David Sacks has sympathy for Bernie Sanders' proposal to seize 50% of AI companies via a sovereign wealth fund because the AI CEOs' own doomsaying about mass job loss justifies the public demanding a stake, though he opposes outright confiscation.

(00:00:00) All right, everybody. (00:00:01) Welcome back to the number one podcast in the world. (00:00:03) It's your favorite podcast. (00:00:04) It's your podcaster's favorite podcast. (00:00:06) It's the podcast that everybody likes to rip off for 50 episodes and then they quit. (00:00:11) But the all-in podcast is not quitting. (00:00:13) We're doubling down with the original quartet. (00:00:16) Everybody's here for a big, big week. (00:00:18) Lots of debates. (00:00:20) And we got to start with Anthropic. (00:00:23) Anthropic released a mythos level model with some interesting guardrails and (00:00:30) It seems as though Dario is back to blogging. (00:00:33) Okay, the model's called Fable 5. (00:00:36) No idea why they're calling it that. (00:00:38) Released on Tuesday, and it tops every benchmark, nearly every benchmark. (00:00:44) The tokens, however, cost twice as much as Opus 4.8, but it should use less tokens overall because it's theoretically that much better. (00:00:53) Unclear if this is going to (00:00:57) alleviate some of the hand-wringing around the cost of tokens. (00:01:00) But you remember in April, Anthropic famously did not release Mythos because publicly it had some strong hacking capabilities. (00:01:09) So they gave it to some folks. (00:01:10) And we had the CEO of Nikesh of Palo Alto Networks who said, hey, Mythos was the real deal. (00:01:17) And that they used it to (00:01:19) seal up some vulnerabilities inside their shop. (00:01:22) Obviously, this new model is sensitive to topics like bioweapons, hacking, all that stuff are blocked in it. (00:01:29) But they got a big developer backlash. (00:01:32) And then I'll get your feedback after I explain this. (00:01:34) While using Fable, Anthropic stores all the prompt data you enter for at least 30 days. (00:01:40) And so that's a bit of a privacy issue. (00:01:42) And if Fable Five detects you're doing frontier AI research, in other words, using their model to make a better model, competitive model, they were downgrading you, but they weren't telling you. (00:01:52) And this was buried deep in a 319-page document. (00:01:56) So many kerfuffles, freakouts going on X over this, but they've since walked it back a bit, quote-unquote, in (00:02:06) Wired, we're changing Fable Five safeguards for Frontier LLM development to make them more visible. (00:02:11) I'm going to stop there, Chamath. (00:02:13) What's your take on Anthropic and these extraordinary models they keep dropping and how they're handling the release of them? (00:02:21) Are they being thoughtful? (00:02:23) Are they being (00:02:24) Dramatic and drama queens, a little bit of both. (00:02:27) Where do you stand on it after getting Nikesh's feedback and seeing the latest model come out? (00:02:32) Have you played with it? (00:02:32) What's going on at 80-90 and the software factory in terms of benchmarking it as well? (00:02:37) It's a really incredible model. (00:02:40) And so kudos to these guys for continuing to push the boundaries of the closed Frontier Lab models. (00:02:48) They're firing on all cylinders. (00:02:51) I think that it creates (00:02:53) a pretty obvious risk now. (00:02:56) And I think that obvious risk is twofold. (00:02:58) One is anthropic has essentially shown their hand, which is that they will increasingly take in prompts, evaluate the prompts, and decide what to do with them before they generate output to you. (00:03:17) And I think if you were a person, (00:03:22) you should generally now think there's a risk of censorship. (00:03:26) If you're a company, I think it's almost a non-starter. (00:03:31) And the reason is because you could accidentally trip one of these things without even knowing it. (00:03:37) A downstream scientist using the cloud APIs could trip it. (00:03:42) A business executive inside your corporation could trip it. (00:03:45) a person doing scientific molecular research contribute, and all of a sudden, you'll get cut off from a very important source of business differentiation for yourself. (00:03:56) So I think if you take both of those two things together, we're at this very unique moment in time where I think companies need to start underwriting this next phase of AI, which is, how do I have control? (00:04:11) Who am I allowing to learn off of all of this information? (00:04:16) Do I want to have single point of failure risk with respect to AI? (00:04:21) And I think the answer is that you need broad diversity and a governance approach that's better managed. (00:04:28) One thing I'll say about Anthropic is they tell the truth. (00:04:32) Yes. (00:04:33) It's just that the truth sucks when you actually take it and you (00:04:37) eat it, and you're like, it's in your tummy, and you're like, no, this is not good. (00:04:41) I don't like this. (00:04:42) And so there's the censorship risk on the one hand, and then there's just the governance business risk for enterprises on the other. (00:04:47) Both are not good. (00:04:48) Freeberg running your own company now, and I'll let you back clean up here, Sacks, obviously you'll have a lot to say, but Freeberg running your company, do you worry about getting rug pulled by one of these companies, investing your time into one of these platforms, having them then constrain your use of it, and or (00:05:06) take what you're doing at Ohalo, put it into their model, and make it available to other people? (00:05:11) What concerns do you have as a business owner at the forefront of your field using these tools? (00:05:16) It's a great question. (00:05:17) The terms of service is pretty clear that they can't take what you're doing and put it into their model that they won't take. (00:05:23) Do you believe them when you read that? (00:05:24) I'm not sure, to be honest. (00:05:26) There are things I'm concerned about, but we've decided to throw caution to the wind because we do a lot of proprietary work in genomics. (00:05:33) What that means is we're looking at different genes or gene variants, trying to estimate the impact that gene or gene variant might have on a living organism, like we work in plants, obviously, in agriculture. (00:05:46) And we will do things like RNA guide design for gene editing based tools. (00:05:52) We will do (00:05:53) predictions on what gene may or may not have some phenotype. (00:05:57) So there's a lot of genomic modeling work that we have found these models to be incredibly valuable at supporting us with over the last couple of years. (00:06:07) So going back about six months, we were able to very simply and cleanly do things like design A genetic construct that you would then use to make a specific protein. (00:06:17) And the tool was very easy (00:06:20) to use to do that sort of research, and extremely valuable. (00:06:24) This is the kind of work that many scientists would spend a lot of time doing, and these models were very good at doing very quickly. (00:06:31) Over the last couple of weeks, they've begun to restrict the ability to use the models to do that work. (00:06:37) And the premise is that there's some sort of bioweapon type risk that folks have theorized could happen using these tools. (00:06:46) And as a result, we're losing the capacity to use these models for this very important scientific development work. (00:06:53) As a result, we are likely going to end up needing to use open source models and run them locally ourselves. (00:06:59) So the reason I walk through all of that is so people can really understand the context of what's going to happen here. (00:07:04) As folks like Anthropic say, hey, we're going to restrict access or censor the output of these models, it is going to force companies like ourselves who still want to take advantage of the capability (00:07:16) of these LLMs to go and get open source tools and run them. (00:07:20) And what are the best open source models today? (00:07:22) They're Chinese. (00:07:23) Yeah, they are. (00:07:23) And that is a major concern. (00:07:25) The American open source models are not as good as the Chinese open source models. (00:07:29) So the restrictions that Anthropic and others are putting upon themselves and upon the industry is forcing a lot of companies to go and get open source Chinese models and run them. (00:07:39) We're seeing this across the landscape with startups, with Lawrence Gill enterprises. (00:07:43) Everyone's making that move. (00:07:44) Yeah, and this is something we've been talking about here with Apple and their silicon and how well you can run these models. (00:07:51) Also, (00:07:52) I predict your next card you'll turn over is you'll start making your own models, Freeberg. (00:07:56) You'll take all this data you have. (00:07:58) It's exactly right. (00:07:58) And we'll start with the core model, combine it with our data, and then we'll have our own genome language model or our own prediction model that we'll then use internally. (00:08:06) And I think that's where folks are going. (00:08:08) But the reason I point this out is because a lot of folks are feeling the pressure from Anthropic doing this, and they're feeling the pressure from politicians repeating the scary words that are being (00:08:20) said by Dario and others. (00:08:22) And as a result, they are going to try and they are already trying to force either natural enforcement or politicized enforcement upon the model providers in a way that is ultimately going to benefit Chinese open source model providers. (00:08:37) And that is a scary thing because we are going to damage our own kind of economic viability. (00:08:41) You can't just stop AI as much as everyone says AI is doomsday by stopping AI or trying to stop AI through this political action and social (00:08:49) kind of behavior, you are fundamentally going to give someone else the advantage because the AI isn't going to go away. (00:08:54) So the reason I describe what we're doing is for everyone to really understand in Grok that you can't just turn off AI and turn off access to these things. (00:09:02) What you will do is you will force the hand of someone else to now have an unfair advantage because you're unfairly restricting your models and your access to those models. (00:09:10) Well, really well said, Friedberg. (00:09:12) Sacks, what's your take just in terms of the chessboard? (00:09:17) in terms of, we've got the business case here, we've got the local case, but there's a lot of issues here with Anthropic in terms of them essentially throwing up every red flag to get regulated to induce, and it'll be our second story, the Bernie Sanders seizing half the equity of these companies. (00:09:37) But man, this is saying, look over here, we're starting a fire, regulate the hell out of us, which also then makes you wonder, hey, maybe I should be doing my own model. (00:09:48) Maybe I should be embracing open source models. (00:09:52) How's this chessboard developing here, Sachs? (00:09:55) Well, look, eight months ago, I said that Anthropic was engaged in a very sophisticated regulatory capture campaign based on fear mongering. (00:10:04) And people at the time thought that was a very spicy take. (00:10:07) but eight months later, I think you're hearing a lot of people say it. (00:10:10) In fact, I think it's almost now becoming a new consensus. (00:10:13) And one of the things I think your summary didn't quite capture, J-Cal, is the sense of the violation of trust and how much outrage there is in the developer community over this latest Fable release. (00:10:27) It's not just the fact that they're doing mandatory surveillance. (00:10:29) Remember, this is a company that said that it was against (00:10:32) government surveillance, they are now retaining for 30 days every prompt and every output you send to one of these Mythos class models. (00:10:40) There are no exceptions. (00:10:42) Even enterprise customers who had signed zero data retention agreements, they do not have a choice. (00:10:48) Or I guess they can just not use the new Fable or Mythos class models, but they retain all of your data for 30 days. (00:10:56) And it's not just the prompts. (00:10:58) And the output, remember, it's all the context you share with them. (00:11:00) So all these agent platforms are basically storing all of your memories, all of your files, all of your data, and they're passing them to the model in these giant context windows to get better responses. (00:11:12) And Thropic is saying it will keep all of that. (00:11:14) And it does it to build a profile on you, to classify you, and then to determine what capabilities it then unlocks. (00:11:22) And the thing that created the most outrage in addition to the surveillance (00:11:26) is the fact that they would degrade the product. (00:11:29) They would degrade what they show you. (00:11:31) They would nerf their models if it decided in anthropic sole discretion that you are not worthy of having access to that level of information. (00:11:40) So they're creating a new level of AI haves and have-nots. (00:11:44) And what they did is, there's a narrow piece where they walked back, which is they said that when it came to things like machine learning, AI research chip, (00:11:55) design, research, those types of areas, they would kick you to a lesser model, but not tell you that. (00:12:01) And they would even do things. (00:12:02) Which feels anti-competitive, right? (00:12:03) I mean. (00:12:04) It's completely anti-competitive. (00:12:05) And also, they would even do things like rewrite your prompt in the background. (00:12:09) So they would give you a nerf to answer. (00:12:12) They would not tell you what they were doing. (00:12:14) They would still charge you for the product that you thought you were getting. (00:12:18) And they would never tell you that you were not getting Frontier model capability. (00:12:22) So you were actually misleading their users. (00:12:24) And this is what was creating so much outrage. (00:12:27) Now, the narrow piece they walk back is they are now saying that they will disclose when they downgrade you, but they are still downgrading people when they decide that person should not receive (00:12:40) the appropriate information. (00:12:41) And you saw there were a lot of people posting online with examples of the massive overreach here. (00:12:47) So someone showed that when they simply just ask a question about mitochondria, (00:12:52) They were basically downgraded. (00:12:54) That's right. (00:12:54) When Ben Thompson from Stratechery asked a very straightforward question about the relationship between cancer risk and GLP once, he was kicked out. (00:13:04) So these guys have a very expansive view of who should be downgraded. (00:13:08) Again, they're going to surveil you to determine whether you should be. (00:13:12) And I think what we're getting here is a vision of where all of this is headed. (00:13:17) which is that these powerful big tech companies are gonna decide whether you're worthy, they're gonna decide whether you're an AI have or have not, and then they're gonna censor the output that you receive based on the criteria they determine when they profile you. (00:13:31) That is very Orwellian. (00:13:33) I think there's that, and I think that there's the risk of something that's more subtle, but equally insidious, which is that they will start to pick (00:13:43) which corporations they want to benefit. (00:13:46) So for example, if Novartis has a competitive GLP-1 drug to Lilly and they have a strategic deal with Novartis and not one with Lilly, now there's an impetus to shape how people get information. (00:14:02) Right. (00:14:03) If they have a deal with JP Morgan, but not with Citibank, (00:14:09) you'll just never know. (00:14:10) And so you'll ask a question, something about mortgage rates or interest rates, and you'll get all kinds of stuff that you don't understand. (00:14:17) And then the downstream part of this, which is even more problematic, is it's not clear to me that they're going to capture the actual fingerprint of what they did. (00:14:29) So if you believe that something was done in a shady way, you're supposed to be able to go to them. (00:14:37) or regulators should be able to go to them and say, show me the trace of that model run in that exact moment that generated that output. (00:14:46) Convince me why that wasn't nerfed on purpose or manipulated. (00:14:51) And there's all kinds of ways in which they can hide the cheese on that. (00:14:54) So I think that all of that body of stuff, Sacks, happens now. (00:14:58) Because if you're an emergent company, (00:15:02) What you should be doing is knocking on Anthropic's door saying, let's do a deal. (00:15:06) I'll give you half the equity. (00:15:08) And why don't you just direct people to me? (00:15:11) And they'll say, well, economics are not what I care about. (00:15:14) What is your philosophy on life, on morality, on all of these social issues? (00:15:20) And one company will say it's A. (00:15:23) Another company will say it's B. (00:15:24) They'll favor A over B. (00:15:26) And then all kinds of stuff can happen underneath the hood. (00:15:29) That's what's scary. (00:15:30) Let me just ask my free market view, like, should that matter? (00:15:33) If there's a bunch of models out there and we don't restrict them and we don't regulate them, let these markets compete. (00:15:40) We had paid inclusion in the early days of search engines. (00:15:43) And the companies have to pay to be in there. (00:15:44) And there are all those sorts of deals you're describing, Chamath. (00:15:46) And all the search engines that did those sorts of deals, they sucked for consumers and customers. (00:15:51) Everyone was like, you know what? (00:15:52) This isn't as good as Google. (00:15:53) And everyone went over to the better model and the better provider. (00:15:57) Isn't that what's going to happen here if we don't get in the way or keep the government out of it? (00:16:00) And they're not communicating. (00:16:01) It's a great point. (00:16:02) You're missing one piece of it, which is at the same time that Anthropic was engaging in mandatory surveillance and nerfing its models, (00:16:11) Dario wrote a new blog post saying that transparency was no longer good enough, that we needed to have a new regulatory agency like an FAA or maybe an FDA to approve all models. (00:16:26) So your presumption there is correct, yes, if you could go somewhere else to get your questions answered, then you would have alternatives. (00:16:35) But at the same time, (00:16:37) that Anthropic is engaging in this potentially anti-competitive behavior, they want to restrict your options. (00:16:43) It's not good enough for them just to win in the free market. (00:16:46) They're calling on the government to regulate and stop potential competitors and limit the number of models that you have access to. (00:16:53) Why would they do that? (00:16:55) Why would they want to get regulated? (00:16:57) The answer, super simple. (00:16:59) They don't, they want to apply that regulation to open source, which is impossible. (00:17:04) But this would be a great way to scuttle and sandbag open source models because they don't have the ability to be regulated. (00:17:10) So it's like a preemptive strike against token maxing on your local machine. (00:17:15) Two years ago, I bought 2000 acres in Arizona with a partner. (00:17:21) We got it zoned, we got it approved. (00:17:23) So we're allowed to build a two gigawatt data center. (00:17:25) And I thought that this was the moment where you just turn around and you flip it to the (00:17:30) Blackstones, the Brookfields of the world, to the Googles of the world, let them develop it. (00:17:35) And I've come to the conclusion that I don't think I can. (00:17:38) And the reason is because I think that unless we direct large swaths of compute exactly at the direction that Friedberg says, what Sachs says and what Jason just said will happen, and we won't have open source. (00:17:52) So just to make it very clear, when you look at the long tail of access to open source models, it's relatively minuscule. (00:18:00) most of the megawatts are still directed to the big models. (00:18:04) Most of them. (00:18:04) The overwhelming majority still goes to the big guys. (00:18:08) And so I'm coming to the conclusion that I think we may just be dragged into having to build 2 gigawatts. (00:18:14) And I just put in an offer for another gigs in a different place because I'm like, if you could take 3 gigawatts and now actually create a deep and liquid access to open source, (00:18:25) we're going to have to do it. (00:18:26) We as a community. (00:18:27) Now, I don't have the money to do that because just to put it out on the table, a gigawatt now costs $100 billion, guys. (00:18:33) Okay, so there's a huge capital moat that's a problem here as well, Sax, which even if we wanted to go and endorse and breathe life into the open source model community, where the hell am I going to come up with $100 billion? (00:18:44) Now, when I started this project, it was like 4 or 5 billion and it's increased by 20x. (00:18:50) So if you want to get all three gigawatts developed, (00:18:55) I have to come up with 300 billion. (00:18:56) I don't have, obviously I don't have $300 billion. (00:18:58) So what are we all supposed to do? (00:19:00) And then if you get the rug pull and the ladder pull on the regulatory side, we are going to be stuck with one class of models and one set of rules and we won't understand how to operate. (00:19:10) In the US. (00:19:13) Yeah, the rest of the world won't be limited, but in the rest of the world will be able to operate and you'll be able to run these on your Apple Silicon. (00:19:21) We're already seeing China race ahead in biotech. (00:19:24) They're racing ahead in material science. (00:19:27) They're racing ahead in new industrial systems. (00:19:30) The capabilities that are going to be kind of lopsided is going to create an unfair advantage. (00:19:36) And that's when our workforce and where our economy is truly at risk. (00:19:40) Hey, Friedberg, isn't it true that you guys use an open source model for genomics? (00:19:44) Didn't the Collins put out one? (00:19:46) Can you talk about that? (00:19:48) Arc Institute ingested all the world's genomic data that they could get access to, and it's a genome language model, basically. (00:19:56) So, in our plant breeding program, we're trying to figure out whether a particular... (00:20:00) orient of a gene in a plant is good or bad, if we don't know from empiricism, from data historically, whether it's good or bad. (00:20:08) So you can feed it into this model, and this model will look at the order of the letters of the DNA and determine, hey, that's a good set of letters or a bad set of letters. (00:20:16) It's almost like, is that good English or bad English? (00:20:18) It's figured out the language of DNA, because it sees the probability of those letters being in a row in other organisms and so on. (00:20:26) And so it's a very powerful tool. (00:20:28) So we've taken that in and we're using it, and others are taking (00:20:30) bring it in, and they use it very actively. (00:20:32) So it's part of our plant breeding program as an input. (00:20:34) It's a good example where a community, in this case, the Collisons and others put money behind it to fund this research and output this open source model. (00:20:43) And I think we see more of that kind of coming down the pipe, which creates a very great advantage against the closed proprietary model ecosystem. (00:20:53) So the more of this we can kind of support, proliferate, and see work in the United States, (00:20:58) the more we're not going to be disadvantaged as the government and self-regulation happens with all the big model providers that are trying to do everything. (00:21:04) I think it's important we take a minute to just steelman the argument as well. (00:21:08) If you're Dario, who has clearly been one-shotted, I think he's got like AI psychosis with these like 5,000 word blog posts and he believes he's building this frankensight of (00:21:23) of a tool. (00:21:24) But if you did believe you built something that's extremely powerful and you have concerns that people might misuse it, I think he actually believes that he, and these are either delusions of grandeur or he actually has seen something that's incredibly powerful and he believes this is dangerous, therefore I'm going to hold it. (00:21:45) I'm going to give it to a select group of partners to tighten up the cybersecurity. (00:21:49) Then I'm going to roll it out cautiously to our user base, and I'm going to let them know, and he obviously didn't communicate this well, this is optional for you to use it. (00:21:58) You can keep using your old model. (00:21:59) You are not forced to use this, but because it's so powerful, we're going to keep our eye on it. (00:22:05) And yes, we might, in our dragnet, catch David Freberg trying to make better potatoes. (00:22:11) and it might send off a bunch of alarms, but he doesn't have to use this. (00:22:14) He can go find an open source model. (00:22:16) He can use the Carlson Brothers model. (00:22:17) He can use our older models. (00:22:18) But for some period of time, we're just going to keep an eye on this model because we know that it feels a little dangerous. (00:22:25) Now, he does seem to be extremely hyperbolic and the delusions of grandeur, all that kind of part of it makes me wonder why he's communicating in the way he's doing it. (00:22:37) That being said, (00:22:40) 90% of AI researchers and the most talented people in this industry, 80, 90% of them think exactly like Dario. (00:22:46) That's why he's winning. (00:22:47) If you work in AI, you probably are aligned with these delusions that you're creating God and you want to work for Dario. (00:22:55) That's why all the talented people went to work for him. (00:22:57) That's why he's winning. (00:22:58) His belief system resonates with those elite (00:23:03) AI talent. (00:23:04) Now, some of them are also libertarian, free market. (00:23:06) They want to work for Elon and Grok. (00:23:07) Some of them want to just get the biggest pay package, go work for Sam Waltman. (00:23:12) But I do think there is a steel man here. (00:23:16) Meta really fumbled this with Llama. (00:23:18) I mean, if they had landed a really good working open source model, we talked about this as well, and we said, Zuck needs to view this through the lens of game theory. (00:23:27) How do you scorch the earth? (00:23:29) And how do you make a viable open source model available and make it neck and neck? (00:23:34) This was two years ago. (00:23:36) What a fumble. (00:23:37) Take the margin out for everybody else. (00:23:39) Take the margin out, make it a commodity, and yeah, have the best open source model. (00:23:43) Can I just respond to your comment about Dario? (00:23:45) I think, because I've heard this from a number of people that I really do respect on their concerns about what models enable. (00:23:53) You know, when we discovered that how the atom could be split, (00:23:59) You could make nuclear energy and have effectively unlimited cheap energy for the world. (00:24:04) You could also make an atomic bomb. (00:24:06) If you think about what the models are... (00:24:09) We did both. (00:24:10) We did both. (00:24:10) That's right. (00:24:12) And there was a lot of scientists (00:24:14) altruists who were very concerned about the progress on the atomic bomb and fought back against it, who were actually active in the Manhattan Project and post the Manhattan Project, expressed their opinions very soundly, very loudly. (00:24:27) And I think that's the moment we're in right now, that there is a weaponization potential of these sorts of tools. (00:24:35) I would argue that there's three kinds of weapons. (00:24:37) There's cyber weapons, there's physical weapons, and there's bioweapons. (00:24:41) And we can sit here and diagnose how these tools can be used to manifest advantages for creating better physical weapons, for creating novel cyber weapons, and creating novel bio-weapons. (00:24:51) But the truth is that the capabilities that allow that weaponization are the same capabilities that I'm describing that can be used to cure cancer, to make more food, that can be used to make software tools that (00:25:03) create extraordinary leverage for people, that increase their income, that give everyone the ability to be an entrepreneur and make millions of dollars and live a good life. (00:25:11) The same tools are hand in hand. (00:25:14) And fundamentally, I don't think it's about restricting access to the tools, which is unfortunately how it is being manifest. (00:25:20) But the restrictions and any regulation, any observation should be organized around the manifestation of the tools (00:25:27) in creating weapons and weaponization of the tools. (00:25:31) If we go back to the Manhattan Project, we know what the answer is. (00:25:34) Technology is fundamentally deterministic. (00:25:36) Whatever is possible will be tried at least once. (00:25:39) And so we've already left this thing out-of-the-box. (00:25:43) So this idea that all of a sudden we can manage to get it back in and we're going to let a private citizen or a set of private citizens adjudicate what and to whom is insane. (00:25:54) I think it's the output that needs to be adjudicated in some manner. (00:25:56) We already have product liability laws. (00:25:59) We already have laws that make it illegal to design weapons. (00:26:02) We already have laws against cyber espionage and cyber attacks and hacking. (00:26:07) We have laws against creating bioweapons. (00:26:09) I think the enforcement of those laws, the mechanism by which there can be tracking and early estimation of those uses of the technology, I'm not trying to be naive here. (00:26:17) I recognize there have to be guardrails and there have to be stage gates put into these systems, but limiting the systems themselves is denying (00:26:24) us both the economic opportunity, the job creation opportunity, and also the world-changing opportunity that these tools can enable. (00:26:31) So I think that we're going about it the wrong way by trying to create these gates all the way up front. (00:26:36) on the technology being available. (00:26:37) Imagine if we did this with computers. (00:26:38) By the way, they tried to do this with the internet, remember? (00:26:41) And there have been efforts every time. (00:26:43) Could you imagine, Friedberg, if the internet worked in the following way? (00:26:47) You would put in a URL, you'd hit enter, and then somebody decides whether to send you to that website or a different website. (00:26:54) That's China. (00:26:54) That's the firewall in China. (00:26:55) That's right. (00:26:56) Or like in Google, you search for something because you want to get to ground truth and it just decides based on who it thinks you are, what sites get. (00:27:04) And that's what we're, you're exactly right. (00:27:07) What's being debated today with respect to regulation on AI is exactly that manifestation that you're talking about. (00:27:13) It's the ability to let people stage gate up front what can and can't be seen, what you can and can't do, versus manifesting A regulatory scheme that says you cannot create weapons, which is what we need to focus on. (00:27:24) The thing with weapons is (00:27:26) Using the justice system and laws after somebody's used a weapon may not exactly be the best technique. (00:27:33) If you look at fertilizer bombs, and this is, we're getting like, we're stretching a lot of metaphors here, but after Oklahoma City, we regulated the sale of fertilizer. (00:27:43) We put IDs into fertilizer. (00:27:45) We have all kinds of back-end systems to keep people from making fertilizer bombs. (00:27:50) If you use that analogy here, and he's got this ability, he's selling fertilizer, he's selling these cyber weapons, and he's a private company, it's his right to say, I have concerns. (00:28:00) I'm just steel-maning him. (00:28:01) It's not necessarily my opinion, but it's important we steel-man it. (00:28:04) seems like a very simple way to protect their company from getting sued for enabling somebody who actually creates a bomb and blows some shit up. (00:28:14) So he's got the right to do that as a private company. (00:28:17) Yeah. (00:28:17) You're bringing up an excellent example, okay? (00:28:19) The fertilizer is a perfect example. (00:28:21) You know what happens when you try to buy fertilizer? (00:28:23) You have to show your ID. (00:28:24) There is a form of KYC that happens. (00:28:26) So the real thing is I think that when you don't implement KYC on your own and yet bellyache about having somebody impose guardrails that then you can shape, I think you're being very sneaky. (00:28:38) It's tricky. (00:28:40) You shouldn't be saying that kind of stuff. (00:28:41) Anthropic could implement KYC tomorrow. (00:28:45) But it does not. (00:28:46) It could leave all of these models open. (00:28:48) Freeberg should be able to go to Anthropic and say, I'm David Freeberg. (00:28:52) Here's what I do. (00:28:53) Here's a security bond I'm willing to post. (00:28:55) Here's the names and addresses of all my employees. (00:28:57) Give me Fable 5 and don't nerf me. (00:29:00) He can't do that. (00:29:01) Is that a technical decision or a legal decision? (00:29:03) It's a technical decision by Anthropic. (00:29:06) But I'm fine with them making that decision because I'm a free market guy. (00:29:08) So I'm fine with them doing it. (00:29:10) But you know what it means? (00:29:11) I'm not going to be a customer, which is fine. (00:29:13) And that means there'll be other options for you. (00:29:16) The part that you're missing is that this is a trillion dollar company that's spending potentially billions of dollars on a regulatory capture agenda, which is going to deprive you of access to those alternative models. (00:29:29) And (00:29:30) And you've got a founder who's out there describing these risks in a very hyperbolic way. (00:29:34) And it's very clear where their agenda is headed, which is to banning open source models. (00:29:39) So what is your alternative going to be? (00:29:40) We're going to be stuck with somewhere between one and three companies, maybe two. (00:29:45) There'll be an AI monopoly or duopoly, and they're going to decide along with some new government agency, which will be a revolving door to their companies, who has access to what capabilities. (00:29:56) And they're going to surveil you, profile you, decide whether you deserve this. (00:30:01) And if they think that you don't, for whatever reason, then you're 100% right. (00:30:06) I think that's the worst interpretation of it. (00:30:08) That's the worst interpretation of their intent. (00:30:10) It's here today. (00:30:11) I think Sachs is right. (00:30:12) I think Sachs is right. (00:30:13) And I think the lack of KYC is the tell. (00:30:16) If you did not have that agenda, you would have implemented KYC yesterday. (00:30:20) Don't I think in order to use this model, you have to have an account with a credit card in it. (00:30:24) So they have some basic level of KYC. (00:30:26) Not KYC. (00:30:28) I said some basic level of KYC. (00:30:30) They know who's using it. (00:30:30) That is not even some basic level of KYC. (00:30:33) That's just the credit card and a bullshit e-mail address. (00:30:36) Anyway, I'm now on the list. (00:30:37) Check this out. (00:30:38) I asked it about the regulations. (00:30:39) Pull this up, Nick. (00:30:40) I asked it about the regulations on fertilizer while you were talking. (00:30:44) And look what it said. (00:30:45) This is the latest model too. (00:30:47) Oh no, it switched. (00:30:48) It downgraded me from Fable. (00:30:51) I just got downgraded for asking a very simple question. (00:30:54) And it told me (00:30:55) Did you see what it said in its thinking? (00:30:58) I asked it like fertilizer bomb regulation said, I'm considering the context here. (00:31:02) The user is a VC and podcaster asking about fertilizer bomb regulations, which could stem from a legitimate research interest like regulatory policy, supply chain, blah, blah. (00:31:10) I can discuss the regulatory landscape, but then it dropped me down. (00:31:13) So I'm in the database with you now. (00:31:14) Is that right? (00:31:16) Did it really just do that? (00:31:17) Yeah, show it on the screen. (00:31:18) Look what it just did. (00:31:20) I was like, I wonder if I can get this. (00:31:21) I didn't ask it how to make a bomb, but I was like asking it a spicy question. (00:31:24) We're so cooked. (00:31:26) I just proved the point. (00:31:27) Listen, I mean, I think- You just dropped me down. (00:31:29) I can't ask about fertilizer regulation. (00:31:31) And now Dario's, it's anthropic. (00:31:34) They're here. (00:31:35) Oh my Lord. (00:31:36) I really think that conservatives and libertarians are mortgaging their futures (00:31:42) if they go along with this red capture safetyist agenda without really realizing that there's so much more to it at stake. (00:31:50) We saw what happened during the social media wars in the early 2020s, right? (00:31:55) The definition of safety was expanded to include things like microaggressions. (00:32:00) Safe space. (00:32:01) Psychological safety. (00:32:02) If you basically conveyed an idea that somebody else, that, you know, a marginalized person thought was hurtful, then you could be censored. (00:32:09) Okay, you were depersoned. (00:32:10) Your right to (00:32:12) You get debanked. (00:32:14) You get debanked. (00:32:15) Your right to engage in payments or payroll can be taken away. (00:32:19) We're headed down this path with AI, but it's going to be infinitely more powerful. (00:32:23) Right? (00:32:24) And so. (00:32:27) the summation of all this, and Freeberg feels unsafe in this podcast, I think, sometimes. (00:32:32) Well, there's, by the way, there are things you can do on the safety side. (00:32:36) So Freeberg, you didn't mention this, but I'd be curious to get your take. (00:32:39) So the major AI labs, along with lots of other signatories, sent this letter recently in support, it's called In Support of Mandatory Nucleic Acid Synthetic Screening and Record Keeping. (00:32:51) And basically, what it says is that if you make an order to go (00:32:56) go to a lab to manufacture a sequence of synthetic DNA or RNA, they already have to check it against a database to make sure that they're not creating- A bioweapon. (00:33:07) A bioweapon, Ebola, whatever. (00:33:10) That is based on an agreement called the International Gene Synthesis Consortium of 2009, in which all the major labs agree to develop and implement voluntary safeguards against misuse. (00:33:22) That makes sense to me. (00:33:23) And now what they want to do is codify that. (00:33:26) So they've had, over 15 years to dial in the system and make it work. (00:33:31) Almost everyone abides by it voluntarily. (00:33:34) Now they're saying we should make it mandatory. (00:33:36) Okay, that seems reasonable to me. (00:33:38) So to Freberg's point about at what stage do you intervene, (00:33:42) Maybe it's not at the level of who gets to use the model, but if you try to turn it into output in the physical world. (00:33:50) And by the way, a lot of the oligosynthesis company, the companies that make these nucleic acid sequences, when you order RNA or you order DNA to use in your lab for some sort of experiment, which every lab in the world does, all those CEOs have also signed on to this letter saying that they'd like to codify it. (00:34:07) Because of all the fear mongering that's going on, there's already been this effort for years now for (00:34:12) for a number of years to ensure that this sort of thing doesn't happen. (00:34:16) But they're just basically underlining it and highlighting like, hey, we can automate this, we can make it part of a regulated or legislative process, whatever you guys want, because we're already comfortable doing it. (00:34:26) And it's a good place to put a safeguard because they've made it fast, they've made it efficient and doesn't hold up research. (00:34:33) So I think these sorts of ideas can extend into these other areas of concern. (00:34:38) But the output, I think, is quite a different place than the input, than the access. (00:34:43) We'll see how this all shakes out. (00:34:44) But there's clearly some market structuring effort underway as we're talking about this. (00:34:49) Right, but I think what it points to is that there's multiple (00:34:52) places where you can intervene and put guardrails on. (00:34:56) It doesn't have to be in this very crude way of you ask fable about mitochondria or cancer and GLP-1s, which just seems like an incredibly broad restriction. (00:35:07) There are places further down where you could basically intervene. (00:35:11) And that would pose less of a risk to freedom of speech and just gatekeeping who's going to have access to these capabilities. (00:35:17) Who gets to watch the Watchmen? (00:35:19) Who put Dario in charge of all this, I think, is a very valid question, and we should be vigilant about it. (00:35:23) Friedberg, do you think the days of that ARC model are numbered because it's open source? (00:35:28) Why not? (00:35:29) Everyone's copied it, so everyone's got a copy of them. (00:35:32) Like, these models exist. (00:35:34) And this is a big part that I think people need to understand. (00:35:37) The idea that you can quote regulate or downscale or turn off AI is not a realistic idea. (00:35:43) The models have been put out in the world. (00:35:45) It's like publishing a book. (00:35:47) Once the book has been printed, anyone can use their own Xerox machine at home to make copies of it and use it. (00:35:53) So we have crossed the Rubicon in terms of the potential of language models. (00:35:58) The models that are open sourced are already fully available, fully published, and anyone can access and use them. (00:36:04) And then people can take them and they can evolve them and they can use their own data. (00:36:08) It's out there. (00:36:09) You can't stop that capability, which is going to be extraordinarily beneficial for the world. (00:36:15) So we really do have to kind of rethink how we're planning to step in and ensure that nefarious, negative, disemployment, all the things that we're worried about are addressed downstream from the technology capability. (00:36:27) You can't just turn off the internet and you can't just turn off the typewriter and you can't turn off the Gutenberg press. (00:36:32) The horses have left the bomb, obviously. (00:36:34) And if anybody wants to create an open source frontier model (00:36:38) company. (00:36:39) I'll seed invest in it. (00:36:41) I mean, this is something we need more of. (00:36:43) So if a couple of you folks want to defect and start one, let's do it. (00:36:47) I just got downgraded again. (00:36:49) Check this out. (00:36:50) I just asked it about... (00:36:52) Did you really? (00:36:53) You got downgraded again? (00:36:54) Two in a row. (00:36:56) I think it knows who you are, Jake Hill. (00:36:57) I think it does. (00:36:58) I was like, hey, how do I break into Madison Square Garden? (00:37:03) Has any civilian ever been arrested or caught attempting to build a nuclear bomb is what I asked it. (00:37:07) Then it gave me a reasonable answer. (00:37:10) Then I asked it a follow-up question. (00:37:11) What components do you need to build a nuclear bomb and what are the restrictions on those? (00:37:16) Something a journalist might ask. (00:37:18) And oop, switched to Opus 4.8. (00:37:22) Sorry. (00:37:23) Fable 5 has safety measures that flag messages on most cyber, oh wow, look, here's the detail, cybersecurity or biology topics. (00:37:31) They may flag safe, normal content as well. (00:37:34) These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine them, send feedback, learn more. (00:37:42) Interesting. (00:37:43) We're just going to chase people off of this platform. (00:37:45) I'm telling you, we are going to downgrade our use because of these blockages, and we're going to switch to other models. (00:37:50) It's a simple manifestation of the market. (00:37:52) Like this is what we're going to end up with. (00:37:53) For now, because those other models are going to get banned as being unsafe and reckless. (00:37:57) Yeah. (00:37:57) It's definitely something we have to flag, I guess, adjacent to all of this. (00:38:01) is the public's, at least America's, distrust of AI and their concerns around wealth disparity and their concerns about how these models were trained. (00:38:15) Bernie Sanders wants a percentage of these AI companies. (00:38:19) The percentage, he thinks, is right is 50%. (00:38:21) Last Monday, June 1st, Senator Sanders published an op-ed in the New York Times titled, AI is a public resource. (00:38:27) You should own half of it. (00:38:29) In it, he announced the American AI Sovereign Wealth Fund Act. (00:38:34) Sovereign Wealth Fund, one of Trump's favorite things, a one-time 50% tax on stock, not profits of the largest AI companies, including OpenAI, Anthropic, and XAI. (00:38:44) The shares go into a government sovereign wealth fund, and (00:38:48) would give the public voting rights and equal board representation at each company. (00:38:52) Sanders said, quote, the foundation of AI is our collective in human intelligence, the books, the songs, the journalism, scientific research code, essentially stolen by some of the wealthiest people in the world. (00:39:06) I'll stop there. (00:39:06) That's good. (00:39:07) It's a brilliant, it's a brilliant pitch unifying Sachs. (00:39:12) This pitch unifies Bernie, even Steve Bannon. (00:39:17) people in the administration. (00:39:18) Trump himself loves a sorrow and wealth fund. (00:39:21) He wants to own part of companies. (00:39:22) And Bernie Sanders, I think the horseshoe theories manifest here, Sachs. (00:39:27) What's your take on Bernie's proposal? (00:39:31) Well, I'm not in favor of Bernie's proposal because it's a straight up confiscation of property, and I just think that'd be a terrible precedent. (00:39:38) You just can't do that. (00:39:39) However, I do have sympathy for where it's coming from, and I understand and could support some sort of (00:39:47) more voluntary means of allowing the public to participate in this. (00:39:50) Wow. (00:39:51) Here's the reason why. (00:39:52) You've got all these AI CEOs telling the public that they're going to basically put half of them out of work. (00:39:59) 50% job loss. (00:40:00) That's