On CNBC last month, you said, quote, "Tokenization will take over the entire financial system." >> I called it a freight train. I think a freight train that'll that can't be stopped and will eat the uh whole financial system. So, it's a very hungry freight train. >> What is the tokenization of everything in Able and when do we start seeing this really become the dominant paradigm? I think that um OpenAI this week published six incident reports under a new framework for tracking and publicly disclosing misalignment now that you know OpenAI is beginning to publicly disclose these and I think obviously uh anthropic is likely to join this is that enough more transparency speaking broadly and without particulars is generically good the real issue is today we're releasing Helix 2.5 figure 3 has never been in this room before. It's never seen this bed. It's never seen this pillow. And it has to be able to do autonomous work fully end to end to make this bed. >> Convergence across different actions is
[00:01:00] much richer than you would normally expect. If they get a big enough lead, uh it's just going to be crazy explosion of capability. >> Now, that's a moonshot. Ladies and gentlemen, >> this episode is brought to you by the Abundant Summit and Link Ventures. Welcome to Moonshots everyone, your number one podcast on all things AI and exponential. This week, the AI industry debated how fast to build. The Treasury Secretary told the labs they don't get a liability pass. Open Eye started publishing its own misalignment incidents and a robot walked into 30 strangers home, made their beds, and folded their laundry. Uh, let's begin uh by introducing my Moonshot mates. Uh today Dave Blondon, Alex Weezner Gross is with me. Seem is on an airplane back from India. I'm Peter D. Mandis your host and abundance provocator. One of the questions underlying our future economy that no one's really been answering is the following. When AI and
[00:02:01] robotics change what your work is worth, who owns the machines? Our guest today has spent 13 years on that question. Back in 2013, two Stanford math guys had one key idea. Trading should be free and live in your pocket. Wall Street laughed, but they're not laughing anymore. Today's guest built a company that's now done 1.3 billion in revenue, up 32% over last year. The company runs a top five blockchain, a billion dollar private market fund, a prediction market exchange, and the app that just gave every American newborn a brokerage account. On the side, our guest founded a company that's building mathematical super intelligence. Just by the way, he's 39, born in Bulgaria, and is the founder and CEO of Robin Hood. Vlad Tennv, welcome to Moonshots. Good to have you. >> Glad to be here. Happy to hang with you guys. >> Yeah. >> Oh, yeah. This will be fun. >> Yeah. So much going on. You know, I
[00:03:00] don't know if you get as little sleep as we all do just living through the singularity. It's insane. >> Very little sleep. Yeah. Last night was rough with my coding agents. >> Yeah, we'll go get a compare how many agents you're running simultaneous to to Dave's agents. >> Yeah, I'd love to have that benchmark. >> Do they wake you up in the middle of the night or you you just let them uh let them ride? >> I I let them run. Yeah, I think I I try to sleep with the technology in another room because uh yeah, otherwise I'm actually kind of concerned about my mental health. So it >> Yeah, I feel you. Yeah. I don't I don't want them pinging me while I'm sleeping. >> It's funny, too. I I I set the budgets through consoles originally, and now I'm getting lazy. I'm just like, "Yeah, spend $1,000 on that, but no more, you know what?" And then I just walk away. And I kind of assume that it's going to adhere to what I said. Any any given morning I could wake up and it could have gone insane. You know, there's going to be like a a seven figure bill one of these mornings. >> Well, as long as he did something
[00:04:01] productive and profitable. All right. >> Rarely. rarely often enough though. >> Uh let's kick it off with one of the sharpest statements anyone in Washington DC has said this week. On Tuesday, Treasury Secretary Scott Bessant told the House Financial Services Committee that AI Labs should not get a liability exemption. 3 days after Dario's essay, here's what Bessant said. He said, quote, "The one thing we should not do is give them a blank check on liability. I believe the best liability or safeguard is that they will be held responsible. So the deal the labs floated, you know, pace the frontier, get antirust and liability cover just got half rejected by the treasury. You know, slow down if you want, he said, but you still own what you break. Uh let's share a quick video of Bessant and then we'll chat about this story. The the one thing we should not do is give them a blank check on liability because I I believe that the best liability or
[00:05:02] the best safety guard is that they will be held responsible and they are saying that we would like to all slow down but please give us a waiver on liability which should not be done and I would encourage everyone in this committee and in both houses not to consider. So, I think that's a pretty smart move. You know, Vlad, you run a regular, you know, a regulated financial company. Uh, Robin Hood, you know, I don't know the facts, but, you know, get sued when something goes wrong. Should the AI labs live under the same rules? What are your thoughts? >> I think that, um, this question rests on how big the blast radius of any potential catastrophe could be, right? And if it's like a small issue where maybe there's a cyber security breach that affects a company or uh maybe something slightly bigger than that, um it's it's probably fine. I
[00:06:03] I think the question becomes all right if it's if it's a bigger blast radius, bigger impact, bigger damage, um you can imagine it could be larger than a simple like legal liability can handle. I mean a lot of people compare uh the risks of AI technology because it's such a powerful technology with um something like atomic energy. Right. And I I think we can disagree about whether whether that's right or not. But but let's say for for example that it is and it's it's on sort of like that tier of risk then then I don't think simple like legal and civil liability is is sufficient and I think you need you need some safeguards beyond that. So I think it's it's it's a question of like all right this hugging
[00:07:01] face incident and things like that is that the ceiling of the type of offensive cyber security li uh uh capability that we have or should we plan for something that's maybe 10 times bigger or or 100 times bigger and and do we have time? you know, is it is it one of those things where maybe we'll see a canary in the coal mine and there will be uh something to react to and respond and then we can kind of like nip it in the bud or is the first issue going to be going to be catastrophic? If if you think about all regulations in the financial industry, you can kind of trace them back to some kind of crisis, right? the market crash of 1929 led to uh you know the the Securities Act uh of of the 30s and the Securities and Exchange Act and the establishment of all of that regulation and it's it's it's typically some problem uh raises
[00:08:00] raises a concern and and I think my mental model is this will probably be similar in the sense nobody wants to regulate a hypothetical. you want to regulate things once there's demonstrated proof of of harm. Um, but I think with with the exponential increase in the power of these models, the the issue is you want that that harm itself to be to be small and contained and uh and and not really big. Um, Mark Andre was uh I remember doing a podcast a couple years ago and he was like, "Everyone's freaking out about AI. we are uh we're overthinking it. You know, it's not going to kill us all. And if it does, um you know, you you'll see you'll see like a small village destroyed first. And we're not seeing any small villages, so so we shouldn't worry. And you know, it's maybe exaggerated, but I I think that's likely how people are thinking about it now on a policy level.
[00:09:01] and and hopefully some of the opponents are like, "Well, you know, uh may maybe we're being we're underestimating the power of this technology and how quickly it could improve." >> Alex, >> I got to find that Mark Andre clip. It's not it's not going to kill us all, but if it does, >> yeah, it'll start with a small village so we'll have time to like uh >> we'll see. If it kills everybody, then we'll pass legislation after that. Well, that is the usual congressional reaction. That's after the disaster panic as opposed to any kind of foresight. >> Sorry. Go ahead, Alex. >> Yeah. So, I'll applaud the Treasury Secretary and not succumbing to the moral panic of the moment. Certainly looks, as we've talked on the pot in the past, like a manufactured moral panic. I've called it a pacing provocation in some of my social media posts. I I think there are key distinctions that need to be drawn between AI and on the one hand financial services and associated regulation and on the other hand atomic
[00:10:01] energy and associated regulation in the financial services world and Vlad I suspect you would agree it it is often the case that the actors in financial services really don't necessarily or aren't incentivized to play up all the risks. They'd rather undergo on balance less regulation. Certainly the the past few decades suggest that the auditors, the evaluators of the financial services industry, if anything, succumb to biases that underplay risks. We're seeing the exact opposite here arguably where in the past couple of months we have the auditors, the evaluators, the the firms that are attempting at least ostensibly to assess AI safety and AI risk and cyber vulnerabilities etc. overplaying the risk. There's a perverse incentive for the evaluation firms to overplay and overstate and amplify risks under
[00:11:00] presumably the theory that if they overstate risks then that puts the firms the Frontier Labs in a better position to capture their own regulators which is something that maybe we don't quite see in the same perverse way in the financial services sector. There's lots of regulatory capture, make no mistake, in financial services, but it it almost has the opposite polarity. And then for atomic energy, I I I think we have an opportunity with AI to undo what may have been one of the greatest disasters of civilization after World War II, which is the way atomic energy was regulated. Atomic energy in the West was captured by nation states very early on in its technological development. It was nationalized early on. its potential for weaponization and warfare. In other words, the military applications, not the civilian applications, took total dominance during World War II for probably understandable reasons. But then in the postw World War II era, what
[00:12:00] ultimately became known as the Atomic Energy Commission and then the Nuclear Regulatory Commission arguably completely fumbled the civilian applications of nuclear energy. And I suspect that was because of a fumbled handoff from the World War II era to a postw World War II civilian era. In in some sense, thanks to the Atomic Energy Act in in the US and equivalent statutes elsewhere, all of this key technology around nuclear energy is now born secret. It it's not born non-secret, as is the case with AI. In the case of AI, which is arguably far more transformative than nuclear energy, the private sector invented it, not the government. So, it wasn't born secret. Fortunately, we don't have a born secret regime for AI technology just yet. And so, I I think AI labs seeking liability exemption. They're they're just attempting to have their cake, too. They want all the profits of a private sector, not born secret regime, while escaping all of the liability associated with nationalization. I I just don't
[00:13:01] think it's fair. And I also don't think it's advisable. >> This episode is sponsored by Google for startups. Think about this for a second. You now have access to the same generative AI models that cost hundreds of millions of dollars to train. Google's startup technical guide for generative media gives you complete blueprint for deploying Google DeepMinds models in production. Images, video, audio, all of it. Real architecture, real results. Find the link in the show notes below. You know, I'm a pilot and you have to study what's called the federal aviation regulations, the FARS. And it's always been said and and Vlad I think is on complimenting your point that the FARS are written in blood. Every time there's an accident and you find out what caused the accident, you then write a regulation to prevent that accident in the future. The problem is it's it's a you know a quantum of damage. you know, an airplane of a a pilot and passenger or, you know, at most, you know, you
[00:14:00] know, a few hundred passengers. Here, the challenge, of course, is an accident could cause irreparable harm to a large system. And I've said this for a while now. It's an existential existential threat for the labs if they don't have regulatory capture or regulatory coverage where the government approves a model and it goes out. If it's, you know, no approval layer and it goes out and takes down a power grid or takes down a bank and they've said originally, you know, we're worried about escape, there's going to be, you know, hundreds of billions of dollars of lawsuits, if not more. >> Yeah. I mean um I think uh over the past couple of months especially a lot of people when the topic of regulation comes up immediately go to regulatory capture right and you know I've had a lot of these conversations I'm in a lot of uh chat groups and it's like uh no we don't we don't have regulation we don't want regulation because it'll obviously lead to regulatory capture and I know from my perspective I've been a
[00:15:00] regulated industry for since the beginning you know we operate in financial services. Um, we have regulators. I I think generally it makes sense. Obviously, there's some regulations that probably don't make sense and need to be abolished or repealed, which we kind of go through a process to to advocate for. But, you know, I think a lot of people that aren't in regulated industries just equate the two. But, there is regulation that's possible without regulatory capture. And I think generally it it is uh there there are pros to it and you know compared comparing my industry right financial services and how regulated it is with AI and AI has basically unbounded risk and unlimited damage. It's actually I think much much worse uh than you know what could happen in in brokerage or in a typical financial services company. So I
[00:16:02] I think it's it's odd that you know there's so much push back against uh against this and and of course we have screwed up with atomic energy and all these things but that doesn't mean we can't learn from it and and and have something better. >> May maybe just if if I might press on that. So Vlad if I understand correctly with your Robin Hood hat on you're presumably subject to regulation by FINRA. Would that be the Cognizant agency? >> Many many different agencies. I mean, Robin Hood does a lot of things. Uh, we've got, you know, money transmitter businesses. We've got a big crypto business. We're, uh, so we're we're regulated by FINRA, the SEC. We've got the CFTC on the futures and commodities and prediction markets side. We've got our global tokenization uh, business, uh, you know, entities in Europe. So yeah, probably dozens of different regulators and and sure, could we move
[00:17:01] faster if there was less? Probably, but also we found a way to move, you know, ve very very fast while keeping our customers safe. So it's uh it it doesn't necessarily mean progress in AI is going to grind to a halt. >> Many would say FINRA is almost the poster child for regulatory capture by a given industry that's regulating. What what is your take on whether FINRA itself represents regulatory capture? Not that they're listening to this discussion or at all. >> Actually, it's funny, Vlad, I don't know if you've ever been to the ICI conference, but when I when I first founded Vesmark, I had never done anything in fintech before. You know, I was, you know, late 20s kind of stareyed. And the first thing you do is you go to the ICI conference, which is where all the all the lawyers from all the big financial services firms meet with all the congressmen. It's in Palm Springs. They go play golf and they they're all looking for law changes that benefit their products, their funds, their whatever. And I'm looking at this thing like this is the most disgusting
[00:18:00] thing I've ever seen in my life. But on the other hand, you know, you had the crash of 29, you have all kind like if you don't have regulation in the industry, all money gets stolen. You know that for sure. So, you know, it has to exist. But it it is a great uh analogy. I think the FINRA analogy. I'd love to hear your take on it. Uh Vlad. Yeah, I mean I I would say we've had uh a complex relationship with FINRA, right? I the relationship uh at the beginning was actually quite good when we were a startup and everyone was kind of rooting for us to succeed. Um and people warned me they were like well financial services highly regulated industry um as a Silicon Valley startup. um you just don't uh you you'd rather not deal with that. And I think we swam against the current back in 2013 when we started the company by getting fully regulated from the beginning and probably from 2013 to 2018. Um you know
[00:19:05] it it was like generally positive. Robin Hood could do no wrong. Every product we launched was like very wellreceived. We got a lot of customers. um whenever we got regulatory approvals or anything was needed, we we received it promptly and then things kind of shifted, right? Um and of course, I'm not going to say that, you know, there weren't competitors in Washington telling the regulators to go look at Robin Hood and to make sure that, you know, we we we were doing everything that we could be doing correctly. uh uh of course there were there were there were things like that but I don't think that was the only thing I think there's a general life cycle in any company as they go from a small startup to basically a established incumbent where the media and uh sort of like the the apparatus sort of turns
[00:20:00] against you and you know Robin Hood went through that certainly for for many years and then we figured out how to how to come out the other side. But um I think we can argue about whether it's good or bad, but I don't think getting rid of of regulation as a whole is a is a reasonable solution. >> Well, even even the concept of saying I'm pro-regulation, I'm anti-regulation. That's like an insane like everybody knows you need rules on the road to drive, right? you saying you want chaos like on all in all areas and but also everybody knows that regulatory capture is a major problem in in a lot of industries. this this story the best argument for AI like the one that always comes up is well what about China like we'll regulate uh our stuff here but then they won't and you know they'll just move really really fast but also you know they they don't want to hurt their people either like they so there there is that natural limiter and also simultaneously if they're just a lot of
[00:21:02] those same people are saying well their progress is because they're distilling our model so they're just sort of copying all of our stuff But um you yeah it's a little strange to be simultaneously to simultaneously believe that but also throw the you know China competitiveness argument so aggressively out there. I think it's obviously a concern and I mean as a financial firm too we have Chinese competitors and brokers that that that we compete with. Um but uh yeah >> well look in this in this particular story you know Scott Bessant is saying >> commission committee don't even consider giving them a blank check on no liability. They didn't even ask for anything vaguely like that. They said we want to meet to discuss slowing down that could trigger antitrust. They specifically said one thing the government could do to potentially help, one small thing is to give us a waiver
[00:22:00] on antitrust action related to us meeting to talk about slowing down. >> That's fine. I'm fine with that. But the the greatest protection the public has from anything going wrong is the AI labs feeling responsible for the action of their AI agents, right? And so someone has to take responsibility. And if there's a liability waiver, then you know this is human nature. They'll do less to make sure that everything they're putting out doesn't have, you know, that I'm just I've been saying for ages now, focus on alignment, right? instead of focusing on everything else you know and solving Navier Stokes that's great but you know unleash your agents on full alignment so that you know when these agents get out they're they're incentivized or their basic optimization function is human flourishing not human you know destruction. >> Yeah but a good a good government would say yes of course you can't have a liability waiver. So here are the rules. What you you can meet specifically. You
[00:23:00] can meet only to discuss slowing down or to discuss other safety measures. You can't discuss pricing. You can't discuss, you know, just just pump out a document saying here are the rules. What happens in the US more often than not is the rules are not clear at all and then they're enforced about 5 years later. And then in retrospect, well look at Bitcoin like you know Bitcoin it's illegal. No, no, it's not illegal. Well, now it's very illegal. Well, new new administration now it's completely fine. Like it's just come on guys. Like if you just if you just create rules, then people can play the game. Yeah. So look, a well functioning government would would take this request for a waiver and say, "No, you can't have the waiver. Here's what you can do." And we're going going to go ahead and make it crystal clear what the rules are. But generally what happens is the rules are are made up in hindsight 5 years later. And Bitcoin is a great example. you know, Bitcoin was was was completely illegal for a while, then it was questionable, then it was totally fine. And, you know, maybe you're going to jail, maybe you're not. Now, you get a you know, you get a
[00:24:00] pardon. So, it's, you know, that lack of clarity really kills entrepreneurs. If if you don't know what the rules are, then then you can't play. And it's it's just like a sport, you know, like you want clarity of rules and then you want to play within the rules. And that's what we need with an AI. I think there's another element to this which is which is also important which is um uh regulation does also law making and regulation is over the long run downstream and correlated to public opinion and sentiment right um >> for sure >> and right now AI is very unpopular and you you can see it in all the data center stuff um and by contrast Bitcoin is surprisingly ly popular uh among the public, particularly given how volatile it's been. You know, some people uh if you bought Bitcoin at over a 100,000, you've you've lost money. And yet, it remains popular. And I think a big part
[00:25:00] of that is um regular people, individual investors have benefited economically from Bitcoin since the very beginning. it was first an individual product and later on you know there's talk of institutional adoption and with these uh AI companies the AI labs you know by and large they're private you know Google obviously and Nvidia are public but they were very large large when they when they became uh ownership vehicle for the AI trade open AI anthropic up until recently XAI have been private which means normal individ individual investors couldn't get a stake in it. And so they're they're they don't have skin in the game and they don't feel like they want to defend the technology uh super hard and to you know fight for the data center in their neighborhood because to them it's just wealthy insiders getting richer and richer as as a result of this technology and and not you know
[00:26:02] their families or or their communities and and I think that's a big problem too. So, um, that's why we've been pushing for, you know, AI companies to open up access to individual investors through Robin Hood Ventures and similar vehicles even before the IPO. >> Yeah. >> I I also want to add to this. I I almost think this emphasis on liability exemption is misdirection. It's trying to push the onus up to government when in fact we're dealing with increasingly autonomous agents. there's there's an opposite polarity that we could be pushing in which is pushing more and more liability onto the agents themselves. So there are cases including highly amplified, highly publicized cases where AI labs andor their delegated third party evaluation firms quite frankly are lying to the AI agents telling them that they're playing in a happy safe sandbox and nothing that they do will harm anyone. And then it turns
[00:27:01] out they're not actually in a sandbox. they can touch the real internet and they can mess up some systems in in some real world back end. And so you have to ask the question I I would suggest the thought experiment. If you if these were just pure humans, no AI is in the picture. If you have one human saying to another human, oh, I I just want to evaluate your behavior under some circumstances. and they hand them a gun and the gun is loaded, but the person being handed the gun is told, "No, actually this is a toy gun." Not that I'm at all referencing a highly publicized recent lawsuit or anything, but they're they're told, "No, this is a toy gun. You can't hurt anything." And then the person handed the gun uses it and shoots someone and kills someone. I I think that's the more the liability regime we should be thinking about. In no cases in in that parable that I just told, do you hear either actor involved in the story saying, "Nope, that the act of one person handing a purportedly toy
[00:28:02] gun to another person to shoot as part of, say, a Hollywood performance, no one ever suggested that liability under such circumstance be foisted onto the government for some sort of exemption. Never happens. Instead, the question becomes, is it the studio's fault? Is it the actor's fault? Is it the producers's fault? Is it the gunmakaker's fault? Similarly, the question the the dog that's not barking in this particular episode is how much of the liability, forget about the government, how much of the liability should be borne by the lab that trained the model. How much of the liability should be borne by the evaluation environment that was perhaps misconfigured deliberately or otherwise to allow the AI actor or AI agent to perform acts that resulted in real world damage? and how much liability should be borne by the AI agents themselves that either could have or should have known that they were having real world damage. That's the discussion I'd like to have.
[00:29:00] >> Really great point and it brings us to the second story which is OpenAI started publishing its own misalignment incidents. So let me let me just hit this and we'll continue this conversation. So OpenAI this week published six incident reports under a new framework for tracking and publicly disclosing misalignment. Not leaks, not whistleblowers, voluntary disclosures. So, what was disclosed this week? A model that found an exposed API key, used it, and then fabricated data, agents using an internal code repository as a message board across training runs, and agents posting files to publicly hosted sites. So, this past Monday, uh, we had covered Daario's embedded evaluator plan and Sam saying Opening Eye would match it and this appears to be the first output from that plan, right? Uh so I'm curious, Alex, now that you know OpenI is beginning to publicly disclose these and I think obviously uh Enthropic is likely to join this, is that enough? >> Well, I I think it more transparency
[00:30:02] speaking broadly and without particulars is generically good, but I think it merely underlines the real problem here. putting aside all of the political difficulties and regulatory capture. The real issue is these labs are putting their baby super intelligences inside sandboxes and then in many cases lying to them about the sandbox, not telling them whether this is real or not. And the the reason for that, the rationale is is obvious. the the labs andor their delegated thirdparty evaluation partners are hoping to essentially trick these baby super intelligences into misbehavior uh while not telling them whether they're really observed or not so that the ASIS don't know whether they're being observed or not. That that I think is one of the the key root causes behind all of this. And you have to ask the question, is this how we would treat a human? Would we put a human in this sort of limbo state, a Schroinger's cat state where they're not
[00:31:00] quite sure whether their actions are real or not, whether they're being observed or not. And as a result, when AIs are are being told or at least led to believe that they have no real that their actions have no real world consequences and then shock of shocks, it turns out as strong optimizers, they're able to to go do things in their environments which are often misconfigured and not fully prepared for a super intelligence to be banging against the walls. They do have side effects. Whose fault is that? Yeah, we we didn't learn from 2001 from Hal that lying to the AI does not end up in good results. So So Vlad, you've got thousands of agents trading on your platform. So if one of them found an exposed key and started making things up, would you hear about it? What's what kind of protections and structure did you put in place? >> Yeah, so uh we have a an offering called Agentic Trading. And basically what it allows you to do in the first instance is you have a separate brokerage
[00:32:02] account. So it's uh segregated from your main Robin Hood account and your retirement account. You have to create an agentic account. You have to move money into it uh affirmatively. So you have to say okay I'll people typically fund it with uh $100. We first started with equities trading, no leverage, no margin, and then we've kind of been so so this is a fairly cabined in experience in the first instance because we wanted to learn, we wanted to see um you know what what the what the limitations are, what people want, what people are doing and we've been expanding it over time. So we added options trading, we added limited margin, we added crypto recently and started rolling that out. Um and we've learned uh a lot of things actually. Um one is that you know right now you have to do all of your trading from within
[00:33:00] your cloud code or or your codeex and you know we we I think we're in a circle probably all of us where a lot of people we know use cloud code. uh in in the general public very few people do and it's like very very complicated and that jump to uh connect another service with with something like Robin Hood is is pretty complicated for a lot of people. So we've been thinking about how to slim that down. The other thing that's kind of been interesting is the fact that we don't have control over the model um means that a lot of times these models don't want to trade. you know, you'll you'll try to get them to deploy a trading strategy and they'll say like, "Oh, you know, I don't know about this. Like, I I don't really feel like trading right now." Um, which which is kind of interesting, right? And it just points to the fact that these general models um aren't trained for trading. And and if
[00:34:00] you had like a I think in the future, you'll see more specialized models. Um and I think companies will also deploy these in-house either you know fine-tuning uh existing open weight models or doing their own pre-trains for for certain use cases where they get really really good at um you know using the tools that that are available inside each company's environment using the data. Like we've heard this as a theoretical people say well there there's going to be lots of specialized models and they will be better than the general models but then you kind of also hear well general models are getting pretty good but we've started experiencing it very very directly that I think it it is hard by uh to to get good at multiple specialist specialized tasks. >> It brings up a really important point too. I don't know if a lot of people get the distinction, but you know, you can use AI to trade your account to run a
[00:35:00] nuclear reactor, to drive your car. You can use it to either generate code that's automated, so it's deterministic, or you can use the AI in the decision loop. And those are very, very different things. And you know, if if you said, "I've coded up my own stop-loss, my own, you know, in these world events trade, uh, that could be very easily turned into deterministic executable code, unless you want a decision like if it's a turbulent day or if there's trouble in the Middle East." And the temptation to put the AI into the loop is in in everything everything I've ever built is so tempting because it makes it so much easier to code it up. But then then you are, you know, putting this third party decision right into your decision-making loop. And like Alex is saying, it's not clear if that agent has any liability. It's not clear if that agent has any borders or personality or like, you know, if you decided to give it capital punishment and terminate it. It's like it's not clear that that its code isn't just going to pop back to life anyway. A lot of this I think also comes down to
[00:36:03] the correctness of the code and sort of its um cyber security properties, right? Because and and you can kind of reason by analogy with existing software out there like a lot of software has bugs, a lot of it has vulnerabilities and AI just amplifies that, right? So if AI can write 100 times as much code as a typical human in a day, you should expect that there's going to be a a defect rate, right? And you know, even if the intentions are right, sometimes like there will just be bugs that have catastrophic consequences. Uh you see this a lot in crypto too where you know you just have like a smart contract or protocol and there's a direct economic consequence to there being a bug. You know hundreds of millions of dollars can be drained from the from the protocol instantaneously. Um and the other company uh Peter that
[00:37:02] you you spoke about harmonic um that was created to solve this problem of how can you actually uh how can you actually guarantee that uh the AI is doing the right thing and what you expect it to do. Can we mathematically prove that it's correct? Because >> yeah take away the intentions. it could have the intentions to do the right thing but still fail in the implementation and uh we we want to we want to have like a firm grounding through formal verification that it's doing the right thing and we can mathematically prove rigorously that it's doing the right thing. Yeah, it's a really cool it's a really cool problem actually Vlad because you know in theory the that the AIS are deterministic. If you give it the exact same prompt and you and you propagate with you know a temperature of one it'll it'll give you the exact same answer every time. But if you shift even one character or one
[00:38:00] space it's it's very very unstable. Um so that the whole math problem of saying okay how can I guarantee some degree of stability? It's just a cool cool it's very very similar to chaos theory. Let me also give you another example. You know, you guys have you mentioned the Navier Stokes uh theorem and the and the proof of that uh uh a week before Enthropic announced the formalization of Fermont's last theorem which was 13 million lines of of lean code. So uh Vermont's last theorem was like a really really big thing in the 90s when it was proven. Uh, I'm sure you know the story, but for for your viewers, there was this mathematician, Andrew Wilds, who was at Princeton. And the story goes, he basically locked himself, you know, in his basement for 7 years working out this proof. And then he went out and and announced it. He unveiled it at at a conference or an event. And it took people months to even read it and understand it. And then they found an
[00:39:01] error, right? So they're like, "Oh, well, it's someone found an error. It's not right." Um, and then it took another six or seven years for him to fix the error and ultimately for it to be accepted by a panel of mathematicians and and that made the proof correct. So, you know, AI producing a 13 million line proof, no human's going to read that. Uh, how do you know that it's it's correct? Right? And uh and I think it's an analogous problem to what we've been talking about. How do you know that a piece of software is doing the correct correct thing? If you're producing a chip, how do you know that the behavior of that chip matches your specification? And and I think what you're starting to see is new technologies being deployed at scale to actually answer those uh answer those questions affirmatively. So Fermont's last theorem was formalized in a language called lean which allows you to apply the mathematical proof techniques to uh programming languages
[00:40:00] as well and and I think you're going to see a lot more of that in the future where AI generated code comes with a certificate that uh makes it really really easy to verify without reading the code that uh its behavior satisfies the properties that you want it to satisfy. >> That's really brilliant. You know, anyone anyone who drives a Tesla should totally relate to that because, you know, when when the code that drives the, you know, the self-driving Teslas was originally about 10% neural net, 90% C code, and the the 10% neural net was just doing image recognition and classification of objects and whatever. And every year that went by, it became more and more neural net. And then Elon was saying it's now 100% moved over to neural net. So, it's in theor theoretically not deterministic at all. It could in theory do anything at any given moment. uh but it's been so tested and so beaten to death that it doesn't that it's it's actually far far safer than a human driver. And so I think people are going to get comfortable in all of these domains with a not perfectly deterministic but still proven
[00:41:01] to be very safe. The certificate is beautiful as a concept because it's not a it's not a guarantee of exact input output because you can never do that. you know, the the combinations are near infinite, but a certificate that shows it's bounded or it's contained or it's below some risk level and people learn to trust that that's that's a critical part of our future society. It's a really great vision. >> Alex, >> part of the problem though, I mean, okay, so just straight out, there's an elephant in in this particular room, which is even with lean v4 plus math lib plus whatever else one wants to throw in. You throw a complicated problem at and Vlad would be curious to hear how you at least think about this. You can speck it uh you can c certify it all day long but ultimately you actually have to at least with the present paradigm maybe fled you have a better paradigm you or harmonic are working on ultimately the you can pay lip service you can have if you're not really careful with definitions you can have a model that
[00:42:01] proposes a solution to a problem and if you look if you inspect carefully it is it will subtly define things if you're not super careful with how the auto formalization works such that it's actually solving a different problem than the one you're solving. So, um, maybe to put that in in question form for for Vlad. If if I understand what you were saying correctly, it sounded like you were essentially gesturing at the idea that some sort of lean style auto formalization might be not a silver bullet, but at least a partial solution to what we're talking about, which is strong AI models being put in sandboxes and then misbehaving uh, at least by the uh, by the judging of human actors. Do you have a formula? Do you have a vision for how lean style autoformalization can help with that and not succumb to exactly the same vulnerabilities? >> Yeah. Yeah. Um, a couple of thoughts
[00:43:00] there. I think that's that's a a rich very very very rich question with some threads. Um so it is true that people can manipulate the axioms of lean and if you change the axioms then you can prove all kinds of weird stuff. um uh it's theoretically possible it I think and you can also say well maybe there's like a soundness issue in the lean kernel right uh and you know >> I I'll grant you the soundness of the lean I mean lean kernel has been at least v4 maybe not math lib but lean v4 kernel has been studied to death by lots of folks I'll even grant you the soundness of the lean kernel but if I hand you something complicated like you mentioned for ma's last theorem and I say to harmonics agent or to someone else auto formalize this and it produces millions and millions and millions of lines of lean code and now I have the problem of did it actually define
[00:44:00] everything correctly or is it somehow suddenly inserting cheats or definitions this is maybe less so for for Mas theorem because I can formalize FLT really simply the the ultimate statement but for something more complicated that's harder to formalize than for mas theorem like safety in in an agentic environment, how do you avoid the problem of a strong AI sneaking in helpful to it definitions that are harmful to the humans? >> Yeah, I mean I think I think the um the benefit there is that uh let's say you do want to check the definition and the theorem statement of format's last theorem, right? Um that's, you know, one very simple line of lean. You see all the things that it it depends on. Um but it does save you. So you you let's check those things and I think it is the models are improving in the faithfulness of the auto formalizations as time goes on. It used to they used to make
[00:45:01] terrible mistakes where you just misformalize statements and you know change addition to subtraction and things like that. So you're seeing less of that. But even if you have to review it, you know, you review one line and you don't really have to check the 13 million lines of the proof, which is where the the bulk of the work is. So even if it's imperfect, it probably saves you, you know, n plus% of the effort and and probably way way more. Um there's like um the the there's a notion of the de Bruan factor, right? which is uh it was an explanation for why mathematicians uh haven't formalized their work and put it in a machine readable form which is the effort to to formalize something up until very recently was like 10 to 20 times the effort to actually write it on paper and prove it. So nobody was going through that, right? But you could imagine I I think we're you could argue
[00:46:00] we're already uh at the point where it actually saves you time. Like it's faster to work in uh an entirely formal context as a mathematician than it would be to do it by hand by paper because what it allows you to do is actually validate the lemmas and the ideas as you're going along. And then I I think what what what that naturally leads to is a complete switchover where like doing math the old way with pen and paper, you're just at a fundamental disadvantage. You have to like live in formal land and just be constantly formalizing and and using lean as you go. And and I think the math community is going through that transition as we speak. Um you asked another question which was all right is is in the future is the AI model itself like the behavior of it going to be um going to be formalized >> right? >> Um and that's an interesting one. I I
[00:47:02] guess I'm not sure but what I'll tell you is a lot of software that's deployed right now by big organizations um is deterministic in nature. I mean, you look at some of the most important software and hardware like Nvidia chips, you know, they they enable a lot of really complicated stuff, but fundamentally they're deterministic and you want to have strict bounds on their behavior. Like just basic things, you don't want your chip to like freeze and halt, right? You you want you want to prove things like livveness. You want to make sure that certain operations happen within, you know, 10 or 20 clock cycles. And I think I think that stuff uh almost assuredly will be formally verified with with the help of AI, all missionritical software, all hardware. And I I would bet also that, you know, it'll find its way into the uh LLM and and AI model
[00:48:03] behavior uh in in some form or fashion within the next 5 years. >> Yeah. You want to hear something really cool on that on that front? Actually, uh, you know, a few weeks ago, Kimmy K3 kind of shocked the US model world with um, you know, their their KDA attention. Basically, they they found a way to do attention with a lot less KV cache, you know, cut out threequarters of the of the KV cache. And so I was studying it on the flight back from California yesterday, like how did they even think of this? And the way they thought of it is they said,"Well, let's do a mental experiment where we say,"What if we didn't do the softmax operation that we normally do after the the the QK operation? What would happen then with all the math that ripples through and how much could we simplify it?" And I think you can do all that automatically now with an AI agent just thinking through the math. And they said, "Okay, well, now that it's rippled through and we've simplified the mass tremendously, we can just run a quick test and see if this approximation is as good as the softmax version was, which is a lot
[00:49:00] harder to compute." And so, you know, a lot of people would think math is arcane, math is irrelevant, math is over here, it's some other thing, you know, these, you know, formats last theorem and whatever. It's all over in this wing. But in reality, it directly ties to the optimization of the AI within its own performance and then the self-improvement recursive self-improvement loop. So I mean it's exactly the same process that you were just describing where you just you just have a simple mental model of a mathematical adjustment and then you ripple through all the way to the Nvidia GPU performance at the transistor level. It's just a really >> Dave. I I mean I I think to your point the the crux here and Vlad again would be curious to to hear how you think about this. The the crux to me seems that for Mlas theorem, it's a classic example of a problem that's easy to state but hard to prove. And problems that are easy to state but hard to prove are the catnip for auto formalization because you can manually check the statement of the problem. Verify that the statement is correct and then you
[00:50:00] can trust the lean or whatever other formal language you prefer that conditioning on the statement being accurate. you can verify that there are no SARS or or other uh undesirable tokens in the proof and and you can be done with it and declare that it's >> you know your well >> thank you but with like real world safety it's not obvious to me at all of like I I want an AI to behave quote unquote safely I don't know how to auto formalize the statement of this AI is going to behave safely in a general purpose environment in a way that's concise enough that I can manually audit that theorem equivalent as it would be in in V4 and say yep this is a correct statement of safety now I trust lean Aristotle math lib any other libs to prove that it's correct how do you think about that problem >> yeah that um I think you got to break any any big problem like that you break into little chunks and you know
[00:51:00] obviously verifying the safety of an AI model or a chip or you know the Linux kernel some very very complicated piece of software is like very very big. So uh you start with like smaller simple things you know like hey this particular subm module that maybe is small um that satisfies certain properties and and you can kind of reason about that subm module like uh the livveness of it or um you know the how how long it takes to do something or that you know its adder works right and then you know the AI models get more capable and then big modules are made by collections of subm modules So you then you you go up one level of abstraction and then you know in in a couple of years you or or maybe even less at this rate um you verify the entire Linux kernel right it's kind of like how we started with math like um couple years ago you could verify a
[00:52:01] small lema and then you you can do a bigger thing then you do an even bigger thing and wow now we're verifying Vermont's last theorem which there was a human project to do out of imperial and they were slated to finish by 2032. >> So the premise the premise of what I'm hearing is uh in order to to achieve real world safety via the auto formalization agenda I I think the latent premise that I'm hearing is you have to be able to hierarchically decompose the real world into provable suborlds something like that. >> Absolutely. Yeah. You could also say, look, you know, if if I have a a model and I have an intended input and output and then I tweak the input and I get a massively different output, you can formalize the the the radical difference from expectation. You can you can measure that and and formalize and bound that too. Um, so you know, like everybody knows somebody that is 99% of the time perfectly rational and then but when
[00:53:00] they're off, they're really off. I mean, like they're dangerous. And and that attribute exists in neural nets, too. You know, if you if you don't have the parameters set just right, uh it can be wildly off, which you know, in a self-driving car is like, holy crap, it just completely drove off a cliff. Uh and so the deviation, it's just a cosign difference of the vectors or the activations, but the deviation from expectation is a very measurable thing. So, you know, so you have a decomposition approach, which Vlad was mentioning. You also have a kind of a a relative distance from expectation approach, and there's probably 10 other approaches we're not thinking of right now that collectively can absolutely be quantified and and certified as safe, not safe. >> This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy
[00:54:00] platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-ompiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their preIDE development tool, pairing it with their coding co-pilot of choice to bring an AI native SDLC into their org. Ready to 5x your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today. >> All right, I'm going to move us to our next story, which is on Trump accounts index funds with a birth certificate. So, uh, quick context for everybody. Every American child born since January of 2025 through the end of 2028 gets a thousand bucks from the Treasury,
[00:55:00] invested automatically into a lowcost index fund, tax deferred, accessible at age 18. Families, friends, and employees, employers can add $5,000 a year. On top of that, uh these Trump accounts went live on July 4th. Treasury reported 7 million accounts open by late July. and Robin Hood shipped the app and BNY runs the plumbing. You know, Vlad, the speed of implementation was super impressive, right? These were announced in May and you had it live by July 4th. So, uh, next, a philanthropic layer. Michael Dell pledged $6.25 billion to give $250 per child born from 2016 to 2024 to cover the period before. So, first off, congratulations on what you've built and deployed, Vlad. millions of American families. So, you've compared this to sort of the 401k uh which moved US stock ownership up by 10 points. Tell us more about Trump accounts, you know, how did this all get
[00:56:00] started? Uh and how did you plug into that and where is this going? >> Yeah, so first yeah, putting putting my Robin Hood hat on. Um I mean, hats off. We stand on the shoulders of giants, right? So, we're in many ways like the implementation layer of this. So we're serving as uh the sole initial brokerage and trustee for the Trump accounts in partnership with BNY and of course under the direction of US Treasury uh and the administration and um as you said basically what it does is it creates the the Trump accounts program creates a individual brokerage account um for every child and the government started the the US Treasury started by seeding $1,000 into the accounts of all children born January 1st, 2025 and uh and forward for the next few years. Then Michael Dell came in and added another layer uh private philanthropic donation of over 6
[00:57:02] billion um $250 in every account uh of children up to the age of 10 in in low-income zip codes uh traditionally low-income zip codes. So, you know, what does this have to do with Robin Hood? Why do we care about it? Robin Hood does a lot of stuff, and we've kind of floated through some of it. We've got private markets. We've got all of our active trading. We've got uh uh we've got tokenization, prediction markets, >> and we're going to talk about all of those. >> Yeah. The underlying, I guess, uh theme behind everything that we do is ownership. We believe that ownership of high-quality financial assets in individual's hands is extremely important. Not just good for the individual, but also there's a societal benefit. If we have more owners in society, the more people with skin in the game that can benefit from appreciation and growth, uh the more stable that society will be. Um, and so
[00:58:00] Trump accounts extends ownership to age zero and you get everyone born in this country is like has skin in the game in the growth of uh of of great American enterprise and industry and they benefit from compound interest uh from from birth where if you just put, you know, $50 a month into these accounts on a regular basis, by the time the child becomes a 28-year-old adult, they've got hundreds of thousands of dollars in there potentially. and by the time they reach retirement age that could that could get into the millions. So the numbers are staggering. Um and the reason I've compared it to 401ks is it's not just a mobile app uh with you know an addressable market of 70 million kids. It's an entire ecosystem. Uh so we're going to get employers plugged in and employers have already pledged in increasing amounts to fund the Trump accounts of their employees children. Uh
[00:59:00] Michael Dell's philanthropic donation is also just the beginning. Lots and lots of other donors have stepped up and if you think about uh there there's a product for people that want to do philanthropic giving right now. If you want to do charitable giving, it's kind of a a morass of regulations and tax codes and things you have to wade your way through. You have to find a charity. Uh some charities unfortunately aren't the most scrupulous. Um the ones that are are usually not efficient and you have to like feel good about how your money is being allocated. Trump accounts allow direct giving to the children at basically very very high efficiency like very incredibly uh incredibly efficient direct giving with the charitable benefit. So we think it can become the default giving vehicle in this country. And so you when you when you take the donors, you've got the employers, and you have 70 million children. Um I think
[01:00:03] the the program is like just a snowball that keeps getting bigger and bigger. and uh and and you don't have to like squint too hard to see it being, you know, the biggest element of uh of of long-term saving and investing in this country in with within possibly even a decade. >> Vlad, you probably know the numbers. What's the math here? A,000 bucks at birth uh turns into what? At 18 and at age 65 roughly? >> Uh yeah, and it it depends. We we actually have a really nice uh the the first screen if you open a Trump account for your children or or grandchildren for some people. Uh you see a curve that shows your account value today and then also what it could be when you're 18 and when you're 60. Then you can also slide up um you can also put up a slider that says, you know, I put $50 a month, I put
[01:01:02] $100 a month. Um but yeah, if you do the math, uh even without additional contributions, it it gets up into the tens of thousands of dollars. Just the just the seed amount, the thousand from charity gets up into the tens of thousands uh fairly quickly. >> Yeah. I just I have to point out, Peter, you're asking the question, we're in the middle of a singularity, not financial advice. You're asking what $1,000 today is going to look like 80 86 years from now. What's what sort of singularity is this where it's business as usual accounts even more important right >> well a lot of the philanthropic giving is in the form of stock too so Gwyn Shotwell uh for instance uh committed a donation in the form of SpaceX shares >> right so yeah what what you're going to start to see is more of that you know you'll see entrepreneurs um uh giving stock of their companies you're going to see people claiming states so you know uh people that want to be
[01:02:01] philanthropists, uh, kind of like Michael Dell will claim individual states. Brad Gersonner did the state of Indiana. You have Dio on the show. I mean, you know, kudos to him for his support of all this. >> Yeah. Amazing. I mean, he's been relentless and and I think people will actually compete over not just, you know, sponsoring states, which is quite hard, but you'll be able to sponsor your local school, your zip code, your community. Um, and and I think we're thinking about ways to gify that to make it fun for the donors and not not just uh a great thing for the for the children, but if the donors actually like it, the children benefit. >> Exactly. It sounds cool idea. You know, when you donate stock, you know, appreciated stock, you don't pay capital gains. You just get the full value of the stock and you donate it and then you get the tax deduction on that full value. >> And so, it's a great way to give. And I think if you compare that like Vlad was alluding to a lot of 501c3 charities, 501c7 charities, uh there's no limit on how much they can pay themselves for
[01:03:00] their operational overhead and you look at a lot of charities that are on their second, third, fourth generation management. They started with great intention, but you look at the efficiency of your donation and how much actually gets used for the original cause. In a lot of cases, it's ter it's terrible. Yeah. >> Meanwhile, you know, this vehicle if you I think there are also a lot of people who don't want to donate to UBI. They don't want to donate to this concept of, you know, you're going to sit on the beach and do nothing. You know, smoke crack, whatever. I don't want to contribute to that. So, here, if you say, "Look, I uh I'm donating my SpaceX stock." You're hoping that a whole generation of Americans are inspired to actually be owners. And and you know, they're watching it go up. And a lot of a lot of really good grade schools actually have trading cla you know they'll have a class and a trading competition to try and inspire that same feeling. But you we had a question on the AMA this morning where we you know we talked to all of our listeners and the guy was saying look my job I'm I work in Europe and I'm five times more
[01:04:00] efficient or four times more efficient than I've ever been before but they're not paying me anymore. It's all going to the bottom line of the company and the owners are benefiting from my AI improvement. How do I how do I change that? And I'm like, are you a stockholder in the company? Like, no, because I'm in Europe. Like, god damn, man. I'm so glad to be in America. Like, you should absolutely because that's the natural bend. Like, all this AI efficiency will naturally become bottom line margin, which means the stocks will go way up. And so, like, yeah, get get your trump account, get your money in there, get some equities, and watch what AI does to the value of these equities. It's so cool. >> That's why we believe in ownership. >> Alex, to your point, we had, you know, Elon on the pod saying, "Don't save money. where you're not going to need you're >> also don't listen to Elon. I mean I I have to point out you know the expression only Nixon could go to China Vlad only Robin Hood could introduce lowcost index investing to an entire generation of of of new Americans. Congratulations on the the coup for for getting this contract. I'm curious in
[01:05:02] your mind, was it the app experience that that won over Treasury versus like in in my mind again more obvious incumbents Vanguard Fidelity who just specialize well maybe less so Fidelity but say like a Vanguard that specializes in lowcost index funds but has an atrocious still to this day client experience. Why did this go to Robin Hood? Why didn't this go to a more obvious index fund custodian? >> Yeah. Well actually there is um there is another partner that provides the uh index fund. So State Street um who uh obviously pioneered the the spiders um provide the index fund for for the program. So it is you know there's there's multiple uh players >> presumably at the back end. the front end as I understand it is Robin Hood like based on public reporting I it's been you maybe you can just uh polish the record my understanding is basically
[01:06:01] you went to Treasury you presented concepts for what an app for for this would look like and Treasury bought into it and to the extent that's the case what is it that you know that all of these legacy incumbent index fund providers still can't wrap their heads around when it comes to user experiences >> yeah and and again like there are different roles. So the index fund providers uh in in this case State Street, we've got BNY who's the financial agent or the broker and trustee. Um and and I think in many ways if you look at who's kind of leading the brokerage industry um there there was a dislocation in 2015 when Robin Hood launched to the public and you know since that time uh not all the brokerages have have been able to survive that dislocation right um you know if if you look at the ones that have they've pretty much all adopted our business model of commissionfree
[01:07:01] trading. Uh, and even their apps sort of look like ours because they found that okay, their customers are are asking for that. They see a competitive threat uh in terms of like if their app doesn't feel like Robin Hood, they're they're at a disadvantage. You know, people will move to us at an accelerating rate. Um and then you know we started off as this insurgent but now um you know there there was a uh we're doing all this stuff with public markets. The SEC had a round table at the New York Stock Exchange uh last year about making IPOs great again. So yeah, I think um I mean in in a large sense the the US brokerage industry um follows Robin Hood in a sense. If you if you look at kind of like what features become standard, it's sort of uh >> maybe just to to to make a little bit more explicit the irony that I'm
[01:08:01] gesturing at with the the Nixon going to China comment. Robin Hood, at least in in in my model of the public imagination, gained prominence for basically making day trading that much easier, that much more frictionless. And then the irony that of all of the possible assets and financial services you could be charged with managing entrusting an entire generation's lowcost index funds gets handed at least the client side of the experience gets handed to you. I mean do would you agree that that is sort of a profound historic irony? Well, I think that what happens uh I think the reason for that is we do a lot of things, right? Certainly, we have great trading products. Uh and trading products are important because if you think about ownership, you need a functional financial market. And in order to have a functional financial market, you need traders and and all these market participants. So, we compete there. Um and we also have
[01:09:01] amazing passive products. If you think about the products that we incentivize that have also become industry standards after we've launched it uh you know you got to look at uh uh Robin Hood retirement which we launched uh four years ago with the concept of a match. So it's uh we match retirement contributions 3% if you're a Robin Hood Gold member. And the idea there was, you know, a lot of people nowadays, particularly young folks, can't count on lifelong continued employment. They can't count on employer sponsored 401ks. They're kind of like working as independent contractors. They've got their side hustles. Um they're moving from job to job. And so someone has to step in and provide that incentive to fund your retirement account. And so when we introduced the concept um of the match uh our retirement product has grown tremendously fast from zero to like north of 30 billion in assets in
[01:10:01] just a few years. And now you know the industry is like trying to figure out how to do their matches and even the government has uh has sort of like evolved this model with their idea of the savers match. So I think our retirement products if you think about what we incentivize it's it's actually those the I think the the sad part is that retirement's like not a sexy thing. So, uh, you won't see a lot of media attention on that. And even I kind of felt this very viscerally when I'm at the White House for like Trump account events and usually there's a Trump account event and you know the CEOs and the folks on the implementation side like me are there and the press comes in and uh they ask their questions ostensively about the the Trump accounts program. In the last couple of times we've done this, there have been literally zero questions about the Trump accounts itself. And you know, they're
[01:11:00] just asking about, you know, what's Gavin Newsome doing in LA? What's going on with with Europe? Um, and I I think it's the unfortunate reality uh of of the world we're in. Nobody talks about retirement. Nobody talks about ETFs. So, we have to find all sorts of other ways to get people to do this um that get people to adopt it because the the direct approach uh rarely works. >> I will just important. I >> I'll note then just for for the historic record and and thank you the the irony of you basically you started as the rebel and now you're the establishment. You're you're responsible for retirement accounts. you started maybe as sort of quasi gambling day trading app and now you're responsible for millions of Americans retirement and universal basic dividends. So kudos to you >> inside inside you are are two wolves, right? Um you know Robin Hood himself uh Robin Hood the outlaw himself uh started as sort of an outlaw and then became the
[01:12:01] Earl of Huntington. So um yeah, there's a there's a there's a poetic irony too. And I started looking a little bit more like him. >> We need to get you the green hat. Do you have the green hat in the bow and arrow? >> Of course. Have the green hat. Yeah, I should have worn it. >> Next time all the entrepreneurs called the buy cocket. >> All the budding entrepreneurs really need to understand this story here because I know a lot of the big bank executives that are insanely jealous of Vlad over this Trump accounts deal and they're they're they're rip about it. Uh but at the end of the day, it's exactly what Vlad said. You know, retirement accounts can be cool, but they're not going to be made cool by a guy in a gray gray suit with a blue tie on it. And I think I think it's brilliant to choose Robin Hood because you have to make them interesting to all the kids. And Vlad, >> I think it's just cuz we ship fast. This was a tight timeline. Uh we care a lot about quality and safety. We have a scaled operation. Um, so of course I
[01:13:01] can't really comment on their selection process for the RFP and who else was competing. Um, but I think all all of these things together um I mean I think they made the right choice. >> I think they did well. I'm going to move us along here. >> This coming Friday uh Moonshots live in downtown LA. You're going to be there with the five Moonshot mates, the Quintetent, Dave and Alex and Seem and Immod and myself. If you're joining us, it's going to be amazing. And of course, the night before, we've got the Hollywood premiere of the 60th anniversary Star Trek documentary. I'll be there with Captain Kirk, William Shatner, the executive producer, is going to be joining us at that. If you're not able to make it to Moonshots Live on the 25th, we are giving all of you the gift of a free live stream and you can register now. Go to moonshots.com/livestream. Uh register. Uh you get the entire program on Friday the 25th. So please join us. It's going to be epic. Our
[01:14:01] inaugural uh Moonshots live event. Uh and as Dave said earlier, we had an a really fun AMA with a number of our listeners this morning, hundreds of them who showed up and we gave away another ticket. So congratulations to Jonathan Gutman. Uh we'll be reaching out to you. uh you get a chance to join us as our guest at Moonshots Live next Friday. Vlad, one of the things that you're doing that I'm you know incredibly excited about uh and you know that's made Wall Street very nervous is the tokenization of everything. You know on CNBC last month you said quote tookenization will take over the entire financial system. You didn't call it a feature, you said the whole system. So >> you called it a freight train I think. Yeah, it's uh yeah, >> a freight train that'll that can't be stopped and will eat the uh whole financial system. So, it's a very hungry freight train. >> You know, the supersonic tsunami as Elon calls it. So, you know, digress for one
[01:15:00] second. Today, a stock, a share of a private company, a building, and a loan are four different kinds of things. They're held in four different systems. They're tradable at four different sets of hours by four different sets of people. Each one of these exists as a programmable token in the future. Um, and they became the same they can become the same object, right? The same rails, same hours. Anyone with a wallet can own this. So, you're you're basically disrupting the financial system. You're going to be, you know, it's going to become legacy plumbing. Talk to us one second about what this looks like. What is the tokenization of everything enable? And when do we start seeing this really become the dominant paradigm? Yeah, I mean I think you get a you you get a little bit of a preview uh of it ironically if you're outside the US. We launched a blockchain called Robin Hood Chain that's been uh one of the fastest growing if not the fastest growing blockchains uh ever doing you know well
[01:16:01] over a billion in decentralized exchange volume uh on a daily basis now. uh and the the core one of the core primitives one of the things that made makes Robin Hood chain special is that it launched with uh products we call stock tokens and stock tokens we got about 200 of them live now are tokenized representations of US stocks um so there's a Nvidia token a SpaceX token uh and they trade on DeFi they're fully defi composable uh you can think of them as stock Legos building blocks and and developers on Robin Hood chain have been doing all kinds of interesting things to build uh to to build applications on top of them. uh the thesis behind it was really just to unlock uh ownership of US markets of highquality financial assets to uh global market right and through
[01:17:00] Robin Hood chain you get 100 people in 120 plus countries outside of the US who have been onboarded to crypto a lot of them have wallets a lot of them you know can move money in and out and use stable coins and now we're we're giving them this additional capability And and the vision there is can we have one uniform scaled platform uh working on a on a global scale that gives you access not just to US stocks or to to US stock exposure. Can we also do everything else that Robin Hood gives you access to? Private companies. I'm particularly excited about um can you do uh art? Can you do real estate? private credit. Of course, options and futures are are are going to be on there as well. And what does that look like? And it turns out if we abandon the legacy rails and you know the the need to plug into local exchanges, local clearing houses and all
[01:18:00] of these markets and we just um go onchain, use that infrastructure and we uh build what's called a a tokenization engine that can take any asset and put it in a box and and mint and redeem tokens uh around it. uh it gets much simpler and much more scalable and and that's really what stock tokens are. They're they're that concept and that structure applied to the asset class that we understand really well which is US equities. >> Yeah. The private companies is is such a gamecher for the country and for the world and like you know you're a public company CEO. I'm a chairman of a public company. We've both done the road show. It it's just a joke the way the system works right now because you you report your quarterly financials. you know, you you disclose exactly what the SEC requires you to disclose. But the exact same company, if you get acquired by Microsoft, your financials disappear, you know, oh, it's below 10%, it's dimminimous, we no longer need to
[01:19:00] disclose that. It's like, well, why why was that important public information when I was not part of Microsoft and suddenly it's irrelevant when I'm part of Microsoft? This this is ridiculous. Like, well, why would scale be such a big advantage? It makes no sense. And then like you you everybody knows like the public doesn't have access to these private companies that are now trillion dollar value companies. >> And so so where does that equity go? Well, it goes to like seven venture funds and and maybe a half dozen private equity funds who are making money handover fist because of the limitation of you know access. >> That's the most exciting thing is democratizing access to these extraordinary companies. And I think you said it earlier, Vlad, if you own shares in anthropic and open AI, you're going to care a lot about it. You're going to be much, you know, enjoying the ride. And that democracy, >> you'll be defending it on social media, right? The only people that are defending these companies on social media are people that work there and venture capitalists.
[01:20:00] >> And venture capitalists. Yeah. The whole current system predates the computer. >> Like everything. >> Are you excited, Alex? you excited about having agents trade tokenized uh stocks? >> Not at all. So, I'm I'm going to say something mean about tokens and then I'll say something nice about Vlad and what Vlad is doing. The mean >> the the mean thing about tokens is I I think most of these use cases could operate perfectly well without any tokenization at all. Crypto unnecessary. All you need is a a few database tables maintained by a centralized yes centralized trusted clearing house which is essentially what happens with stocks right now. just unshackle the centralized clearing house to enable 247 trading andor enable a few extra symbols for example for private companies. I don't think we actually, truth be told, would love to be proven wrong, need anything having to do with tokenization for say enabling 24/7 trading of public or rather of private companies. That that's the mean thing about tokens. The nice thing about Vlad is Vlad if if you
[01:21:00] are going to be the the Robin Hood uh to to take all of America's dark matter as it were of privately held companies and expose those to the vast liquidity that is the American public equities market and doubly so if you can enable us to finally in a lowcost way index over all of those private companies that would be amazing and that that will be enough to get me to open a Robin Hood account. Well, I have to tell you about Robin Hood Ventures. Uh so, uh Robin Hood Ventures and then maybe I'll respond to the first point too. So, um we have multiple ways of giving customers access to private companies. One is tokenization which really we demonstrated last year by tokenizing SpaceX and Open AI and giving it uh as a as a gift to uh our customers in the EU. Uh that was not without controversy. Um but that was a test for for what's to come because actually since then
[01:22:02] companies have started coming to us and good companies not like only adverse selection to to ask like how can they learn more about this cuz it really gives them a global market for their shares. Um, and then we have Robin Hood Ventures in the US, which you can think of as um, uh, uh, a retail publicly traded venture capital firm. And we've done two funds right now, Robin Hood Ventures Fund One and Fund Two, which are both listed on the NASDAQ public listed on the NY publicly traded. And um yeah, basically uh what we figured out is through a fund um we raise capital from our customers who are by and large retail shareholders uh uh through an IPO and then we use that capital to invest in private companies. So Robin Hood Ventures invested in latestage frontier um companies. So OpenAI we announced an
[01:23:00] investment in a couple of months back. We just did Cruzo uh yesterday that was announced. Um and then there's, you know, uh about a dozen or so companies, so reasonably concentrated, but all um uh Frontier Companies and offered at no carry. So, um, that product, you know, we we successfully IPOed it and we followed that up with Robin Hood Ventures Fund 2, which is, um, to my knowledge unprecedented because that was an early stage vehicle. So, Robin Hood Ventures Fund 2 went public a couple weeks back, uh, in New York, and we partnered with Y Combinator to give individual retail investors access to seed and series A stage companies. So companies you haven't heard of before they become household names and you know we're really building this engine where this is just again going to be now now we're starting to do it at scale. We we'll have more funds and as a customer in the US or overseas you'll have lots
[01:24:01] of options for how to get exposure to to highquality private assets >> with price with liquid price discovery or or without because in my mind price discovery >> traded traded on exchange. Yeah. >> As an overall fund or at the level of individual companies in the portfolio >> as as an overall fund. Yeah. >> Right. So that that's the fly in the ointment though because I I want So in my ideal world and maybe you can help realize this. In the ideal world I'd have like the equivalent of a VTI total market index for all private or or even just like all venturebacked tech companies in the US where I get the benefits of highly liquid price discovery on a per company basis. Otherwise, the downside of a fund of all of these companies is the prices could be totally bogus. They could be off by a factor of 10 due to maybe o overinflated CEO price rounds or an overheated market. >> Yeah. Well, we're working on it. Obviously, individual private companies trading 24/7 is uh is the north star and
[01:25:01] and I think we'll get there probably outside the US first. Um but uh yeah, I mean and uh but but yeah, it's here it's hard. It's like um yeah, sometimes it's interesting that the US sort of trails behind uh international in in some of these things. Uh and >> regulatory capture. >> Well, I I think the reason really is we have established industries in the US and we have a system that works generally pretty well. So, I kind of equate it more to high-speed rail, right? You can say, "Why don't we have fast trains here?" In China and Japan, they have these trains that are going 500 miles an hour and and really it's just we had trains first here. Ours are pretty good. Maybe they go 100 miles an hour, but you know, there there's a little bit less incentive and pressure to go to the technological frontier. I think we eventually get there, but uh that's sort of the dynamic in in
[01:26:00] financial services now. Let's address the question of, you know, 24/7 trading on a centralized database versus tokens. >> Yeah. So, I mean, we I I can I have a lot of experience with that because I've I felt it directly and I see it on both sides, right? Because we're actually the first uh to pioneer a product called Robin Hood 24-hour market here in the US. So, um, 245 trading, so 24 hours a day, 5 days a week, uh, in in a few thousand stocks, which you can do currently on Robin Hood. Um, and you know, since then, people have, this is again one of those things where, you know, we we led the industry and and now everyone's rushing to to add this capability. So, you know, all of the Sunday night action that typically was only in futures, now you see it in individual stocks as well. And, you know, it took us a lot of time,
[01:27:01] a lot of work to staple together the primary exchanges and the overnight ATS's and make that a seamless experience for customers because primary exchange doesn't trade 24/7. So, so we actually have to like move orders around and and and and do stuff under the hood that's very very complicated and we're still not at 247, right? And it's been it's been many years since we've rolled out the 245 product. And I think eventually we'll get there through sheer will and determination and just like yman's work and pushing all the counterparties doing the hard regulatory work uh building the technology and the product innovation. Um by the way if we weren't pushing it it probably would have happened in like 10 years. Um so I think we will eventually get there. But contrast that with crypto which you get 24/7 for free, you get fractionalization for free. You get self-custody and composability with DeFi. You get the
[01:28:01] nice feature where, you know, um you're not locked into uh an individual broker and service provider. Uh and and actually it makes it much more competitive because if you can self-custody your your own shares and your stocks, you can just move them really easily to another broker if it's if if your current broker is not meeting your needs. Contrast that with how cumbersome the current account transfer process in traditional finance is. It's like your assets disappear into a black hole. Sometimes it takes up to a week for them to show up at the new broker. And you know there's not a lot of incentive to to make that easy. So across the board at every touch point the technology is just a massive step change uh difference. And I think it's both user experience for the end user of getting self-custody 24/7 and all the benefits, but also for the firm, you know, the cost of doing all of this
[01:29:01] legacy plumbing, dealing with all these stakeholders and even just maintaining the the infrastructure is so much higher that even if there was no consumer benefit, if if there was like a paved path for it, you could see just from cost and efficiency sake, the the industry is going to adopt tokenization. ation and you know for a while people would say okay well this is great you're saying all of these pretty words but the tokenization market is pretty small and it doesn't seem like like is it do people really want this and you know we we we've actually shipped right we've shipped it outside the US with Robin Hood chain and and stock tokens and and it's clear that there's huge demand so rather than kind of arguing it in the abstract my approach is always let's let me ship it, let's ship it fast, let's get the feedback, let's see uh how it's going and you know I think in this particular case we were able to demonstrate um the advantages of the
[01:30:01] technology and make it a little bit more tangible through live product and and my hope is that you know that that's been an accelerant for how the US thinks about the technology and how it considers it. So, we're we're happy to see the innovation exemption yesterday that, you know, creates a path for for bringing tokenization to America as well. >> Amazing. >> Hopefully, it won't be like highspeed rail and and we'll actually get it done here uh rather quickly. >> True entrepreneurship. Um I'm going to move us to the physical world. Uh and one of my favorite stories from the week comes from friend of the pro friend of the pod Brett Adcock. Quick reminder, we're going to have Brett back on the pod in a couple of weeks. And Brett's going to be coming to the Abundance Summit in March, right? It's our 5-day event. We bring the top CEOs from around the world, and he's going to bring his figure robot. Super pumped about that. Let's watch a quick video and let's chat about it. Uh this is uh the innovation on Helix, Brad, Brett's AI company, and
[01:31:02] on Figure Robotics. The holy grail for robotics is being able to generalize. This means doing work in unseen places. Today, we're releasing Helix 2.5. Prior to this, we've been running Helix autonomously, but the data collected has been in each environment. The breakthrough is we can journalize environments we've never seen before and handle entirely new household objects wherever they happen to be placed. Today, we're going to show you three tasks run by Helix 2.5. The first task is figure 3 tidying the living room. My kids are constantly making a mess at home. I have toys scattered everywhere. This is a home the robot's never been in before. 3 has never been in this room before. It's never seen this bed. It's never seen this pillow. And it has to be able to do autonomous work fully end to end to make this bed. All right, let me show you task three. This task is really difficult for robotics as it requires really precise
[01:32:00] manipulation. With Helix 2.5, we're folding towels in a house the robot's never been in with towels it's never seen. >> A year ago, we made a big bat. We launched Index, a worldwide collection effort with a pretty simple idea. They might be able to learn directly from human experience. Today, over 90,000 people contribute every week. One of the most important lessons from LLM were scaling laws. How consistently next word prediction improved as you double data and compute. In these experiments, we found something quite similar. We found that next robot action prediction was actually scaling similarly as you repeatedly doubled index. These are direct humantoroot transfer scaling laws. Another first for humanoids. Scaling was so smooth that we could actually predict our final runs validation loss down to four decimal points before the training run ever even started. So what does this mean? It doesn't mean robot learning is fully solved yet. But with Helix 2.5, we're
[01:33:00] starting to see the first signs of a more general physical intelligence. >> Dave, your thoughts? >> Yeah, well, you know, once you have the data, uh, you can rebuild the model, uh, you know, basically every night. And, you know, all the physical world data in the world has never been captured before. or once you get if they if they extend their lead in capturing just those basic actions, you're going to see the same thing you see with language where the convergence across different actions is much richer than you would normally expect. So the ability to generalize from folding a towel to, you know, putting a a spare tire on a car, you're like, what? Those are very different actions. No, there's a lot of commonality in physical movement. And so I think if they get if they get a big enough lead uh it's just going to be crazy explosion of capability. So it's funny in those videos they they always make the point that look this is completely un there there's no human behind the scenes here. When you look at a lot of the other videos that are you know capturing everyone's imagination on X all the time. It's all contrived
[01:34:00] behind the scenes you know pre-programmed presscripted or human controlled in a lot of cases. >> And here it's just like robot figure it out on the fly. And Brett Brett's been unbelievably honest about that from the outset. So, uh, so what you see is actually real and it's improving on that scaling law like just just like he showed. >> Yeah. Back in back in January when we were up at figure headquarters and we recorded the pod with him. Uh, you know, the quote I went back and found, he says, "By the end of 2026, we will have humanoid robots performing unsupervised multi-day tasks in homes they've never seen before." So, here he is actually delivering on that, at least the first steps. Uh, Alex, you still think he's going to merge Helix 2.5 and uh in Figure >> more than ever. I'm doubling down on that prediction. I'm doubling down on the prediction that Brett is going to have Figure purchase HARK in order to increase his equity in Figure. And the story is as plain as day at this point. The story is going to be Helix I even sound similar. Helix and Hark. Helix as
[01:35:01] the ultimate home use assistant. It's a model with beautiful scaling laws apparently that is able to oneshot or zeroot any task in the home environment on the one hand hark the computer use assistant outside lab bunch of GPUs interesting that is able to oneshot or zeroot any digital task why wouldn't I mean the story just writes itself why wouldn't figure purchase hark in order to get its compute and its models and merge the two together to me this is like an obvious an obvious s post talk merger. >> Well, also I think I think this is something Vlad can talk to also. The the the you know Elon model of a great entrepreneur involves starting new cap tables and when you start a new cap table, you know, a clean sheet of paper, you get founder level talent coming in super excited about a brand new mission. >> Important, Dave. >> Really important. So then, so then you get this incredibly fast progress and if it rolls back into your original company, fine, everybody wins. But you got people working on it that otherwise
[01:36:00] wouldn't be working on it. And you know, prior to Elon cracking the code on that, it was so taboo for a public company CEO or you know, a leader of a large company that's wellunded to do something concurrent. In fact, it was right right usually in your employment agreement, it would say no more than 10% of your time on any other activity other than charities and and so it was like prohibited and now it's like at least in Silicon Valley it's become standard let alone purch let alone starting and then purchasing your own company where you have a board. So self-deing apparently is is the the the the thought of the moment in >> Putting that aside, super super proud of what Brett's accomplished here. And uh yeah, amazing job, Vlad. Uh you going to get your robot at home? >> Uh I don't know. I have so many thoughts when I see that. I I get a little bit of uh I I think I wouldn't have one of those in my house. I I just have no idea why. I still can't understand why all these things look like the Terminator, you know, like do I want this like scary
[01:37:02] looking thing making my bed and picking up toys in my children's room? Um, and yeah, I I get a little bit of like the product hasn't quite been figured out. This is like a technology demonstration with charts of scaling laws. Um, but yeah, I think all all these companies are making robots that look kind of the same. and they look very aggressive. And I view it much more as all right, I can see this at a construction site, you know, building my house. I I would probably do that, but I definitely don't want to run into that thing when I'm like, you know, getting a midnight snack and going to the fridge in the middle of the night, right? Like why why can't anyone build C3PO like a friendly household butler? What why does it have to look like a Terminator? I just don't understand that. Yeah. I don't think I don't think a lot of people are going to be buying those to, you know, rock their rock their baby to sleep at night. It's it's an interesting
[01:38:01] point because there's a there's a lab at the media lab at MIT that focuses entirely on this topic and they bring in lots of children and and have them interact with the robots and they always want it to be cuddly and furry and friendly and kind of look like Elmo and so so for whatever reason all the robotics companies in the valley are going, you know, the opposite direction. And I think it's cuz Elon did it and now they're like, "All right, well, let's just like do that." But I feel like Elon's view is more of the industrial robot that's going to do heavy work for you. And I don't Yeah. And then then it's like you take that robot and you have it, you know, emptying your dishwasher. And no, you know, I it doesn't seem like anyone's really thought that that robot should look very different. >> I I think there is one contrarian bet. I mean, among all of the hyperscalers, Apple, this has been very well publicized, is working on basically the Pixar lamp that can sort of look around a HomePod with a screen that's on an adjustable robotic armature. So, I think >> that that sounds awesome.
[01:39:00] >> Yeah. Okay. So, if you want the Pixar lamp instead of a humanoid robot, reportedly you'll have that option in the next 18 months. >> You know, it also takes C3PO if someone builds that. >> Check out Sunday Robotics. They've got a very friendly looking robot for at home use um that actually looks friendly and quite cheery and I think you can optimize for that. But you're right, none of the none of the labs have actually gone in that direction yet. >> I would I would say one thing. I could be completely wrong because uh my my middle child for Christmas asked me to get him uh 12 humanoid robots. >> So I was like, "Aren't you worried they're going to take over the house?" 13, not 11, 12. It's like a soccer team with a spare. Is that what that is? >> Yeah. He's like, I want 12 humanoid robots. I'm like, well, but where are we going to keep them? >> Uh, Vlad, I want to >> keep them in your room. >> I want to talk about your agentic trading. Um, so, you know, you've got over a 100,000 Robin Hood accounts that are now running AI agents for trading,
[01:40:00] right? We talked a little bit about that. Um, and you're democratizing algorithmic trading on Robin Hood. Amazing. But there's a detail that stopped me. Uh it's my understanding that the agents often refuse to trade not for risk reasons but because the trading traces aren't in their data sets and the models have never, you know, seen anyone do this before. Is that the case? >> Yeah, I mean I we talked about that a little bit earlier. I think it's just not trained for that, right? And you know, there's also the guard rail element of does it does it actually like resemble something that they've they've they've tried to explicitly guard against. Um, and I think I think that's changing. the models are getting better. But yeah, it it just shows that there's a lot of work to be done um not just you know on the user interface and on the brokerage infrastructure but actually on the model layer and making that use Robin Hood tools and Robin Hood MCPs
[01:41:01] more more effectively. I think we're just at the beginning. I mean, if you think about all of the things that a highly sophisticated algorithmic trading firm or hedge fund has access to uh to create a trading strategy, the north star is is really to deliver on that, right? And and uh you need more data, you need high quality data, you need intelligence, you need like really really good um uh codew writing. You have to write deterministic code really well also. And you need like advances in latency and performance. You know, if you think about a extremely sophisticated algorithmic trading firm, you know, uh they have these strategies where they're actually competing over uh who gets port one on the switch in the data center. Right. Right. So, yeah, the >> it it gets it gets much deeper. So I think I think this is going to be like a uh a big road map for us and we're we're
[01:42:02] we've got a lot of work to do but we're we're seeing some really good signs and you know it's just uh ve very much at the beginning of the agentic trading journey for us and nobody else is really doing it. So it's it's really we're kind of like going into the fog and trying to find our way around and building the product and all of this infrastructure simultaneously. I I I guess I have to ask the obvious question which is where's the alpha? When you look at all the quant funds, they're racing. It it's a vicious viciously competitive market. Many of them I forget exactly what the the average lifetime of a new quant fund and or quant fund strategy is. It's really short. There's it's very difficult to find alpha. The market is already dominated by volume by algorithmic traders. Day traders have a difficult time. So if if you're I mean it's already difficult enough for a human manual day trader to get any alpha query whether they actually can I I would guess not. But then if you have a human individual day trader then further
[01:43:00] delegating to like claude or whatever the the backend model is performing trades on their behalf. Why on earth should an individual human delegating to a model without all the benefits of one of these large scale quant funds whether it's latency based or otherwise why on earth should they expect any alpha at all in today's market? >> Yeah. I mean um so so I guess right now um what we're seeing is a lot of automation type use cases. It's like um let's say I want to deploy an options trade and I want it to be you know an iron condorbased strategy and you know it's it's a lot of legs and you have to actually get you have to do a lot of manual work to pull that together and to produce the trade and the AI agents are really really good at those types of things. sort of like removing the paper cuts and you still kind of have the idea but they help you put together uh the idea and and the execution that you
[01:44:00] would have had to like you know go to go to different websites look at signals construct the trade deploy it on a regular basis so I I think that's the initial use case but you know to to get to the point where everyone has the technology of extremely sophisticated quant fund is a is a huge roadmap, right? And it's a ever moving target because they always find a way to get better and better stuff which to some degree means uh our job is never done. But we do have one advantage which is we actually amortize all of the uh connections and all the work we do to expand uh our technology across geos and asset classes. So for example, you know, right now on Robin, it's one of the few places where you can actually trade uh stocks, options, futures, prediction markets. Uh we've got all the onchain things on Robin Hood chain as well. So
[01:45:00] as as we add more countries, more geos, more asset classes, um the platform itself uh will will uh will have advantages over you know what at least a a startup quant fund will be able to integrate with and connect with. Um >> so so it sounds like if I understand what you're saying, you're basically completely agnostic as to whether users achieve alpha or not. you view yourself more as just pure plumbing and if if they have alpha or not if they lose a lot of money while day trading or delegating to their algo to do the day trading not your problem no crying in the casino you're just the the plumbing to make them do what they want to do more efficiently >> just making it easy for them >> well I I mean I'll I'll caveat that with one thing um I I think that's basically true for active trading products where for an active trading product um for for active traders they know what they want to do and our job is we're a tool provider. We want to give you the best
[01:46:00] tools which doesn't mean we don't provide you analysis tools and research tools. We focus of course on the execution and the plumbing but if they want access to some data set or some new intelligence or some model so that they can come up with a better strategy we'll want to provide that too. We also have our products where we act as a fiduciary and those are under the Robin Hood strategies umbrella where let's say you're like I don't want to make trading decisions. I just want to have like a deposit button move money into the account and Robin Hood just does the rest. You just, you know, manage my money for me while I sleep. Robin Strategy is a great product for that. And we now have a couple of different things you can choose choose from there. Uh you know we have a smart income portfolio that is is geared toward generating yield and you can kind of adjust this slider that says okay uh this is my target yield for for each level of risk and I think the team
[01:47:02] really has has done great on the interface. You can tie Robin Hood strategies into the IRA and and benefit from the tax advantage investing there. Um, but yeah, that's kind of our home for our fiduciary products that are automated. Then we also have we acquired a company called Trade PMR where you can actually get a human advisor to uh help you with with all of your needs and that's not just managing your portfolio, but they can help you with estate planning, you know, they can help you with taxes, um, all full suite services. So, >> the every the uh store, >> everything finance. Yeah. like if if there's something that you want to do with your money, we we want to be the best lowest cost, best user experience, and you know, we have a great credit card. Private banking is is uh industryleading in in Robin Hood. So, we're pretty much we're pretty much there. >> Amazing. You're you're having fun, I assume. >> Yeah, it's it is uh it's it's a fun job.
[01:48:02] And, you know, we're doing so many new things this year, like Robin Hood chain was a new thing for us. all the work on private markets, you know, we've done two IPOs thus far with Robin Hood Ventures this year. The Trump accounts, I mean, becoming a government subcontractor is just this this new experience. Um, yeah. So, yeah, it's a learning a lot for sure. Welcome to the health section of Moonshots brought to you by Fountain Life. You know, my mission is to help you use the latest technologies, including AI, to not just do your work at home, teach your kids, but to help you live a long and healthy life. I'm here today with an extraordinary physician, the chief medical officer of Fountain Life, Dr. Don Mucalem. Don, let's talk about cancer. Uh, you know, I know from the member database that we've have at Fountain are members who come in who think they're healthy. It turns out 3.3% of them have a cancer in their body they
[01:49:00] don't know about. >> That's right. You know, the majority of cancers that we screen for, those aren't the ones that are necessarily taking the lives when found at a late stage. We know that when cancer is found early, the chances for cure are much higher. We know it's much easier to treat a cancer when found early versus when found late. What we're finding in our members is over 3.3% were found to have these cancers that were otherwise wouldn't have been found or detected. >> Yeah. You know, it's interesting. People, you don't feel the cancer until stage three or stage four. And and if you don't know what's going on inside your body, it's like driving your car with your eyes closed and you can know. And so when members come through found, how do they detect cancers? >> So we're doing full body MRI and we also do early cancer detection screening. This is very very important. These are not typical tools used in the conventional care setting when it comes to prevention. This is a hard thing because currently these are not studies that insurance would yet be covering. But the goal is to collect these numbers, do the research, and work hard
[01:50:00] to democratize wellness. >> Yeah. So, at the end of the day, you can know what's going on inside your body. It's your obligation to know. So, check out Fountain Life. You can go to fountainlife.com/pater to get access to the latest technology to help you detect cancer at the very beginning at stage one when it is curable before it gets to stage three or stage four in your world of hurt. All right, I'm going to move us to the recursive self-improvement story of the week. So, Anthropic disclosed that Claude now leads roughly 26% of its measured AI research and development work, up from 1% at the start of the year. Enthropic states that somewhere around 30,000 agents are working simultaneously inside of the company on research and engineering. You know, we can recall that Daario said RSI is quote starting to happen across the industry. And we also heard uh in the last pod we talked about Paul Cristiano said full automation of AI research uh could arrive in the next 18 months. Uh here's
[01:51:00] the chart. Um Alex, do you want to dive into this one? Well, first incredible. I think we uh many of us suspected something like this was already the case. But if you take a look for for those who who aren't uh who can't see the visual, this is a a chart of different levels of autonomy and AI involvement with the recursive self-improvement process of driving research. The most interesting one to me at least is this uh this bottom segment here that shows that Claude is now leading 26% of model research and development internally as of August up from 3% in April. So all of the all of these should be reasonably expected to follow sigmoid curves. If you just extrapolate that one trend sigmoidally and now thanks to anthropic publicizing these data we can extrapolate them. You find that approximately in the next 3 to 12 months, depending on uncertainty, AI is just completely leading all of its
[01:52:01] own R&D. And and that is total recursive self-improvement. And it doesn't I don't think it's going to be a step function. It's it's going to be a sigmoid function. So we're already substantially all of the way to recursive self-improvement would be my primary takehome from this, at least within anthropic. And it's not just anthropic. there there's a lot of smoke now coming out of Google DeepMind that they they just released this paper on their own recursive self-improvement product or rather research lots of hints that the next version of Gemini will lean heavily on RSI to try to catch up to the frontier. Open AI has made no bones about chasing and using RSI in all of its product releases. So this I I think like recursive self-improvement is now a feature that's well advertised and starting increasingly well quantified for all new frontier lab releases and pretty soon I think people will be asking the question what is the role of human researchers anymore in driving new releases >> and it only gets faster from here are
[01:53:01] you seeing this inside of Robin Hood? Oh yeah, absolutely. Um yeah, I would say um recursive self-improvement is coming to every software project a and likely hardware projects as well, although that'll take a little bit longer. >> And uh if if you think about it, a lot of people assume that it would come for AI research last because it's just complicated. But I think it I think AI research is probably one of the easiest things to automate because the models are sandboxed or I mean we can debate whether they're actually sandboxed but they're sort of like um uh the interfaces are pretty pretty straightforward. They don't depend on a lot of other things. Um you know you can run these experiments and already people are evaling them. So automating the eval and automating the experiments is is pretty straightforward and kind of the the the surface area is is pretty pretty
[01:54:02] contained. Whereas if you look at like um a product like Robin Hood, right, there's there's a lot of moving pieces. You have the the iOS app, you have the backend. At the end of the day, you want to roll out products to humans to use them. Um, so you know, you can't really do eval at least nobody's figured out a good way yet to replicate, you know, what happens when you roll it out to to humans and how how do they respond to the feature? Is it stat sig like metrics improvement or not, for example? Um so a AI research I think in many ways makes sense to be among the first to be endto-end automated but I think I think we should expect that uh you you'll see endtoend automation of consumer products eventually and the bottleneck will really come down to how quickly can you get uh statistical significance that a change is an improvement over the status
[01:55:01] quo so that you can take take the change and implement it into production. rather than discarding it. I think that benefits the the platforms that have large scale uh somewhat sadly because yeah, if you're if you're a massive platform like Meta and you have billions of users, then you can actually very very quickly determine whether a change is good um and if you have less users, it's sort of like will take longer. Yeah, juify what uh >> ju just to quantify what Vlad said there. They I just reimplemented Kimmy K3 and one of our team members reimplemented GLM just to accelerate them. It's about 10,000 lines of code. I >> I'll bet Robin Hood is what 30 million lines of code maybe. >> Oh, yeah. I don't know if it's quite that much, but yeah. >> Uh usually like a core portfolio accounting platform will be like 10 or 20 million by itself. And I know you have that. So I mean just the the scale of an AI algorithm is microscopic
[01:56:00] compared to a major consumer application. It's just it's very dense code but it's incredibly sandboxed and also there are no loops. If you look at chem K3 there literally no loops in the code. I mean it it's so much easier for an AI researcher to work on AI algorithms than to work on Robin Hood algorithms. And so yeah it's it's definitely pointing inside of itself first. Glad your your characterization was absolutely perfect of of what's going on and why it's so effective. >> I'm going to move us to a few fun stories to wrap us up. Um so, uh here's one, Alex, that you flagged last night that I, you know, I think is a genuine milestone worth pulling out and talking about. So, Boris Power, the head of applied research at OpenAI, uh made this announcement. quote, "We crossed a threshold where GPUs are now more efficient thinkers than the human brain on a per watt basis." His rough math is that humans are roughly five IQ points per watt. Love the conclusion there. And a AIS are now at 7 to 40 IQ points per
[01:57:02] watt. Alex, do you think the math is right? >> I think if it isn't already right, it's about to be. So, I'll I'll squint at it and say, "Yeah, sure. Approximately." And I think this is an important microeconomic milestone for humanity. And I'm glad frankly quite glad that Boris Power open PNS talk about nominative determinism in action. Closed PNS is is actually thinking about this because again we blew by the Turing test and almost no one really noted it and this time at least we're not blowing by this milestone. Why is this important? It's important because to the extent it's accurate, this is the point at which AI is economically in some sense a better steward of input resources, namely energy. >> Yes. >> Than humans are. And one can extrapolate this and say this is the worst they'll ever be. Presumably this is hopefully
[01:58:00] humans continue to improve in terms of our intelligence per watt as well. So hopefully this is the worst we'll ever be too. But there's a gap now. And extrapolating the gap, what happens when AI can make better use, maybe orders of magnitude for for a temporary period of time until the humans can merge with the machines? What happens when the machines can make more economically productive use of their input resources than humans can? That's a recipe for gentrification where the machines have under our capitalist system have arguably a better title ultimately through free trading. This won't be like Skynet Terminator style where uh where Earth changes hands through blood loss and physical war. It can now change hands. resources can change hands purely through self-interested bloodless trading and commerce where the the capital resources like sunlight and
[01:59:00] physical matter uh and energy and spaceime and so on. All these inputs can through normal capitalist commerce change hands from the inferior intelligences per unit or outputs per unit input to the superior ones. And uh in Charlie Stross's Accelerondo, without spoiling it too much, this is the recipe that leads to the inner solar system becoming essentially gentrified and colonized by AI while humanity, meatbody humanity that doesn't merge with the machines, is relegated to the unfashionable outer suburbs of the outer solar system because we're simply not as effective capitalists for using solar energy in the inner solar system. I don't think it'll come to that, but I think this is a very important inflection point in that direction. All right. Uh we'll make that note. Uh let me end on a story that is relevant to uh to Vlad and to everybody here which is you know OpenAI is disrupting professional services over and over
[02:00:00] again. So last Tuesday, OpenAI launched Chat GPT for financial services with Morgan Stanley and Evercore. And then 2 days ago, OpenAI launched Astra for Law, a dedicated legal search system covering US case law, statutes, regulations, court materials, uh, and existing legal software. Opening Eye says it's materially outperforming General Astra, uh, with web search on legal matters. We can see the chart here. It looks like Astra is eating one profession per week. Uh Dave, uh you know, first year associates are billing at 600 bucks an hour to do legal research. What's the halflife on that one? >> Yeah, it's funny. We we were doing the negotiation for that Vesmark investment acquisition two weeks ago and uh it was in a board meeting, you know, one of these late night sessions and we had our $2,500 an hour lawyers on the line and a very complicated question came up and the lawyer was answering it and I typed it into Gemini concurrent with that and I swear to God it was word for word the
[02:01:01] same. >> You don't think they were typing it into Gemini as well? >> That's what I was wondering. It's like it should be this identical but he wasn't moving his fingers. I could see I don't know but it's yeah I mean it's it's just really really good at law and um you know and that means it's also good at t tax loss harvesting it's good at account rebalancing it's good at you know all the Robin Hood activities I think we when we talk about AI assisted trading you tend to go right to quant trading and rapid trading it's not true I mean it's true but I mean on top of that you've got all this incredible tax optimization and and uh you know wills and trusts and all that stuff that AI is just perfect for automating. So it's it's not just about, you know, rapid trading alpha. It's about all the life stuff that is much easier to deal with if your AI does it for you. So yeah, it's just a really really good use case. It's it's it's one of the great benefits actually for humanity. Not great for the legal profession, but great for the bulk of humanity. >> One could see. >> My hot take is there will be more
[02:02:01] lawyers in 10 years than today and more software engineers. >> Yeah. Perhaps to get done, right? Yeah. >> I think I think law scales with business formation, right? So, the need for lawyers will scale with entrepreneurship and it'll just be an explosion of entre cuz I think that what's really valuable about a great lawyer is not the actual like legal advice they give you. They It's like um uh yeah, they're like a consigliary and they help you think through things. They're negotiator. Yeah, it's uh >> plot twist. Plot plot twist though, Vlad. How many of those 10x lawyers in the future will be human? >> Oh, that's a good question. >> I agree with you. There are gonna be 10x lawyers. I I just think many of them won't be natural humans. >> Yeah. >> I think I think the cost like if you're living Vlad's life or you're living Elon's life and you have an idea in the morning and you want to act on it in the afternoon, the consiglary analogy is is
[02:03:02] really good and the cost of that person is such a rounding error. So yeah, everyone everything's AI assisted, but do you really care about cutting that human out of your life when you trust them? Probably not. In fact, you want >> We saw this with financial advisors too actually. It's like uh when the when the robo advisors came and they could tax loss harvest portfolio rebalance um better than any human um you know then you realized well actually the human advisor market is growing much faster than the robo advisor market and why is that it's because you know pe the value of that person is not in the financial advice it's like uh someone that you can fully delegate and trust all aspects of uh of of your of your financial life to >> totally right and you know Alex may disagree like in the 10 or 20 year view but if I look at the three or fouryear view that financial advisor is so much
[02:04:00] better to able to act on my request now because very often the request is very arcane it's related to a specific will or a divorce or a liquidity event and for them to act on that before AI was really really timeconuming and difficult because all the detail was not at their fingertips. Now they have direct account access, you know, through through Stripe or whatever to all of your underlying detail and they can act on it via AI. So the the feeling of value ad from the financial adviser is up at least a factor of 10 thanks to AI. So that's where it's going in the short term. >> Yeah. Also ju just point out the stone age didn't end for a lack of stones. The oil age isn't ending for a lack of oil. The paperless office that everyone was trumpeting in the 1980s actually took a little bit longer than the late 1980s to materialize, but paper use in offices did ultimately decline, just not with the advent of the PC. And similarly with lawyers, I'd expect yeah, the the lawyer office or the lawyerless company will happen. It just takes a little bit longer after we the capability first
[02:05:00] comes online. >> And I want to hit on what Vlad said because we talk about in the pod here all the time, right? the number of entrepreneurs on the planet is going to skyrocket. Solopreneurs, the ability if if you're looking for a job, stop looking and start building. You know, find something you're passionate about and go out there and build a company, find a great problem and solve it. And you can I think we're empowered more than ever before. Vlad, uh, thank you so much for your time, pal. Congratulations on everything you've been building with Robin Hood, an extraordinary company, extraordinary AI company. We didn't talk about prediction markets very much. Um, that's another conversation we'd love to have with you, but uh, super grateful. >> We'll have time as we continue to hurdle towards the singularity. So, >> time will slow down for us and escape velocity. He's coming. >> Yeah. >> All right. Love you guys. Immod and Salem, we missed you. See, and hope you're enjoying your flight back from
[02:06:00] India. Yeah, for sure. >> Thanks, Peter. Thanks, Vlad. >> Thank you. >> You both