10:05) making it sound like this is the idea of the Trump administration. That's not true. No, not Speaker 7 (10:09) the Trump administration. This is what the frontier lab slash corporations are pitching to them. And that there might be the rumor is there might be some people who are open to that in the administration. Again, this is a back channel. Speaker 4 (10:22) Here's what we know is that in the past week, you had virtually every Trump administration official come out and say some version of what I just said, which is first and foremost. the AI companies have to take responsibility for their own products and that we will not waive product liability. I think Speaker Johnson said it, the congressional... Besson said it. Besson said it in the most recent interview. Everybody said, if your game is to get a waiver of product liability or antitrust liability, if you're trying to form a cartel, if you're trying to collude, forget about it. We're not going to do it. So I think they completely threw cold water on that whole idea. And remember what President Trump said is... He just tweeted today, we have the DOJ. That's a guardrail. The administration is a guardrail. The administrative state, civil lawsuits, criminal lawsuits, all these things are there as potential liability to make sure that you make the right risk-reward trade-off. And again, at the end of the day, if those guys release an unsafe product, it's not going to matter what the United Nations agrees on. So we need them. to internalize the responsibility. Speaker 7 (11:28) Listen, come on, be candid with me. You got the inside goods here. Are they trying to lobby for that? Are they, do you think they're lobbying to get this protection? Yes or no? Speaker 4 (11:37) Come on. Well, I have heard, this is not like first person experience. They have not asked me like, hey, we want product liability waiver, but I have heard from people in Washington that they have been seeking something like that. Speaker 7 (11:48) It's hard to pin down, right? Yeah. Okay, fair enough. Freeberg. Your thoughts here as we are going to get into the first story here with the token maxing and the prices going down, but I'll give you the final word here on what America should do in terms of dealing with the corporations that are pretending to be labs who cannot take ownership of their own Frankenstein creations, which I think J.D. Vance, you know, I thought had a spectacular appearance. He was crisp and looked quite presidential, if I was going to describe it. Speaker 4 (12:21) You threw a lot of fastballs. Speaker 7 (12:23) And he knocked him out of the park, by the way. Fastball, you know, home run hitters want fastballs. You can't hit a fastball if it's out of the strike zone. So I gave him the three hardest questions I could think of, and he proceeded to knock all three over the fence. I think it was... Speaker 6 (12:35) JD's excellent. I mean, no, Speaker 7 (12:38) on the Sunday shows, they played the clips of the questions I asked him. Bang, bang, bang, Freeberg right over the fence. But, you know, he said, basically, if you're creating Frankenstein, here's an idea. Stop. Speaker 4 (12:51) And then he added, if you've already kind of let the cat out of the bag, then create anti-Frankenstein, basically. Create the safeguards. Speaker 5 (12:58) It Speaker 4 (12:59) makes total sense. I Speaker 5 (13:00) think Zuck gave a good interview on this. They took their time doing this Muse release this week. They said they wanted to get it right. They wanted to make sure that it was safe, that it didn't give people capabilities to do bad things. That's the responsibility that they took as a corporation. Anyone that's outputting a model should take the same responsibility. Speaker 5 (13:23) And Speaker 7 (13:24) this is an organization that understands it, having paid a series of fines from Cambridge Analytica to Speaker 5 (13:32) recently. We have a lot of laws that make it illegal to do a lot of things. All right, let's move on here. Speaker 4 (13:39) We have laws, okay. By the way, can I add like one philosophical point to this? So I think that implicit in, let's call it the Dario position, that let's call it the regulated free market that we have. His fundamental contention is that that will lead to a race to the bottom, that competition jeopardizes safety. It's in all of his blog posts, and many other people reflect this idea that competition equals lack of safety. And that is implicitly a critique of our economic system. It's basically a left-wing critique of our system, and it goes back a long way. I remember Back in the days of the Cold War, when we really had this duel between two different systems, do you want the sort of the free market system or do you want the communist system? A lot of people made the claim that the capitalist system would do things like jeopardize safety and beauty, that our buildings would always be uglier because the greedy capitalists wouldn't want to spend a dime on beautification or good architecture, things like that. In fact, it turned out that the Soviet system was the one that produced the cold gray landscapes. Why? Because people want beauty, they want safety, and therefore, if you have a well-functioning market economy, then it gives people what they want. And obviously, the American system won out in that battle. So I just fundamentally disagree with the idea that competition means lack of safety. As Zuckerberg and Jensen were saying, there are huge incentives on both the upside and downside to be safe. On the upside, no customer wants a product that is unpredictable and unreliable. That doesn't do what they say. No enterprise wants an agent that's going to leak their data or hack a competitor. And on the downside, as the president made clear, they've got tons of liability. Product liability, civil liability, administrative liability, criminal liability. So I just firmly disagree with this idea that competition is a bad thing. Competition is a good thing that can be harnessed to get the right result. And Speaker 7 (15:48) just to build on that, Shamath, our friend Nikesh from Palo Alto Networks, friend of the pod, he released in the past week or so, Unit 42, he's building, there's a business opportunity here to build a product to do continuous cyber defense and also a friend of the pod, George Kirch from CrowdStrike. He also released Falcon, his new product, to do cyber defense. Chamath, if these guys in the frontier labs slash corporations were so concerned about cybersecurity, why didn't they ever release a cybersecurity product and Palo Alto Networks and CrowdStrike are on it? They will. Speaker 6 (16:30) That's probably more in the second topic that we'll talk about, which is as the models essentially cluster in terms of capability, there's no differentiation. And so they'll be forced to move up the stack. But let's wait on that for one second. You have a point Speaker 7 (16:43) here you want to make, and then we'll go to the first. I do. Speaker 6 (16:45) The point I wanted to make is more philosophical and sociological, which is. Speaker 6 (16:53) The last time we heard the word lab in the social consciousness was during COVID. And there was a lab that was responsible for a leak of a virus that then killed 15 million people all around the world and caused $45 trillion of damage. It's the Wuhan lab. And I think it's important to recognize that there has been zero transparency or culpability around this. And I think part of why people are comfortable calling themselves a lab is because of that fact, that a lot of these young people that work at these places live through this and saw absolutely no responsibility taken by the lab that caused this leak. And now we're talking about leaks of different kinds and we have... organizations and corporations that try to fashion themselves as labs as well, but they're not labs. These are corporations. I just want to remind everybody, these are for-profit corporations that have absorbed hundreds of billions of dollars of debt and equity, that have thousands and thousands of shareholders, that have billions of dollars of market cap. And if I were them, I would not want to call myself a lab at any cost. And I think if you saw the recent... speeches that they gave at the United Nations, they're still using that terminology. They are not labs. These are for-profit corporations that are subject to product liability risk, and they should act accordingly. Speaker 7 (18:21) Yeah, and we've had a conversation here for a year about this. If this stuff is so dangerous, you don't have to release it, or my proposal, know your customer, give logs. Do not compare yourself to the Wuhan. Speaker 4 (18:35) lab. Well, Chamath, it's bad PR, but also I do think it's part of their like virtue signaling routine where they just don't want to say that we're companies, we have a for-profit incentive. They always want to say that we're either a research institution or they say things like we're a public benefit corporation, which no one really knows what the hell that means. So it's this patina of virtue signaling that obscures what's really going on. Speaker 6 (18:58) It tries to obscure the word corporation, but legally you're not allowed to hide it. I think they should have when they see we are all running corporations all in for all be it very small but mighty is a corporation I think we should call it a lab is a corporation lab is Speaker 7 (19:14) the all-in lab where we discuss different formulas no we're experimenting with ideas yes Speaker 4 (19:19) this is a this is a meme workshop but you know what Speaker 7 (19:22) these ideas are getting so dangerous that we don't want any liability for our opinion welcome to the all-in lab podcast here as we do topic zero, Freeberg. Is there a mindset in scientists' minds that needs to kind of have a hard stop and then the corporation kind of takes over? Maybe they could put the labs on the side here and then say, hey, these are the adults doing the corporation stuff. What are your thoughts here if you were going to architect how anthropic operates in the world? And maybe you could speak to scientists' views of the world versus capitalists and management teams. Speaker 5 (19:59) I Speaker 7 (19:59) don't have a good answer for you on that. Okay, well, you're doing it. You're the answer. Look at Ohalo. Let Speaker 5 (20:03) me kick off the first topic. What I wanted to do was just take a step back and look at the insane flurry of model releases happening. Yes. I think the pace of model releases happening on both the open source side and the premium side, the Anthropix, Groks, OpenAI's of the world is unfathomable right now. The list is insane. And just like the last 10 days. September 9th, DeepSeek 4.1 flash gets released. $1.20 for a million tokens output if they're hosting it. Key value cache from V1 of DeepSeek. Remember what a big deal DeepSeek was when it came out? The key value cache then was like 390,000 bytes. Now it's 890. Insane drop, insane efficiency improvements. On September 20th, Quen, which is the Alibaba model, comes out, version 2.1. It's an open weights model that outperforms Nano Banana 2, which is Google's big breakthrough insane image generation model. This thing is free, open source. You can put it on your desktop computer and run it at home to generate images for free. And it's made by Alibaba. Xiaomi put out Mimo on September 22nd, two days ago. Mimo Pro. performs on par with Claude's Opus 5, which is like this insane model that everyone thought was going to change the world and break the internet. And GPT-56 Sol across most benchmarks. It's a 309 billion parameter model. It is entirely open weight. You can download this model and put it on your server and run your own instance without paying anyone any money except for the cost of running the server. This model called Bonsai 2 came out on September 17th by Prism ML. This is actually a fork. of Quinn, which is another one of the Alibaba models. It's a 27 billion parameter model 98 % of the performance of the big Quinn model. It's 5.9 gigs in size, and you can run it on a local computer, you can have an NVIDIA chip, or have a good Mac Studio computer. And you can run this insane model on your desktop computer for free. And then the big crazy models, the big closed source models, Anthropic releases Opus 5.5 on September 22nd. OpenAI Astro comes out with Sol and Luna on September 23rd. Grok 4.7 comes out on September 21. Metamuse comes out on September 22nd. Any one of these stories would have broken the internet a year ago. And they all happened in the last 10 days. You now have an entire catalog of open weight models to generate visual information. to do VLA, where you can actually instruct it and it can control a robot. You have models far beyond the most advanced LLM from a year ago that are open weight that you can install and run on your desktop computer. We are past the point of Bernie Sanders put the super intelligence away. It can never be allowed in the real world. This stuff is out. The open source, open weight models are free. They are on the internet. You can go and download them and you can run them on your computer at home. I don't know in what world we think we're going to have a police state that's going to come bang down your door, take your desktop computer from you, and test it to see if you're running an LLM on your local computer. The efficiency wave is here. And the pacing at which these models are coming out, now all of this stuff happened in the last 10 days. It's just mind-blowing. We're now in this kind of frontier space where things are accelerating. The performance improvements, the cost decline, AI. is becoming so performative and so affordable, it is going to be ubiquitous, and we are all going to benefit from it. This is not just about the anthropic and open AI is with the closed world, making billions of dollars selling enterprises, hosted models, there is so much proliferation right now. So I just thought it was worth taking a step back to recognize where we are, because in the context of we have to stop