Open source and decentralized AI training models, like BitTensor's subnet 62 Ridges AI, represent a disruptive orthogonal attack vector that could outcompete frontier labs by achieving near-frontier performance at a fraction of the cost through distributed contributions.

(00:00:00) How many PRs do you think are going to get pushed to the core structural internet in 100 days? (00:00:04) What's the over-under number? (00:00:06) Because I'll give you a number. (00:00:07) You're going to say zero. (00:00:08) My answer to that is. (00:00:09) No, I'll say like 10,000, but it's going to be a meaningless thing. (00:00:12) But if it prevents your browser history from being released to everybody in the world, Chamath, that may be something that you're willing to, you know, let 100 days pass on. (00:00:20) I think you got Chamath's attention when you said browser history. (00:00:22) What about the dick pics? (00:00:26) Chamath is, he's going to release them himself. (00:00:31) We'll let your winners slide. (00:00:33) Rain Man David Saturn says. (00:00:38) We open sourced it to the fans and they've just gone crazy with it. (00:00:45) All right, everybody, welcome back to the number one podcast in the world. (00:00:48) David Freeburg is out this week, but in his place, the one, the only, (00:00:55) Our fifth bestie, Brian Gerstner. (00:00:57) I mean, why don't you ever give me puts a little namaste in your payday anymore? (00:01:01) You used to be stopping. (00:01:02) You used to be the greatest moderator, but now it's just, it's kind of late. (00:01:06) You know what? (00:01:06) These guys beat me up. (00:01:08) They beat me up and they just beat the joy out of me doing this program. (00:01:13) It's because you're a Rokana apologist now. (00:01:16) No, I will get into it. (00:01:18) OK, save it for the fucking show. (00:01:20) Rokana apologist. (00:01:22) Just because I said like, hey, they've stopped retard maxing and they've started doing like some logical things. (00:01:29) Yeah, okay. (00:01:29) Here we go. (00:01:30) It's great to be here. (00:01:31) Great to be here. (00:01:31) Good to have you. (00:01:32) Good to have you here. (00:01:34) And of course, we have David Sacks is back. (00:01:38) Everybody wants to hear from David Sacks. (00:01:39) We missed you last week, bestie. (00:01:41) We didn't beat the joy out of you. (00:01:42) We just tried to beat some of the hot air. (00:01:46) Any fluff that you can put on the show that just involves you talking and saying nothing is... (00:01:53) That's the stuff we got to. (00:01:54) Yeah, fair enough. (00:01:55) Okay, yeah, we'll cut it right now. (00:01:58) We'll cut it out and then we'll just put a promo in for the syndicate.com. (00:02:00) Thank you. (00:02:01) Also with us, Jamal Polly is here. (00:02:04) How's your retard maxing going since last week? (00:02:07) Did you have a retard maxing full weekend? (00:02:10) Did you have a good full weekend of just smoking cigars in the back deck and not ruminating about all the chaos you've caused in the last 20 years? (00:02:18) I think I've done generally more good than not. (00:02:23) you have, but there's been some chaotic moments. (00:02:25) Of course. (00:02:25) Don't think about it, Jamal. (00:02:26) You can't, bro, you can't have ups without downs, man. (00:02:30) It's like, what are you there to do? (00:02:31) Just like placate everybody and be a loser? (00:02:33) Are you there to be a winner? (00:02:35) Yes, you're in the arena, but have you stopped going to therapy after realizing that you're ruminating? (00:02:40) What's up with this sudden interest in retard maxing? (00:02:44) Are you like that clavicular for a retard maxing? (00:02:47) No, the world finally caught up with me. (00:02:49) That's it. (00:02:50) What do you, I mean, I've been retard maxing this whole time. (00:02:52) They just didn't have a name for it, guys. (00:02:55) Eli's videos are really good. (00:02:57) I watched two more this week. (00:02:59) Take us through what's so appealing about not ruminating, smoking a cigar, and just living your life. (00:03:05) Because what he says actually works at every level of society and every sort of thing that you may want to achieve. (00:03:13) Even if you're trying to like climb the rungs, (00:03:17) you very quickly learn that the more you want something, the less you're going to get it. (00:03:22) And I think that's like his real message is let go, live life, and just try stuff or don't try stuff. (00:03:29) And I think that detachment is really healthy for people. (00:03:33) I like it. (00:03:34) I like it a lot. (00:03:35) Who's the guy who says this? (00:03:37) I actually didn't know. (00:03:38) Elisha Long, but Eli, I think is how he goes by. (00:03:42) But he's fantastic. (00:03:43) He's got his YouTube channel. (00:03:44) Mark Andreessen found him. (00:03:46) And he's like, this guy is the new guy. (00:03:49) Modern day philosopher. (00:03:51) He gives you a roadmap for how to live your life, right? (00:03:53) New age sage. (00:03:54) What's the name of the guy, the character's name from Dune? (00:03:58) I was into girls. (00:03:59) I know these books. (00:04:00) I was dating girls. (00:04:01) He's the Lizan al-Ghayib of the modern internet. (00:04:04) This is why we need Friedberg here, is to explain these deep holes. (00:04:09) All right, listen, we got a lot to get to. (00:04:11) The basic point is build something and don't ruminate, okay? (00:04:14) Ruminating is just not worth it. (00:04:15) Just everybody go forward. (00:04:16) No, just do stuff. (00:04:17) Stop blathering in your own head. (00:04:19) Just do stuff. (00:04:20) Absolutely. (00:04:20) All right, listen, speaking of doing stuff, Anthropic is withholding its newest model, Mythos. (00:04:26) I'm using the Greek pronunciation, its newest model, Mythos, saying it is far too dangerous for any of us to have access to it, according to the company. (00:04:35) The model autonomously found thousands of vulnerabilities, including bugs in every major operating system and web browser. (00:04:42) This little study they did included 20-year-old exploits that had been missed by security audits for decades. (00:04:50) Some examples, they found a 27-year-old vulnerability in OpenBSD used in firewalls and critical infrastructure. (00:04:56) They found a 16-year-old bug in FFmpeg. (00:04:59) That was missed by automated tools after 5 million scans. (00:05:03) The Linux kernel, (00:05:05) all kinds of bugs they found, they released a hype video hyping up why they were not going to share this model. (00:05:14) Here's Dario, come on the program anytime, brother. (00:05:16) But as a side effect of being good at code, it's also good at cyber. (00:05:20) The model that we're experimenting with is by and large as good as a professional human at identifying bugs. (00:05:29) It's good for us because we can find more vulnerabilities sooner and we can fix them. (00:05:33) It has the ability to chain together vulnerabilities. (00:05:36) So what this means is you find 2 vulnerabilities, either of which doesn't really get you very much independently. (00:05:42) But this model is able to create exploits out of three, 4, sometimes 5 vulnerabilities that in sequence give you some kind of very sophisticated end outcome. (00:05:50) All right, Brad, by the way, that set they're using there, that's the same room those guys play Dungeons and Dragons in every Sunday. (00:05:56) Brad, you're an investor in this company. (00:06:01) Is this virtue signaling or is it reality? (00:06:04) Is this a good move by them to not release this model and be thoughtful, give it to a handful of people, and just... (00:06:12) I actually think they deserve a ton of credit here, and let me walk you through why. (00:06:21) The company could have just released Mythos, broken a lot of core things on the internet. (00:06:26) Oftentimes in Silicon Valley, we say move fast and break things. (00:06:29) In this case, it means just releasing the model to move further ahead of your competition. (00:06:33) But here the company realized it would wreak havoc. (00:06:36) They ran their own vulnerability testing. (00:06:38) They saw that it would allow offensive hacking and people to expose browsers and browser history, expose credit cards, you know, on the internet. (00:06:46) So, you know, what I like about this is they didn't need government to hold their hand on this. (00:06:51) We have plenty of government regulations. (00:06:54) They know it's in the best long-term interest of the company and the industry. (00:06:58) You know, so they set up Project (00:06:59) Glasswing. (00:07:00) It's an AI-driven kind of cyber coalition, Apple, Microsoft, Google, Amazon, JP Morgan, 40 of the most important companies. (00:07:09) And their goal is very simple. (00:07:10) Let's spend 100 days using advanced AI to find and to fix and to harden these software vulnerabilities before hackers exploit them. (00:07:19) Now, what I think this represents, Jason, is a threshold that we're crossing. (00:07:25) Mythos and SPUD, which is going to be out from OpenAI any day now, which is the first Blackwell-trained model at OpenAI, they represent the beginning of what I would call AGI models. (00:07:37) These are models with massive step function improvements and intelligence, and they're just too smart to be released immediately. (00:07:46) You know, and by the way, there was nothing that said that every time you finish a model, you got to immediately release it GA. (00:07:53) So they set up this idea of sandboxing, building defensive alliances, in order to move away from that regime. (00:08:00) I think it shows, and Saxon and I have talked about this a lot, so I'm interested to hear what he thinks. (00:08:06) It shows you can trust the industry and market forces in coordination with the government. (00:08:11) They were talking to the government about this, but they're not relying on some top-down regulation. (00:08:17) in order to do this. (00:08:18) They laid out a blueprint that seems to me very pragmatic, that now that we're at this threshold, we're going to sandbox these things. (00:08:26) I think that OpenAI will end up doing the same thing. (00:08:28) I think Google will end up doing the same thing. (00:08:31) It's an aggressive way to keep the pressure on and win the race at AI while making the trade-offs to protect safety. (00:08:39) So, you know, I think you're always going to have to make these trade-offs. (00:08:42) I think in this case, it was a great move by Dario and team, and I think they deserve a lot of credit. (00:08:47) Sacks. (00:08:47) When you look at this, we had Emil Michael on the program a couple of weeks ago, it might have been four or five weeks ago, and we had a very thoughtful discussion about, hey, if the government is going to have these tools, Anthropic wants to withhold them, and what is the proper relationship there? (00:09:04) You have to think (00:09:05) that the government, and I know you don't speak for all parts of the government, if you were just going to run through the game theory, they must have gone to the government and said, listen, this thing is so powerful, it can put together two or three hacks, create a novel attack vector, and this is incredibly dangerous. (00:09:21) What if China has it? (00:09:22) And if this thing is as powerful as Dario says it is, then this is an offensive weapon as well for us to take out, let's just pick, you know, a prescient issue. (00:09:33) the North Korea's ballistic missile program. (00:09:36) This is equivalent, the way it's being described, as the Manhattan Project, perhaps. (00:09:41) So what are the chances, two-part question for you, Sachs, that China already has this and is using it? (00:09:46) And do you think Dario is doing the right thing by regulating themselves? (00:09:52) I think Anthropic has proven that it's very good at two things. (00:09:57) One is product releases. (00:09:58) The second is scaring people. (00:10:01) And we've seen a pattern in their previous releases of at the same time they roll out a new model or new model card, something like that. (00:10:09) They also roll out some study showing really the worst possible implication of where the technology could lead. (00:10:17) We saw this last year, about a year ago, they rolled out this blackmail study where supposedly the new model could blackmail users. (00:10:25) There's been a whole bunch of these things. (00:10:27) Actually, I went back to Grok and I just asked, Hey, give me examples where Anthropic has basically used scare tactics, and it's a pattern. (00:10:35) Okay, it's a pattern. (00:10:37) Okay. (00:10:37) These guys, I'm not saying it's not sincere, but they have a proven pattern of using fear as a way to market their new products. (00:10:47) And (00:10:48) If you think back to, again, my favorite example is this blackmail study, where they prompted the model over 200 times to get the result they wanted, and that result was clearly reverse engineered, and it got them the headlines they wanted. (00:11:02) And I would say the proof that it's reverse engineered is we're now a year later, there's a bunch of open source models out there that have the same level of capability that anthropic model had. (00:11:14) And have you seen any examples of blackmail in the wild? (00:11:17) I don't think so. (00:11:18) So in other words, if that study were true in the sense of being a likely outcome of that model, I think you would see examples in the wild of that behavior, and we haven't seen any of that in the past year. (00:11:30) Now, let's talk about this specific example with cyber hacking. (00:11:35) I actually think that this one is more on the legitimate side. (00:11:39) I mean, look, the reason why I bring this up is anytime anthropic is scaring people, you have to ask, is this... (00:11:44) a tactic? (00:11:45) Is this part of their Chicken Little routine? (00:11:47) Or is it real? (00:11:49) Are they crying wolf or not? (00:11:50) I actually would give them credit in this case and say, this is more on the real side. (00:11:55) It just makes sense, right? (00:11:56) So that as the coding models become more and more capable, they're more capable of finding bugs. (00:12:01) That means they're more capable of finding vulnerabilities. (00:12:03) And like one of their engineers said, that means they're more capable of stringing together multiple vulnerabilities and creating an exploit. (00:12:10) And so I do think that over, say, the next six months, we're going to have this (00:12:14) call it one time period of catching up where AI-driven cyber is going to be able to detect a whole range of bugs that maybe have been dormant over the past 20 years across a wide range of systems. (00:12:28) And so I do think that there is real risk here. (00:12:32) And I do think, therefore, that having this pre-release period makes a lot of sense where they're giving the capability to all these software companies that have existing code bases (00:12:42) to use the tool to detect the vulnerabilities for themselves so they can patch them before these capabilities are widely available. (00:12:49) And by the way, it won't just be anthropic that makes these capabilities available. (00:12:54) We know that like, let's say the Chinese open source models like Gimme K2, it's about six months behind. (00:12:59) So we have a window here of maybe six months where we're still in this pre-release period where I think companies that have large code bases can get advanced access to this model (00:13:12) And I guess OpenAI is going to release a similar thing in the next few weeks. (00:13:16) I do think that every company or IT department or CISO that is managing code bases should take this seriously and use the next few months to detect any, again, like dormant bugs or vulnerabilities and roll out patches. (00:13:34) If everybody does their job and reacts the right way, then I do not think it will be the (00:13:38) the doomsday scenario that Anthropic is sort of portraying. (00:13:42) But it's one of these things where the fear might end up being a good thing in order to drive the correct behavior. (00:13:50) So I ultimately think this is going to work out fine, but you do need everyone to kind of pay attention, use the capabilities, fix the bugs, then we're going to get into (00:14:01) big arms race between AI being used for cyber offense and AI being used for cyber defense, but it'll be a more normal sort of period. (00:14:09) Chamath, we have Dario and a number of the participants here are taking this super seriously. (00:14:16) They're making a big statement. (00:14:17) Sack's very nuanced, I think, take there. (00:14:20) What's your take on how do these companies have it both ways? (00:14:24) hey, this shouldn't be regulated. (00:14:26) This should be regulated. (00:14:28) If this is in fact a cataclysmic, oh my God, they're going to hack everything. (00:14:33) What if the Chinese have this right now? (00:14:35) That would speak to more government, either coordination, regulation, or some kind of relationship between the CIA, the FBI for domestic stuff and these companies, because (00:14:48) It is a non-zero chance that the Chinese have an equal capability here. (00:14:52) We're assuming they're behind, but who knows what they're doing behind closed doors. (00:14:56) So what's your take on this? (00:14:57) Is it The Boy Who Cried Wolf, or is this the real deal now? (00:15:01) I think it's mostly theater. (00:15:03) Okay. (00:15:05) In February of 2019, when Dario was still at OpenAI, they did the same thing with GPT-2. (00:15:14) that was a 1.5 billion parameter model, which sounds like a total fart in the wind in 2026. (00:15:22) But at that time, this 1.5 billion parameter model was supposed to be the end of days. (00:15:28) And it was supposed to unleash this torrent of spam and misinformation. (00:15:32) And that was the big bugaboo at the time. (00:15:34) And so what happened? (00:15:35) They went through this methodical rollout over six or nine months. (00:15:38) They started releasing the smaller parameter models, and then they scaled up to the big 1.5 billion parameter model. (00:15:44) And at the end of it, was a huge nothing burger. (00:15:47) If you actually think that Mythos is capable of doing what it says it can do, two things are true. (00:15:54) One is a very sophisticated hacker can probably do those things right now with Opus. (00:16:01) And 2, if these exploits are this easy to find, whether you use Opus or whether you use Mythos, the reality is you'd have to shut down the internet for about 5 years to patch them all. (00:16:15) So when you see like a large multi-trillion dollar GCIP bank, (00:16:20) It's a bit of theater. (00:16:21) Why? (00:16:22) What do you think they can actually accomplish in two months? (00:16:25) Do you actually think that if there's these vulnerabilities, it's all going to get fixed? (00:16:30) Let's give them six months. (00:16:31) Let's give them nine months. (00:16:33) But the reality is that capitalism moves forward, the funding needs moves forward, and the need for these guys to build adoption moves forward. (00:16:43) And that's going to supersede what this is. (00:16:46) So I do think that Sachs is right. (00:16:49) that they have figured out a very clever go-to-market muscle here and a go-to-market motion that activates hyper-attention and hyper-usage. (00:17:00) And so I give them tremendous credit. (00:17:02) And I'll maintain what I've maintained before. (00:17:04) Anthropic is shooting the lights out right now. (00:17:07) This is like Steph Curry going bananas. (00:17:10) From everywhere on the court, these guys are hocking threes. (00:17:12) Clay Thompson. (00:17:14) It's all in that. (00:17:15) Okay, so huge kudos to Anthropic. (00:17:19) But we've seen it before. (00:17:21) We saw it when these folks were the principal architects at OpenAI. (00:17:25) We're now seeing the same playbook here. (00:17:27) I think we'll look back, and I think what we'll say are these two things. (00:17:31) One is, if we're really going to patch all these security holes, we need to shut down the internet. (00:17:36) for some number of years, honestly, literally years. (00:17:39) And the second is an advanced hacker can probably do this today with Opus if they really wanted to. (00:17:46) Okay, hey, Brad, I'll get you in here for the last word. (00:17:49) I'm going to go with, yeah, maybe they did cry wolf before, but based on what I see with these models advancing and using them, and I'm using a lot of the open source ones right now from China, (00:18:01) I think that this is like code red kind of moment. (00:18:03) This is DEF CON. (00:18:04) Like we should be taking this deadly seriously. (00:18:07) And I think these companies got to coordinate with the CIA and this is equally a defensive as offensive opportunity. (00:18:14) Do you think this is... (00:18:15) You're asking for the nationalization of AI now? (00:18:18) No, I actually, I don't think it should be nationalized, although I did see people sort of insinuating that. (00:18:24) I think these companies need to build a group, Brad, that (00:18:28) work and coordinate with the CIA. (00:18:29) I assume that they're already doing this. (00:18:31) I'm assuming Emil Michael and Trump and everybody have these people in a room and that they've given the DEFCON and said, hey, how can our government use this to stop bad actors? (00:18:43) And (00:18:44) this is already being coordinated with the CIA and the FBI. (00:18:46) I am 100% certain of that, Dario went to them and said, look what we found. (00:18:51) This is the real deal. (00:18:52) I'll give you the last word on this, Brad, since you're an investor in both companies and you know them quite well. (00:18:56) The Frontier Model Forum, which was put together in 23, is cooperating on anti and adversarial distillation stuff as we speak, right? (00:19:05) They don't want to make it easy on, you know, so Google and (00:19:09) and OpenAI and Anthropic, they're coordinating on this stuff. (00:19:13) there are times where I've pushed back on Anthropic because I thought it was, perhaps regulatory capture or something else. (00:19:19) This is very different in my mind, right? (00:19:21) He could have easily, Dario could have easily come out and said, oh my God, we passed a threshold. (00:19:25) We need to have a government moratorium. (00:19:27) Remember, even our friend Elon called for a six-month moratorium in 2023 because of civilization risk. (00:19:34) This guy didn't do that. (00:19:35) Instead, he said, okay, what should we do? (00:19:38) I'm going to get 40 of the leading (00:19:39) companies together. (00:19:40) We're going to spend 100 days sandboxing, hardening the systems, and then we're going to keep pushing forward. (00:19:45) What do you honestly think is going to get accomplished in 100 days? (00:19:48) How many PRs do you think are going to get pushed to the core structural internet in 100 days? (00:19:53) What's the over-under number? (00:19:54) Because I'll give you a number. (00:19:55) You're going to say zero. (00:19:56) My answer to that is. (00:19:58) I'll say like 10,000, but it's going to be a medium. (00:20:00) But if it prevents your browser history from being released to everybody in the world, Chamath, that may be something that you're willing to let 100 days pass on. (00:20:08) I think you got Chamath's attention when you said browser history. (00:20:10) What about the dick pics? (00:20:14) Chamath is, he's going to release them himself. (00:20:16) Right now Chamath's like, hey, Chinese hackers, here are my dick pics. (00:20:21) Please put them out. (00:20:21) Oh my God. (00:20:23) We have to be out there complimenting when they're doing the right things or relying on the market rather than running to the nanny state and saying, do more of this. (00:20:30) So (00:20:30) This to me was just an example of a good balance. (00:20:33) I'm sure we're going to have plenty of debates about this in the future, but this is 1 I would like to see more of. (00:20:39) This is why, to use your word, Jake, I tried to have a more nuanced take, is because we have no choice but to take this seriously. (00:20:46) Whether it's total theater, whether it's fear-mongering, and they do have a pattern around this, we can't take the risk, right? (00:20:53) And it does logically make sense that (00:20:56) As these models become more and more capable at coding, they're going to get better at cyber. (00:21:00) And there's going to be that one time period where you're moving from pre-AI to post-AI, and you need a patch for that. (00:21:07) So my guess is we're going to see a lot of patches over the next few months. (00:21:10) I think that will resolve the problem. (00:21:14) I think this is a case where I'm going to give them the benefit of the doubt. (00:21:18) I think that, you know, I've criticized them in the past. (00:21:22) I think that blackmail study was embarrassing to the level of being a hoax. (00:21:26) But I think in this case, I'm going to give them credit and say that I think that it's legit. (00:21:32) So it's not the anthropic hoax. (00:21:33) This could be legit. (00:21:35) I, you know, looking at this. (00:21:36) We have no choice but to treat it that way. (00:21:38) Of course, yeah. (00:21:39) I mean, even if two things could be true at the same time, Sachs, they could have used this tactic before. (00:21:45) It could be performative, like the video with the dramatic music in the background. (00:21:49) It does. (00:21:50) have a little bit of drama to it. (00:21:52) And the way they presented it is very dramatic. (00:21:55) But it does make logical sense that the one company that made the bet on code, bigger than anybody else, would be the one who would discover this quickest. (00:22:05) And in 100 days, that's a pretty good, that's a pretty big advantage versus the hackers. (00:22:10) But let me say one more point there, Chamath. (00:22:13) The most important thing that people haven't talked about here is (00:22:17) The amount of code being pushed right now because of these tools is 10x, 100x in most organizations. (00:22:23) So we need to have this type of security embedded in these new coding tools to do it in real time. (00:22:29) That's the opportunity. (00:22:30) There should be real time correcting of this. (00:22:33) If this is real, they pick the wrong companies. (00:22:36) Meaning there are energy companies, folks that control nuclear reactors, (00:22:42) There are airplane companies that are flying hundreds of thousands of people in essentially manufactured missiles of like streaming gas going at 500 miles an hour. (00:22:54) None of those companies were the ones that were included in this. (00:22:57) And so I think if you really thought that this was end of days, at a minimum, we can agree, maybe we should have expanded the circle a touch. (00:23:08) maybe those are customers of the ones they're including here. (00:23:10) Anyway, this is a really important story. (00:23:13) We'll obviously track it in the coming weeks to see what turns out to be reality. (00:23:17) And Dario, do come on the program at some point. (00:23:20) Hey, Brad, will you get Dario to come on the program? (00:23:22) I've invited him like three times. (00:23:23) I got his phone number. (00:23:23) He's ghosted me. (00:23:24) I don't know why. (00:23:25) Wait, he's ignored you? (00:23:26) I literally got an introduction from the number, like one of the number one venture capitalists in the world. (00:23:31) He's on the cap table very early. (00:23:33) He just won't respond. (00:23:34) I don't know why. (00:23:36) I would tell you, Dario's podcast with Dworkish, who I think is an excellent podcaster, I've listened to that three or four times, taken notes every time. (00:23:44) It is a really exceptional piece, really exceptional piece of work by them. (00:23:49) All right, let's keep moving. (00:23:50) We got a lot on the jock. (00:23:50) You may once again be tarred with your affiliation with us. (00:23:54) Poor you. (00:23:55) I mean, I don't care. (00:23:56) Literally, I've got friends on both sides of the aisle. (00:23:59) I have friends. (00:24:00) Of course you do. (00:24:01) Even J. (00:24:02) Cal. (00:24:03) Even J Cal has friends everywhere. (00:24:05) Let me ask Brad a question here, just while we're on the topic of Anthropic. (00:24:08) There was a really interesting story or tweet, I guess you could say, by the founder of Open Claw that... (00:24:14) Peter. (00:24:15) Peter, yeah. (00:24:16) What's his name? (00:24:16) Peter Steinberger. (00:24:18) Steinberger, yeah. (00:24:20) Renowned coder, created Open Claw, which is kind of the thing that launched the sole agent era now, I guess you could say. (00:24:27) In any event, he said that Anthropic was cutting off his access (00:24:33) This is on the docket. (00:24:35) It's a little bit nuanced. (00:24:38) Everybody using OpenClaw would take their $200 a month subscription to Anthropic, which was essentially like people were using more tokens and it's an average. (00:24:47) The people from OpenClaw, it is very verbose, and those people are 100X the usage of the average subscriber. (00:24:54) So he said, you can't use your 200, you have to use the API. (00:24:58) You move from the $200 plan to the API, add a 0 to your token use, or more. (00:25:04) And so they essentially ankled OpenClaw (00:25:08) And then 10 days later or less, they released or announced their new agent technology, which is, according to them, a safer, better version of OpenClaw. (00:25:18) So hey, all's fair in love and war, and they have basically shot a huge cannon across the bow of OpenClaw. (00:25:26) Can you just explain that exactly? (00:25:27) So I think you're right that they systematically copied feature by feature of OpenClaw, incorporated that into Claude, and then the coup de grace was basically cutting off OpenClaw. (00:25:38) Yes. (00:25:39) Can you just explain exactly what they did? (00:25:41) Okay, very simply, when you buy a subscription to these services, they have blended your usage across many users. (00:25:50) So there's, you know, 9 out of 10 users use less than the tokens they're paying for, and the top 10% use much more. (00:25:56) When OpenClaw became a phenomenon, the number one open source project in history on GitHub, with all of this usage, people went crazy. (00:26:05) And you heard me talking about how crazy I went for it. (00:26:07) Those people with the $200 subscriptions were using 2000, $20,000 worth of tokens. (00:26:13) So they said, you can no longer use your subscription to, you know, either your professional or enterprise subscription at $200 and plug that into your OpenClaw. (00:26:22) You now have to go to the API and pay per usage. (00:26:25) So no more like unlimited essentially. (00:26:28) If you use Anthropic's own agent harness, are you part of the bundled flat rate? (00:26:34) You can assume that that's what they'll do, which if you were thinking on an antitrust level, might be token dumping or price dumping. (00:26:40) I'm not saying like I'm ratting them in with it. (00:26:44) price dumping or bundling. (00:26:45) When you price something under the market price in antitrust, that would be price dumping, right? (00:26:50) And if you were to bundle, it would be like the bundling issue. (00:26:54) Critically important, you can use OpenClaw via Claude API. (00:26:58) And every company has a right to set the price for its products. (00:27:00) It's just saying that you were under their current regime, they were selling dollars for 10 cents via OpenClaw because these were such power users. (00:27:09) And now they're just saying, we have to price this rationally, but we're happy to have you guys use the API. (00:27:14) Okay, but Brad, when you use the open claw competitor that Anthropic now offers, are they subsidizing that? (00:27:22) Are you paying? (00:27:23) We don't know yet because it's enclosed banner. (00:27:25) So in other words, what I'm saying is if they charge for API usage, their own first party agent harness or system, then that would be apples to apples. (00:27:34) But if they end up charging the bundled flat rate, let's say, for their stuff, but then (00:27:42) charge the metered rate for third-party stuff, you could make a bundling argument. (00:27:46) Sure, And you could say it's anti-competitive, assuming that Anthropic has dominant market share in coding, which I think most people would say they do at this point. (00:27:55) And assuming that it's the same product. (00:27:57) I mean, the reason most enterprises will probably use the Anthropic version of this Agentic product is because it meets all of your security parameters, right? (00:28:08) So Altimeter runs, (00:28:10) a lot of stuff on Anthropic, they're already integrated within our data warehouse, our data lake, things of that nature. (00:28:16) So just letting OpenClaw loose on the altimeter, you know, data set would not be wise. (00:28:22) And so it's a different fundamental product. (00:28:24) No, I get that. (00:28:25) And I think that Anthropic has a huge advantage, let's say cloning OpenClaw and just building it into Claude. (00:28:31) I'm not denying that. (00:28:33) To me, that would be the reason why they don't need to do price discrimination is because there's already a very good reason. (00:28:39) to use the, let's call it the bundled offering on a featured basis. (00:28:42) But the question I'm specifically asking is whether they're giving themselves a price advantage. (00:28:48) Because I think Brad is giving the most generous interpretation. (00:28:52) You're taking a more cynical one. (00:28:53) I'm with you, Sacks. (00:28:54) I'm 100% on the cynical side. (00:28:55) Open Claw is so powerful. (00:28:57) It's got so much momentum that not only is Anthropic trying to ankle it. (00:29:03) I believe when Sam Altman bought it, was (00:29:06) And he didn't buy OpenClaw itself. (00:29:07) He hired, Aqua hired Peter. (00:29:10) I believe it was to subvert the open source project to get Peter's next set of genius ideas inside of OpenAI as opposed to letting them go there. (00:29:18) People are going to say I'm A conspiracy theorist. (00:29:20) But this is the number one focus. (00:29:22) And let me just give you a list of who is trying to kill OpenClaw slash compete with them. (00:29:28) Obviously, you have Anthropic, but also Perplexity Computer launched. (00:29:32) It's awesome. (00:29:33) I've been using it. (00:29:34) Anthropic has this Claude Manage Agents. (00:29:37) They dropped that on Wednesday, April 8th, yesterday. (00:29:40) Today's Thursday when we tape, you guys listen on Fridays. (00:29:44) And then you have Hermes. (00:29:46) Agent, that was released on February 25th. (00:29:48) That's also open source and very good. (00:29:50) So that's in the open source camp. (00:29:51) Alibab is coming out with one. (00:29:53) That's going to be based on their Quinn model. (00:29:55) Then you have Elon who said he's got something called Grok Computer coming out of MacroHard, which is a play on words for Microsoft. (00:30:03) In addition to that, Amazon and Apple are preparing new releases of their retard maxing assistants, Alexa and Siri, that will be less retarded in this new version. (00:30:14) and then nothing out of Satya and Microsoft yet. (00:30:17) So the number one goal, I believe, in the large language model, frontier model space is to kill this open source product. (00:30:26) No, I mean, come on, like why they're building multi-functioning agents that can move from answering questions to actually doing something for you. (00:30:35) Like you got to do that because that's what consumers and enterprises wants. (00:30:39) It doesn't mean that it's about killing OpenClaw. (00:30:41) It's just this is an obvious thing that they have the right to do it. (00:30:44) But this is a giant movement to stop it, because this is the equivalent of having an open source Android-like player in the market. (00:30:52) And that could be incredibly disruptive. (00:30:54) I believe open source is going to win the day on the large language models and take 90% of the token usage. (00:30:59) And I think the entire Frontier model space (00:31:02) could be undercut by open source. (00:31:03) And I think they realize that SLMs, the smaller language models that are verticalized now, that will run on desktops and laptops and is even starting to run on the top ones, that is their biggest competitive threat. (00:31:16) And I hope it happens. (00:31:17) And all due respect to your investments, Brad, I think this technology and the interface is, you know, he plays bets. (00:31:23) But I think it's imperative that the agent level, which is (00:31:28) Essentially, your entire life, you don't give that to Anthropic. (00:31:31) You don't give that to OpenAI. (00:31:32) That's your entire business, your entire life. (00:31:34) It is foolish for you, Brad, to give your entire business and all the knowledge you have to Anthropic through that, unless you're just doing it to boost your investment in those companies. (00:31:44) But I would be very concerned if I was you with putting all of your knowledge that you've earned over a lifetime into any of these large language models. (00:31:52) All right, Jake, let me ask you, can I ask a question? (00:31:55) Thank you for that impassion. (00:31:57) Thanks for coming to my TED Talk. (00:32:00) Yes, thank you for that TED Talk. (00:32:02) I have a yes or no question for each of you. (00:32:06) Do you believe that Anthropic has dominant market share in coding right now? (00:32:11) Yes or no? (00:32:14) No. (00:32:14) In coding? (00:32:15) Yes. (00:32:16) They had the lead, but not dominant. (00:32:18) I think it's a trillion dollar market, and these guys have less than 10% of it today, so it's hard to make a case. (00:32:24) What percent of coding tokens do you think that Anthropic is providing the market right now? (00:32:30) Greater than 50%. (00:32:30) Yeah, that's true. (00:32:31) Okay, that's called dominant market share. (00:32:34) I don't know about that. (00:32:36) Greater than 50% on the market. (00:32:37) You got to look at what the TAM is. (00:32:39) You got to look at what the TAM is, right? (00:32:42) There are a lot of people who provide, you know, that are in the business of helping people write software. (00:32:48) I'm not saying it's a permanent condition, but if you're telling me that today, (00:32:53) Anthropic is delivering over half of the coding tokens. (00:32:57) That's clearly a dominant position in the market for coding. (00:33:00) It's an early market. (00:33:01) It could change, but... (00:33:01) If I were representing them, David, I would say nine months ago, everybody call us... (00:33:07) out of the game. (00:33:08) We were being destroyed by OpenAI. (00:33:10) In 3 months now, people are saying we have dominant market position. (00:33:13) This is the fastest changing, most competitive market in the world. (00:33:17) I think it would be very hard pressed to walk into some district court and make the case that these guys have somehow already formed a monopoly against Amazon, Google, Microsoft, OpenAI, et cetera. (00:33:29) Well, I'm not saying it's already a permanent monopoly, but (00:33:33) I am just asking about market share. (00:33:35) And I do think you guys all agree that- Let's get Shemaf, go ahead. (00:33:38) They probably have 50 to 60% market share, because I think Codex is actually quite broadly used as well. (00:33:46) But that belies the more important point, which is AI-enabled coding, I think, is still 5% of the broad market. (00:33:53) So it's kind of a nothing burger. (00:33:55) Yes, they're leading, but they're leading in something (00:33:57) that isn't that big yet. (00:33:58) Now, you would say, how could it not be big? (00:34:01) And what I would say is, because most of the stuff that's being written is still white sheet de novo code. (00:34:08) And I think the ugly truth is, I don't care what model you have, but the long horizon ability for any of these models to actually build enterprise-grade software is still shit. (00:34:20) S-H-I-T shit. (00:34:22) And that's the actual lived experience. (00:34:25) Not for me, but when I call on our customers, half a trillion dollar banks, 100 billion dollar insurance companies, none of these guys are like, wow, it just works out-of-the-box. (00:34:34) It doesn't work. (00:34:36) So most of it is still hand-tuned. (00:34:39) So until I can honestly tell you that we can point a model at this with the right guardrails, which I can't today, what I would say is it's a small market that will become large as these models become better. (00:34:54) But we are in the world where we have 50 years of accumulated tech debt as a world. (00:35:01) And I suspect when you enumerate the number of lines that represents, it's hundreds of trillions of lines of just pretty marginal mediocre code to bad code. (00:35:12) On top of that, we have all these legacy languages. (00:35:14) I'll tell you one of our customers, they have to go and get 60-year-old pensioners to come into the office to interpret cope. (00:35:22) No, I'm not joking. (00:35:23) This is a- COBOL, Fortran. (00:35:24) This is a $100 billion a year revenue company. (00:35:28) And that's how they solve these problems. (00:35:30) It's not Opus just solves it. (00:35:32) So I would just keep in mind that most of the tech debt in the world that exists, 99% of it is still poorly addressed by these models. (00:35:41) we are untying this Gordian knot, it's going to take decades to do it right. (00:35:46) So all the breathlessness about all this other stuff, I really think it's not where the money is. (00:35:50) It's not the big time stuff. (00:35:52) And you can tell me, oh yeah, it's going to be the future. (00:35:54) And I would say, tell this business there's $100 billion a year of revenue and 50 million billing relationships that all of a sudden you're going to open claw your way to a solution. (00:36:04) It's bull . (00:36:05) Not to say that you can't have a great chief of staff and not to say you can't do some useful stuff and trickery and have a good knowledge base. (00:36:13) I'd like that too. (00:36:15) But the core things that your lived experience sits on today is a mess of tech debt that will get very slowly replaced. (00:36:23) And that's just the reality of life. (00:36:25) And there are competitors that are extremely disruptive. (00:36:28) I'll tell you about one, we talked about BitTensor, TAO on this program a couple of weeks ago when we had the (00:36:34) Jensen interview, you brought it up, actually, Chamath, there's a project that's subnet 62, it's called Ridges AI. (00:36:40) And what they're doing is a competitor that is not only open source, but anybody can contribute to it. (00:36:47) They spent about $1,000,000 in TAO, like rewards, and in 45 days, they hit 80% of what Claude 4 is. (00:36:55) And they did that in under 45 days. (00:36:56) The way that works is they give rewards for people (00:37:00) who, and they can do this anonymously, make that coding product, which is like Codex or Claude Code better. (00:37:06) That flywheel is racing right now with participation in the same way Bitcoin is. (00:37:12) So you're going to see a lot of open source and these crypto open source combinations. (00:37:17) And anybody who's not investigated this, I highly recommend you investigate this. (00:37:23) I do think you're right about one specific thing. (00:37:25) I would put zero, literally the probability 0. (00:37:29) of any important company worth anything more than a dollar, having and outsourcing their production code to an open source project. (00:37:37) That'll never happen. (00:37:38) However, what will happen though, is when you look at the cost of training this 10 trillion parameter model on Blackwell, and when you look in the future, let's just say in six or nine months, that a 15 or 20 trillion parameter model is going to get trained on Vera Rubin, (00:37:56) I think, Jason, where you are right, I have zero, and just to be clear, I have no investments in this at all. (00:38:03) I do, to be super clear. (00:38:05) I'm just observing, because another project other than BitTensor that someone brought up to me is Venice. (00:38:09) The concept of open source training and orchestration (00:38:14) is a hugely disruptive idea, which is the complete orthogonal attack vector to this idea that you have to raise 10s and 10s of billions of dollars to train your models. (00:38:25) Because if the capital markets run out of 10 and $20 billion checks to give people, the only solution is to be totally distributed. (00:38:34) So I tend to agree with you, Jason, that there is going to be, at some point, a very successful open source project for pre-training. (00:38:43) Absolutely will there never, ever be an open source way where a real company that has any skin in the game says, here, guys, re-engineer my code base as an open source project. (00:38:53) Never going to happen. (00:38:54) Yeah, I think the coding tools will. (00:38:55) And if you look at the history of open source, Brad, you actually, I think, had a lot of bets in this space. (00:39:00) Linux, Kubernetes, Apache, Postgres, like Terraform, like these open source projects are deep inside of enterprises, deep. (00:39:08) And we're sitting here 15, 20 years ago, the same argument was made. (00:39:12) Nobody will ever adopt these inside the enterprise. (00:39:14) You got to go with Oracle, whatever. (00:39:16) And fair enough, many people do. (00:39:18) But I think this $29 Ridges subscription to do this versus 200, it's starting to take hold inside of startups. (00:39:28) And that's where I always look at the tip of the spear. (00:39:30) Startups love to use open source products. (00:39:33) I think this could be the next big thing, but listen, I invest in things that have a 90% chance of going to zero. (00:39:41) Do your own research, no crying in the casino. (00:39:44) Can I just make a final few points? (00:39:46) So just quickly. (00:39:47) So #1 is, with respect to this market for code, or code tokens, whatever you want to call it, might be 5% today, meaning 5% of the code's AI generated versus human generated. (00:39:59) I think it's going to 95%.