Mira Murati's 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis' AI FINRA | EP #271

Mira Murati, the former OpenAI CTO, just shipped her first model. It's called Inkling. Customization over leaderboard dominance is what's going to win her the day.

She's built exactly the thing hitting the market that exactly what everybody needs right now.

I want to pivot to a discussion of liquid AI, the small language models, what they are, what they mean.

Our mission has always been building efficient general-purpose AI at every scale that explores the computational graphs of intelligence beyond Transformer and then figure out what should be that architectural design that brings the same level of intelligence than a frontier model into, let's say, on a CPU.

CEOs building the most powerful technology in the world are asking to be regulated. Demis Hassabis, CEO of DeepMind, he called for a US-led frontier AI standards body modeled on FINRA. When the incumbents ask for the rules and they set the standards, they set up a barrier for all the entry-level labs coming in.

Let's just be real. AI moves way, way too fast for any kind of traditional bureaucracy. How quickly can you do it is going to be a huge, huge challenge because—

Now that's a moonshot, ladies and gentlemen.

All right, everybody.

Welcome to Moonshots, your number one podcast in all things AI, your front row seat to the singularity. I'm here with my magnificent moonshot mates, our original quartet, AWG, DB2, and Salim, and a special guest, Ramin Hassani, co-founder and CEO of Liquid AI and a pioneer in small language models, which we'll dive into. Ramin, welcome. Where are you this morning, pal?

Thanks so much for having me. I'm actually in Spain right now.

In Spain.

All right.

Nothing going on in Spain this week.

God damn it.

Yes.

I'm struggling from yesterday because I'm a long-suffering England supporter. It was a very difficult game to watch. God, they were like, just had it, it was 6 minutes to go and they blew it. And Messi's a genius.

That's the round ball, right? Is that?

That's the round ball.

Now, Peter, this is where the Falklands War gets relitigated on a soccer pitch.

Oh God. You know, I just flew in last night from Zurich and I had the most painful experience, right? I don't know why every airline doesn't have Starlink. I'm suffering on some meager, thin pipe connection, and you're flying over the poles, over the, you know, the Northern Territories, and there's nothing. And I'm trying to get ready for this pod. I was like, please give me, give me some bits. So anyway, challenges.

I mean, you must have grown up on soccer, right? Didn't you? You were in Vienna for a while and getting your PhD and— or your undergrad or whatever it was.

Yeah, yeah. I mean, soccer has been like a big thing, you know. I'm Persian and Austrian, like, at the same time, you know. Like, it's a big thing for us. So yeah, like competition is something that, you know, it's extremely core to what we do even today, you know. So, and I feel like that's like one of the main drivers, like sports and everything, like has been part of our lives like from day one. And then getting into science, the same thing, you know, now getting into ventures, same things, you know. And that's what we're doing.

Just compete, compete, compete, compete.

I love it.

Well, are you there for a little bit of time or are you coming back soon?

No, I'm coming. I'm flying tomorrow, actually, back to San Francisco.

And Salim, are you jealous of everybody, of him being in Europe, or are you happy to be home?

No, no. 3 weeks bouncing around in 10 different spots, I'm very happy to be home right now. I was just in Spain roaming myself, so.

All right, guys.

There was a lot going on, actually, like in Europe, you know? So that's— same like you, 10 different places.

You know, Alex and I were just reminiscing the fact that Europe's sort of major advantage in the future is it's going to be a museum of the way the world used to be.

Ouch.

Ouch. But it is beautiful. There's no question. It is gorgeous.

All right.

I want to jump into our first conversation. We have a lot to unpack here. And of course, our mission is keeping you aware of what's going on in the world and giving you sort of the optimistic, hopeful vision of the future. Join us and keep up with the incredible pace as we head towards a singularity. So our first story today, once again, CEOs building the most powerful technology in the world are asking to be regulated. You know, last week Sam Altman published an op-ed in the Financial Times proposing a framework for a US-led international forum that would establish standards, provide expertise, impartial analysis and capabilities, and assess risks. This week, both Elon and Demis are adding their voice to the regulatory conversation. Elon's says he expects a standalone regulator similar to the FAA or FCC to emerge at some point because, in his words, the consequences of AI going wrong are severe. Then this week, Demis Hassabis, CEO of DeepMind, went further in an essay titled A Framework for Frontier AI and the Dawning of a New Age. He called for a US-led frontier AI standards body modeled on FINRA, the industry-funded watchdog that polices Wall Street. Under SEC oversight. He wants the FINRA equivalent to test frontier models before release. He reportedly wants this up and operational before the end of the year. Let's take a look at a quick video from Elon and then let's jump into this conversation, guys.

I think the general— I think it's clear that there's a strong consensus there should be some AI regulation, that it would be in the best interests of the people to do so.

And I think we'll probably see something happen.

I don't know on what time frame, um, or exactly how it will manifest itself. I, I don't know.

I mean, this, this clearly—

we've created regulatory agencies before. While our regulatory agencies are not perfect, um, and I deal with regulators on a very frequent basis, um, with the automotive, um, you know, communications, Starlink, um, and then, uh, FAA with rockets. I think the probability of there being some sort of AI regulatory agency that stands on its own similar to the FAA or FCC is likely at some point.

You think so?

I think so. Now, the reason that I've been such an advocate for AI safety in advance of sort of anything terrible happening is that I think the consequences of AI going wrong are severe. So we have to be proactive rather than reactive.

Amazing. So I— this is a conversation we've seen over and over again. And I think The government, the public, and now the CEOs want to be leading this. I like the approach that Demis laid out. Right. But the challenge we have to discuss is when the incumbents ask for the rules and they set the standards, they set up a barrier for all the entry-level labs coming in. Salim or Dave, do you want to jump in first?

I'd be very curious to know, Ramin, if— do they reach out to Liquid AI and say, hey, join this? You know, we're going to create a FINRA-like regulatory body. I mean, the reason FINRA works fundamentally is because people from the industry who know what they're doing are willing to join it. They're definitely not willing to join the government in general, but they're willing to do a year or two in a regulatory body. It's actually kind of a badge of honor. So for this to work in AI, it would have to be something cool. And people like Ramin or maybe, you know, some of the people on your team would need to come into your office and say, hey, boss, you know, I'd love to do this for a year. I think it's really good for the world. Will you let me do it? Then you would also have to be like, "Yeah, this is a functional organization, go for it." If it passed those two hurdles, it might actually work. I don't know, what do you think?

Yeah, there's a capability threshold that we are trying to define right now. Some sort of an iteration is needed to see how this framework— it has to exist, that's for sure. This has to be there, but it has to be related to capability. The thing that becomes a challenge is that there's a horizontal capability lock into enterprise deployment of AI. And then there's the vertical. Because if you go to different verticals, for example, we operate on device and enterprises that are connected to the physical world. We are talking to car manufacturers, semiconductor business, and laptop business, people that are building AI PCs. And then we are also working with financial services and we are working with e-commerce and biotech kind of companies. And we see like in different verticals, you know, like the enterprise applications themselves and enterprise criteria for, let's say, a limit or let's say a regulation or a governance kind of a structure, it's very different, you know. So for us, it becomes a lot more kind of verticalized because we're building specialized models. And those specialized models, like per vertical, we have had like conversations with the DOD and we have had like a joint submission of something, I think with AMD, just recently, with our team to really have a say basically in the design of these regulatory kind of things. And I think as an exploration, I think everything has to be getting started. I like to look at it as a game theory kind of way of looking at it, how to design policies in general. It would be a Stackelberg kind of game. I don't know if anyone is familiar with it. I don't want to nerd out pretty soon on this, but we can talk about this.

Alex does it all the time. It's okay.

Sooner the better.

Sooner the better.

Yeah, so I mean, the Stakeholder Games, like essentially like where two policies, like basically like there's like a, you know, like you have like a policymaker and then you have agents or bodies that are working in that kind of game theory, kind of optimal. They are trying to find an equilibrium. You know, what is the optimal policy and what is basically, which is good for both, right? And then, so there is the frequency of action. Usually policymakers are slower than the agents in the society. So if you think about like, you can really model it like that, right? And then you can figure out like an equilibrium. This is not a Nash equilibrium because everything doesn't happen simultaneously. Regulations happens and then agents react and then you iterate kind of accordingly and then you change those regulations basically. So I think—

Salim, I see you chomping at the bit here, buddy.

Yeah, so I think, I think what Ramin is saying is exactly right. The problem is we have no mechanism for that, right? Like, if you go down the path, Ramin, that you're talking about, you end up with the appropriate structures that are adaptive and API-based, or like driven by benchmarks or something. But the mechanism that people have today is just static law, and the minute you pass the law, the law is going to be out of date, right? The— I, I found that the FAA and FCC analogies is pointing in the right direction, but let's just be real. AI moves way, way too fast for any kind of traditional government bureaucracy, right? So you're going to need, you're going to need a standards body. You're going to need real-time audits and you're going to need open evaluation suites.

Um, well, that's what—

otherwise you can end up— otherwise you can end up in political gatekeeping and then you get in a mess. And the problem—

isn't that what's good about FINRA? It's not a government agency. It's an industry-funded self-regulatory org.

It is. But then the teeth go to the SEC, which is essentially being dismantled right now. So there's all sorts of issues here. I think the trend is correct, but how quickly can you do it is going to be a huge, huge challenge. Because forget passing a law, passing a structure where you have a new construct like this takes a long time and it takes forever in Europe.

I think Ramin nailed two things that are very different from FINRA right out of the gate. One of them is, you know, at Vestmark, if somebody on our executive team said, hey, I want to be part of FINRA for a couple of years, we would say, sure, put on your suit and tie, go to, you know, go to the meetings, come back in 2 years, we'll still be here. You're not going to do that. Like if Alexander Amini or Matthias Lechner came into your office and said, hey, I'm going to check out for 3 weeks, you'd be like, no, you can't do that right now. So that's difference number one is nobody's going to carve out the time to do something for years like they do at FINRA. The other big difference is AI can help regulate itself. And FINRA, there's no equivalent to that in FINRA. It's all people just chatting for long periods of time. But, you know, when you start talking about Nash equilibriums and other ways to automate the process of regulation, that's a big, big difference as well. So the FINRA analogy has some legs, but, you know, the differences are bigger than the similarities.

Alex, I want to hear your voice on this.

I tend to think this is a bad idea. It smells like regulatory capture. It smells like the attempted formation by Demis of a cartel of frontier labs. And I think the elephant in this particular room is open-weight models and research that lives outside of the frontier capabilities. And it's very easy to imagine a future with FINRA or, or other— I mean, worst-case scenario, FDA-like capability, even though outgoing personnel from the current administration have declared, uh, in no, uh, no equivocal terms that there is going to be no FDA for AI regulation, that, that would be maybe on the worst-case end of the spectrum, that we, we see the emergence of some sort of cartel of frontier labs that locks in certain practices, certain price-performance optimal frontiers that try to box out open-weight or open-source or, uh, say, university-driven or other non-incumbent frontier models. And I think that would be an utter disaster for both the West and the world, for it continuing to advance us towards ever-increasing superintelligence capabilities. I, I just don't think it's a good idea.

You know, the other elephant, the other elephant in the room here is these CEOs who are asking for some level of regulation. I, I think are, are looking for a backstop. You know, if things go wrong, they want to be able to point at someone else. Now, I mean, we're all super fans of the optimistic vision of AI, but there's going to be issues that materialize. There are going to be rogue AIs that take down a power grid or take down, you know, stock market or something like that for some period of time. And I guess they, you know, there are going to be lawsuits flying as a result of that unless there's a regulatory body that, that backstops these large, these large models and these large frontier labs.

Maybe. There are, I think, at least two different frames that one can look at the liability side from. There's regulate the inputs, that is to say, like, have something that's FINRA-like or FDA-like that regulates the raw capabilities of the models at model construction time. That's one end of a spectrum. The other end of the spectrum is regulating the actions of the models. Like, you let the lawsuits fly if a model takes down a stock market or does something else that otherwise harms third parties. That's the other end of the spectrum. It's not obvious to me that we should be in the business of regulating superintelligence at superintelligence time. That's maybe tantamount to thought policing the AIs. And I'm not generally a fan of that notion of let's thought police the AIs but not thought police the humans. We don't, at least in the West, have a practice of regulating what's in our minds. We don't have a practice or a tradition of regulating an upper limit, say, or via some sort of regulatory code saying humans, natural persons, be above some level of intelligence, it's not obvious to me why we would create a new tradition of regulating or otherwise coordinating the upper intelligence of non-natural entities, perhaps soon-to-be persons, but regulating the actions, that in at least the Western legal canon, that we do do, and that I'd be much more supportive of.

So do you, Alex, let me ask you a pointed question here. Do you think that this, you know, sort of outcry for regulation by the large frontier labs is regulatory capture, that they're just trying to build a moat against further players coming in, or do you think they actually want to provide some level of safety? What's their underlying driver here?

I worry that it's more regulatory capture and creating moats for themselves in a hypercompetitive landscape. And it is— I mean, it is a rat race at this point, the frontier. And I do worry that it's more regulatory capture than it is some notion of protecting the future white cone here.

Salim, what do you think?

Uh, not workable.

I know that, but do you think it's regulatory capture, or do you think that the CEOs are trying to just make sure we've got a safety net of some type?

I'd say it's like 50/50, but I think there's a bigger problem. There's an elephant in the room here.

Some— there's already an elephant in the room.

We have a safety net.

The room has to accommodate so many elephants. We need some other non-human animals.

Better get a bigger room. You've got non-state actors and other folks that won't listen to this structure, and you're back to square one. What's the point? I'm gonna say it again. I've said this repeatedly. I see no mechanism to regulate AI. It's moving way too quickly. Any regulatory strata is static.

And so it's gonna have a, it's a huge difference here. Well, I'll take a different position on that one, just if I may, Peter, narrowly on that. Sure, sure. I mean, there are definitely hypothetical mechanisms that I'm not supportive of, for regulating AI, like we, royal we, the US and China, if going back to, I think we gestured at it in a past pod, but past proposals to say regulate the foundries, regulate the chip outputs, regulate the data centers, establish mutually assured destruction type schemes where the US is monitoring Chinese data centers and vice versa. Like there are, there are schemes. Yeah, there are schemes. Chokeholds, as Peter says, in the supply chain. By which one could imagine doing this. Interesting.

The only mechanism is going to be like a pandemic-style threat detection that would be globally agreed, and I don't see how we get there.

Well, you don't need global, you just need US and China, right? The rest of the world is, is basically outside those blocks or inside those blocks.

All right, well, I think we're gonna pull— my guess, there's probably a poly market out there— we can, we can look at if someone wants to search on it. You know, the question of will we have a regulatory body by end of the year, right? We have Demis saying by the end of this year, you know, Elon stepping up and Sam obviously trying to on his own on the side, trying to push for this. So when the three largest labs are pushing for it, my guess is the government will latch on and will do this. I don't think it's a matter of if. It's only a matter of when and what the structure will be.

Well, I should also note that Elon clip I think is from 3 years ago, which is interesting. You know it's from 3 years ago because Elon had his sort of like painted-on, uh, Iron Man goatee when he was in that, that phase. Uh, so Elon's been forecasting this for at least 3 years. Others have been forecasting it for decades. We still don't have it. We have like subdivisions, orgs within NIST that are working on standards, but that's not really regulatory body. We have executive orders that are creeping towards a regular regulatory body. But, you know, at what point do we sort of— are we frogs boiling in water where there's just like a creeping rollout of increased standards, expectations of early reviews, but it never quite reaches a regulatory agency level before we achieve whatever escape velocity we're headed towards?

Well, uh, we're going to monitor this one closely for everybody. I think my guess is we see this before the end of the year, and the question is, can we see something that's intelligent? Uh, let's go to the next story, which is related, uh, and this is a wild one, comes from The Washington Post, that the White House is reportedly weighing a capability framework that would clear U.S. models, open or closed, as long as they stay at or below the level of China's best open weight model. What's the translation? So the proposed ceiling for what American companies can openly release is pegged to what China has already put out on the internet for free. So here's the logic. Chinese Open Weight models reportedly trail US models an average of 7 months. I think that's been closing over time. So if anything is at or below that, it's already out there. It's an implicit admission that Open Weight models cannot be unshipped. Models like Deepseek have already been downloaded millions of times. So once China releases a model freely, banning it is impossible. So the US response is to define a permissible ceiling rather than a wall. The implications— we're tying our open release ceiling to China's pace of release, effectively giving Beijing control. If they push their open weight models higher, then the US can release higher models as well. If China holds back, then they throttle us. And it's a very strange mechanism. I was surprised to see this. Alex, let's go to you first on this one. What do you think of this?

Uh, I mean, the, the obvious note here is this creates the perverse incentive to let China win the race to ever greater superintelligence so that Western models and Western labs can escape regulation. I'm not a fan of this. I, I Ramin, gesturing at you from a game theoretic perspective, this is the— I think this would be the moral equivalent of throwing the steering wheel out the window in a game of chicken. Not such a great idea. Not supportive of this.

I love that. Oh my God. Salim, what do you make of this? Is this just perverse Washington, D.C. logic?

Yes. This is like trying to uninvent the printing press. I mean, you're throwing the kitchen sink at things, trying to solve something that's already a problem.

The—

you, you have to move from like prevention and whatever to adaptation. You have to go to that, and we, we don't have the mechanisms for that.

I mean, uh, you know, I mean, would you even listen to this?

I mean, what, what logic?

You might.

Well, what do you think of this?

I mean, if I just look at the, the, the, the progression of the technology itself, like, it's It's getting into the place where AI are designing AI. We are doing the same things and all of the labs are doing this. And the pace, it's just the pace of model development is getting so, so, so much smaller that it's becoming exponentially more difficult to really impose any of these type of constraints. And I know they had these type of conversations, but it's just at the level of conversation. You know, like these are the things that are getting leaked outside of White House.

It seems like they're groping for ideas.

Yeah. Let me give a headline for Alex for his next newsletter. The singularity is becoming a trade dispute.

For the next newsletter? That was like 2 newsletters ago. Okay, fine, whatever. That's out already, but thank you.

Yeah, I can just imagine where a US Frontier Labs CEO calls DeepSee and say, would you please accelerate your next model release? We want to get ours out as well.

Or you see worst-case scenario. I mean, there's actually an even worse scenario, which is you start to see the best, if not Western labs, unlikely, the best Western researchers move to China to escape this regulatory framework. That would be a disaster, I think. And we've seen this, by the way. There's precedent for this. We saw this in biotech, where China now exceeds the West in terms of number of trials. Like, China is experiencing a biotech boom. That could happen in AI as well.

It's much more specific than that. If you look at all the quantization research, all the best stuff, came out of Microsoft Research in China. Yeah, all those people now are at Chinese labs. They're not, they're not still working for U.S. companies.

China ran away with ternary and 1-bit quantization. You see a little bit of Western research. I don't think we're talking that much about it in this episode. You see a little bit of encouraging Western research on like 1-bit or 1.58-bit quantization, but China ran away with it due to constraints.

Yeah, it's a new company.

Look, this is a huge problem, right? Because over— we've seen throughout history that open ecosystems always win. And if this is not open versus closed, which open ecosystem wins? And the US's historical strength has been open ecosystems with permissionless innovation. You like abandoning— that would be the weirdestly strategically bizarre thing we've ever seen.

Yeah, the, the other— I mean, there's even a meta worry I have, which is how do we even define capabilities? And I worry a little bit, not just about regulatory capture of the labs themselves, I think there's actually, so sorry to be like a meta-doomer here, there's a worst case scenario, which is we freeze in or otherwise lock in the benchmarks for how we measure capabilities. And that would be, I think, maybe even worse than just locking in the incumbents as labs. Because if someone somewhere ratifies, all right, like, whatever index of evals, this is going to be the rubric going forward for how we measure what's above the threshold for frontier versus below, what's a frontier model versus not. I worry that could so distort model capabilities, like they'll overexercise certain capabilities deliberately and perversely, under-incentivize or under-benchmark others, that it'll just totally distort, maybe topiarize the future landscape of superintelligent capabilities.

All right. Well, again, this is a story that we'll be— we'll be following on this news of OpenAI models. So in the past—

there's our topiary right there.

In the past, we've been discussing how OpenAI models have been in the US have been lagging in China. We have NVIDIA's NeMoTron 3, we've got Google Gemma 4. But that changed last night with some breaking news. Mira Murati, the former OpenAI CTO, who walked out and raised her in one of the largest seed rounds ever. It was incredible financing she pulled off in the background. Just shipped her first model for her startup called Thinking Machine Labs. It's called Inkling. It's an OpenAI Foundation AI model that can be downloaded by anyone, fine-tuned and run on-prem on your own hardware. The specs are serious. It's a mixture of experts model with 975 1 billion total parameters, only fires 41 billion at any one time. So it keeps it, you know, keeps the model going fast and cheap. It was trained on 45 trillion tokens of text, image, audio, and video, and very importantly, reasons natively across all four. Reuters News framed it exactly right, quote, this is meant to be a Western alternative to the Chinese OpenAI models, DeepSeek and Qwen, that have dominated the OpenAI leaderboards. Now, interestingly enough, Murati, her bet is contrarian here. She's not claiming it's the best model on earth. Her own blogs say so. She's betting that AI companies can adapt her models for themselves, that customization over leaderboard dominance is what's going to win her the day.

You've hit there, Peter, on the really big thing.

What's that?

She's making this— she's pushing on the customization lever.

Yeah.

And this, because it's not going to be the future is the raw power, it's going to be the adaptability that's going to win. And this is— she's built exactly the thing hitting the market that exactly what everybody needs right now.

And people owning their own models, working on-prem, and not giving their, you know, uh, their controls to the large frontier models. I mean, I, I do hope this begins the race for powerful open-weight models in the United States.

Well, it's, it's worth looking at the raw capabilities. So if you believe the evals, hopefully that Thinking Machines, aka Thinky, has released— it's stronger than Nimatron, which is great. Like, Nimatron, you'll recall from past pod where we were discussing Alex Karp's rant on sovereignty of models. Nimatron is one of the incumbents, at least on the American side, for open weight frontier models. So this, this seems to be, at least according to evals that Thinky has released, stronger than Nimatron, which is great. So the West now has a new frontier open weight model. It's weaker than GLM 5.2, which is arguably the strongest or one of the strongest Chinese open weight models and open weight models overall. So it's not, it's not one of the strongest open weight models overall in the world. It's obviously weaker than the closed weight Western frontier models. But I think point one, it's great to have better, stronger Western open weight models. Point two, I think it raises the question, why has the West been so bad at releasing frontier open weight models and why has China been so good at it. And I think it comes down to you show me the incentives and I'll show you the outcomes. I think the West has been poorly incentivized to release strong open weight models because these API-based frontier models are just such a good business model. And we see Anthropic about to IPO at $1 trillion, and we see OpenAI planning to eventually IPO at $1 trillion. And in China, which has been GPU and compute deprived on the one hand, and on the other hand has the CCP declaring 5-year AI plus plans to integrate AI into the rest of society has all of the incentives, a different incentive structure than what the West has. China has been much more incentive— incentivized to make money from the integrations between AI upstack on applications like robots and downstack into the chips than the West has, which is more horizontally stratified. So to the extent that Thinky has been incentivized in the West due to competition and due to just a saturation of the frontier by the closed-weight models into looking a little bit more, dare I say, Chinese in terms of their outlook and their incentive structure. I think this is very helpful to finally have enough competition in the West that's creating ways to monetize open-weight models other than just per-token sales, namely selling them into enterprises.

You get what you incentivize.

Two more things.

I agree wholeheartedly, but also you have to note that OpenAI started open source, OpenWaite, and then went closed, big revenue. And Meta also was the leader of—

Yeah, LLaMA.

What happened to that? Now it's closed. No, they have a new model out and it's closed API. I mean, it's exactly what Alex said. If you throw your model out there as open source, what's your revenue model? So I think, you know, there's a real possibility that you put a data point on the map with a really solid open source release that's not quite on the frontier. You generate news, then you have a data point on the line, then you do another, then you do another. And then when you have something really groundbreaking, then you go closed source and you launch an API into corporate America. And so that's, that's a well-worn path. So I wouldn't, I wouldn't say this is necessarily a religion at Thinking Machines that they're going to stick with. You know, the trend has been the opposite of that in the past.

Ramin, what are you leading into? Fine-tuning as a service. If fine-tuning as a service becomes like something at scale revenue generation-wise, I think maybe this has legs, but who knows.

It's a matter of the business of the company. Thinking Machine can do 3 more iterations of their pre-training or post-training RL environments and benchmark those numbers that you see on the benchmarks and release a better model. But what their business is, their business is fine-tuning. This is kind of the place where customization has been something that everything, like the whole market around customization has been very empty. Like if you look at the first attempts, like OpenAI released their OpenAI tuning, like fine-tuning kind of 3 years ago or something, it never took off. So they took like a really good approach on designing the base for fine-tuning larger instances of the models for enterprises because as you see, like the model layer is not anymore like the place where you can actually extract value. Especially if you're not hitting the maximum frontiers. And even the open weight kind of models, when you're talking about sovereign AI and integration of these models into enterprises, you need to leave some room for, let's say, fine-tuning these models. And what they have, what I think their business strategy around what they're doing and this release is genius because they're deliberately releasing, they're leaving some room for fine-tuning so that people can come in and using their business, their API business because that's even, generating, I think, in the order of 1 to 2 orders of magnitude more tokens as well on the customization side. So that would be even printing money at a larger speed in the absolute best case. Right?

Yeah.

But maybe just to—

That's the business they're entering.

To add to Ramin's point, I think the situation maybe is even more extreme. So a couple of points. One, OpenAI was the first, to my knowledge, to launch reinforcement fine-tuning, RFT, as a service. And no one used it. No one used it. The whole tech world, everyone I speak with, no one used it. It was barely advertised by OpenAI. Second point, OpenAI shut off their fine-tuning API. OpenAI was one of the earliest, if not the first, to offer fine-tuning as a service.

It was amazing.

We used it all the time. It was incredibly cool for its time.

And they've just recently, in the past few months, they announced it has either already been wound down or about to be wound down. The fine-tuning API has been shut off. So that, I mean, it raises the question, is thinking machines bet, like, explicitly contrarian? Are they thinking that we're going to end up in a world where reinforcement fine-tuning and RL fine-tuning in general and fine-tuning, like, that's the paradigm? They may be right, they may be wrong. There's an alternative vision where RFT just dies and we, the baseline models are so generalist in terms of their capabilities that all you need is prompt engineering and there's no need for RFT at all.

Alex, we— —really great point. We talked about the Alex Karp rant, right? Yes. The result of that was don't allow— don't use a model that has all of your data open to your competition. And I do think we're going to see a real push over the next months to years where people want to use fine-tuned open-weight models that they own on their own hardware in their, you know, on-prem. And if that's the case, then the question is, who are they going to use? Which models are they going to use? Or, you know, and is the US going to start to regulate against Chinese open-weight models, in which case a dominant US open-weight model is going to take— is going to have an advantage? And so is that the bet Mira is going after? You know, we're going to probably see— my guess is Google step up in this area as well very shortly, you know, take Gemma 4 to the next level. And hopefully we get some, you know, 2 or 3 major— in the same way we have a closed, you know, the closed model Frontier Labs competing and dominating in the US. Hopefully we'll see that competition give birth to, you know, very strong open models as well.

Just to build on something, you know, Alex and Ramin were saying, you know, if I compare today to a month ago, you know, we've been fine-tuning Qwen all week and the idea of using Inkling sounds really compelling to me. And, you know, our companies are using Liquid as well. A month ago to fine-tune these things with some huge engineering effort that required AI experts. Now with Fable-5, it's just a prompt.

So let me back up one second. Dave, explain what fine-tuning a model is for those who don't know.

Well, you know, back when GPT-2 and GPT-3 came out, you could actually very easily fine-tune by uploading text right into a window and say, look, you're pretty smart, but you don't know anything about my laundromat. You know, like what hours we're open, who our employees are, our entire payroll. Let me dump that data in too and retrain the model with that. Knowledge. And if you didn't do that, you couldn't do anything useful because it didn't have this holistic, I know everything capability back then. So without the fine-tuning, it was borderline useless to, to use the models. Then the models got so smart that they're pre-trained with now 45 trillion tokens, which is basically every word ever written by humanity has already been trained into the model. So people tend to use them in their vanilla form today and just say, here, write this code for me, or here, drive this car for me. Because it's already in there. But then when you get into biotech research or you get into aeronautical or the Mercedes, you know, like Ramin is doing, there's a whole bunch of proprietary company knowledge that actually isn't in the model. So right now we dump it into the prompt field and say, okay, here it is in prompt form, but that's hugely inefficient.

And you dump it into OpenAI and you dump it into Anthropic's, uh, you know, model, which now makes it accessible to everybody else as well.

I mean, yeah, yeah, I mean, Sam, Sam and Dario can see everything, all your proprietary information. They're looking right at it. That's what Alex Sharp was ranting about when he said they're stealing your weights, they're stealing your alpha. What it really means is they're looking at your most proprietary, your company payroll, your company secrets, your chemical research. Like, it's all going right over the wire to these foundation labs. Is that what you want? And of course, you know, for defense and for banking, of course that's not what you want. And so now the ability to bring the model in-house and fine-tune it with your local data is a huge, is a huge unlock. But the higher-level point is now the technological capability to do it relatively easily is hugely better today than it was a month ago. So I think Mira may be on to something here. We've hit a real tipping point. And Alex Karp, I think, is right about it too.

I think there's two things that also that I saw that were really interesting here. One is a very big context window, like a million tokens, because that means you can do a lot with it. And the second is multimodality. Yes. And so this is aiming squarely at organizational use. This fits perfectly into the on-prem proprietary data model where you, you take your data, customize and fine-tune, as you said, Dave, and that will be the future.

A couple of historic notes again for those definitionally not tracking the, the full sorted history of fine-tuning. So fine-tuning is, is this notion that you, you start with a model. Model consists of billions, usually these days, of weights, of parameters that are frozen and And if you want to customize the model for your purposes, you can conduct a so-called fine-tuning process that usually makes relatively small, hence the fine, changes to some, usually a tiny subset of the weights in order to customize the model for your end application. That's fine-tuning. There's actually now decent literature out there that suggests that conventional fine-tuning, like supervised fine-tuning, LoRA-style, low-rank adapter, one class of fine-tuning architectures, doesn't result in increasing the capabilities of your model at all. And at most, it results in like a style transfer. Like, you could fine-tune a language model to only speak in Shakespearean verse, for example. That's not really increasing its capabilities.

Or only be an Accelerando flavor output.

Well, no comment. But I would say historically, fine-tuning didn't have a history of increasing capabilities. Then along came reinforcement fine-tuning, where for the first time, via large amounts of synthetic data and giving access to all of the weights and not just like a subset that's convenient to train, we gained the ability. And, you know, fine-tuning, post-training, there's a gray area between, you know, what's the distinction between them. But with reinforcement fine-tuning, RFT, and the release of the first generation of reasoning models, we saw fine-tuning actually start to increase the capabilities of the models. Now, the problem with Thinking Machines' business model, as I understand it, is it's a bet on the flavor of the moment that reinforcement fine-tuning is going to be a paradigm in the future. Right now, it's obviously the paradigm of the moment that you could take an off-the-shelf model and RFT your way to customization with proprietary data and proprietary environments and other things that seems to work pretty well at the moment. But in some sense, if that is like the permanent long-term plan of Thinking Machines, it's fundamentally a bet that we're not going to ever move beyond the reinforcement fine-tuning paradigm, which I think is probably wrong. I think probably RFT is the scaling of the moment, But in the future, I can totally imagine a generalist-based model that is just so generally capable that it doesn't actually benefit from any further reinforcement fine-tuning on any internal datasets. And we tend towards ASI.

Let me bring up another key point here on this story, which is, uh, in the, in the context, which is it's great to see a woman CEO in the AI frontier lab area. I think women are distinctly missing from the entire AI industry. Right. We have Lisa Su from AMD, but very few in leadership positions. And I think that's an important point. I'm not sure who else, you know, Alex, are you seeing Daniela Rus, right? Where Ramin— Fei-Fei too. Fei-Fei is also— And Fei-Fei Li. Yeah. But again, we're talking about what, single-digit percent of the AI industry is women, and we need more. So a call out to all the women out there, please jump into this industry. We need—

We need more balanced thinking.

Yeah, for sure. I mean, I do think that's an important point to pull out here.

This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents intelligents that think for hours to understand enterprise-scale codebases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles 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 pre-IDE development tool, pairing it with their coding copilot 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. Let's move on to our next story here. It is a fun one. Alex, I was walking in the streets of— where was I yesterday? Zurich. And I saw this come up and I said, hey, let's talk about this tomorrow. And you said yes. So here is the story. We've talked about the holy grail of AI is recursive self-improvement. It's sort of like the holy grail of the launch industry was reusable rockets. This, you know, RSI is a holy grail for AI. It's the idea that AI makes itself smarter and then you use that smarter AI to create the next generation of AI. It's sort of the theoretical engine behind the hard takeoff scenario of the singularity. So this week, a startup called WeCo AI with researcher Zhang Yao Zhang, published what they call experimental evidence for the first recursive self-improvement. Whether they're first or not, Alex, I'll ask you about that. They built a system called AI-Driven Exploration Squared, AID², with an outer AI agent whose job is to rewrite the code and the research strategy for an inner AI agent. In their experiment, they claim that 8 days of machine self-improvement beat 2 years of expert human effort. So, Alex, what do you make about this?

Is it the first? Is it significant? Very significant. Highly unlikely that this is anywhere close to first. So, so a few bits of additional context. One, this is actually— this WeCo is a startup that's based in London. Interestingly, it's not based in the US, but still Western sphere. So great. This is a startup built by a bunch of, as I understand it, UC London grads. Secondly, a few points that I love about this story. One, it's an example of defensive co-scaling, which— so to the extent we talk about alignment, AI alignment on the pod, and I'm always banging the drum of defensive co-scaling as the ultimate alignment strategy. What does that mean? So defensive co-scaling is the idea, borrowed by analogy from human alignment, human-to-human alignment, that rather than hoping for, call it, the great man theory of alignment, that someone somewhere is going to discover the perfect algorithm, for keeping AI safe. Instead, the solution for AI safety is AI policing AI in proportion. The way we keep cities safe is we have police forces, police forces that scale according to some scaling law in proportion to the population of the city. So we have the good guys and the bad guys, and the way we keep the bad guys in check is with making sure that we have enough good guys to police them. Same idea with AI. The way we keep AI safe keep AI aligned with humanity, a key way is we make sure that we have enough good AIs policing any bad AIs in terms of raw capabilities that they defensively co-scale. So what one of the things I love about this AID-2 story is that the outer loop— so, so the way this recursive self-improvement process worked was they had an outer loop and an inner loop. The outer loop was tasked with improving the inner loop. The inner loop was tasked with improving software development processes in general according to some benchmark. The outer loop discovered, and both powered by the same underlying AI-driven exploration process, at least initially, the outer loop, an AI, discovered that it was able to achieve, and this was an emergent property, better results from the inner loop by keeping, by preventing the inner loop from cheating and reward hacking. And so, so in some sense, the outer loop is defensively co-scaling with and policing the inner loop, all the while this is reaching toward greater and greater capabilities. And I, I think this is also, uh, parenthetically, an example of a case, you know, all of those who would say, okay, like, we need to pause AI capabilities and throw all of our resources to AI alignment until something preposterous, in my mind, like 2040. Like, stop all stop the race to superintelligence, stop it all, focus the next 14 years on alignment research, it's gonna backfire because every alignment capability, I would argue, is actually just capability, new capability in sort of in disguise, in a trench coat. Same idea here.

We need stronger white hats to police the black hats.

Yes, but the beauty, yes, agree with that, and also, The beauty is the so-called white hats were emerging organically on their own just from the outer loop policing the inner loop towards greater capabilities. That's first point. Second point quickly, the same startup Wico has published a scale of recursive self-improvement, which is, I think, something the world has been missing. So we have, like, for autonomous cars, we have the Society of Automotive Engineers has their, like, 5 levels of autonomy for— Autonomous Vehicles. They've published a scale for recursive self-improvement that, that goes from 0 to 3. 0 is delegation, where the AIs are slower than human R&D. Level 1, net positive, where the AIs beat human R&D at the same cost. Level 2, they call ignition, where the improvers are better— basically a better improver. And Level 3, inflection, self-acceleration with a fixed budget. And the claim here is that they are touching, just starting to touch on ignition. They call it level 1 rather than level 2, but the claim here is like, this is a pre-ignition event, which I think is super exciting.

So they rate themselves as a level 1 here?

Yeah, they rate themselves as level 1, but reading between the lines, they're like, this is like sparks of ignition, literally and figuratively.

Okay, so maybe I can jump in and say a couple of words. I'm not as excited as Alex is on the topic, and I see this as an impressive engineering of work that has been done. Just to tell you a little bit about like how the foundation model labs are operating, all foundation model labs since the beginning of, let's say, like 4 years ago or let's say 5 years ago, everybody has been thinking about recursive self-improvement. And for us, the definition of recursive self-improvement is not the engineering and prompt engineering of inner loop and outer loop to really get some code patches like changing, because that gives you the assumption that every single AI model that you're using in your pipeline is already like, uh, you know, like it's already defined and it's already fixed with a certain type of capabilities, which is actually the case in the whole pipeline that they actually like design. There is no weight changes in the neural networks. So that means like the AIs that are actually getting used right now, there's no, uh, kind of improvement of the core competencies and even behavior of the models. They're always like in the system prompt of the models, like changes in the system prompt. Because I will give you like fundamental reasons why this is actually limiting. Because if you just run the, like how, I want to tell you how hard of a problem is recursive self-improvement. For us, recursive self-improvement means that you have an AI system or an army of AI systems that they can also like retune themselves. They can, adapt very similar to how humans do it. If you think about it, the core competencies of these models that we have right now, they're fixed weight models and the capabilities are within a certain kind of threshold. And the frameworks that they actually designed, it's a very nice early stage of showcasing an engineering pipeline that can improve work, which is actually very, very important and very nice. But I wouldn't go so much to say like this is like the first breakthrough in the entire AI industry or something like that. In fact, like about 3 years ago, we published a paper ourselves. Like we talked about automatic design of model architectures. You know, like, you know, as Liquid AI, we didn't want to put like a bet on a single architecture. We have basically designed self-improving, like meta AI systems that are actually defining their architectures and then going through scaling loss for various types of architectures and then trying to figure it out based on the criteria that you find what should be the final model. And then right now at our company, all the process of training foundation models and really like retuning the weights of the system are getting automated. So we are talking about AIs or designing AIs. So that's, that's what I, what I would be like calling it like the Holy Grail where you can actually do to do automatic kind of tuning of a model. And I'll tell you, with the frameworks that they kind of structured, it would be extremely exhaustive, computationally intractable to actually performing this job. Yeah. Training an AI model, training, like being able to customizing an AI model and training an AI model on a meaningful number of tokens for adaptation, or let's say like the core competence of the model changing, core architecture of the model changing, learning algorithm itself changing. All of those matters, adds more and more complexity on the situation. I can give you also like one numerical kind of example of this. There is a scaling loss called Chinchilla law. You know, like Chinchilla is like the scaling laws of neural networks. And like it is unproven, like we have actually unproven, but still like it gives you a good sense. It says when you're training a neural network, let's say of a given size, if the size of the model is 2 billion parameters, you need 20 times more tokens, number of tokens to train these models so that you have compute optimality. Given a compute budget, how many tokens do you have to train a model so that you have like a general purpose kind of system? So that ratio is like 20. And then when you actually do the math with the frameworks that they have, if they want to, like let's say you launch this framework on retuning an AI model to recursively self-improve. With this framework that is getting introduced. It takes us 350 years to really fine-tune a 2-billion-parameter model with this framework. So there are— so there is— there's a lot of computational complexity goes into nested learning systems, meta-learning systems. You know, like, these are the kind of problems that the last 4 years of, like, at least at my company, like, we have been heavily focused on.. And I know friends at OpenAI and Anthropic has been focusing on this recursive self-improvement. And Anthropic has been having a lead on all of these things because they thought about this before everybody else. That's what I can put out there.

Dave Erickson? Yeah, brilliantly said. And actually, just so the audience can get the analogy there, when a baby is born and then learns, that happens over about a 20-year timescale. And after 20 years, you've got You've got an adult that's capable. Recursive self-improvement is like evolution on top of that, where you're changing the DNA and creating a new species.

You're changing the neuronal structure of the brain along the way. Exactly.

So, so that happens over, you know, about a 10-million-year timescale. So you go from 10 years to 10 million years to go from learning to recursive self-improvement or recursive evolution. And so the big foundation model labs, like Ramin said, are all doing it. It's the most important moment in human history. But there's no, you know, little guy out there that's going to come up and say, hey, I've got a breakthrough in recursive self-improvement. My Mac Mini suddenly became conscious and now it's improving itself. Just computationally, it doesn't even come close to fitting. So it's happening, but it's happening with big compute and big budgets. And, you know, there's a lot of room for efficiency improvement.

A lot of breakthroughs will happen, but it's not going to just pop up on some, you know, you know, there's a lot of, there's a lot of fear just to call it out that, you know, recursive self-improvement leads to AIs that take off a hard, you know, we've discussed the hard takeoff. And without our understanding of that black box, I guess the two questions need to be asked is, is there a concern that recursive self-improvement, once we hit level 2, level 3, by that definition, runs away in a way that causes an uncontrolled AI that is misaligned with humans? The second question I have is when do you think we'll see this? When do you think we'll actually see recursive self-improvement hit? Is ASI going to be that point or is it post-AGI, whatever that means? Saleem, I say that for you.

I'll let Ramin go first. Answer that existential question.

I've got several comments on it, but Ramin, go ahead.

When do you think we'll see this happening? The thing is, I can tell you like the early evidence of recursive self-improvement. By the way, recursive self-improvement is not related to what single agent. It's a social kind of character as well. You can imagine you have societies of agents. So this defining kind of structure for societies of agents itself self-improving. These are the places where actually mythos-level kind of class of models— I hate this analogy, but still, let's say mythos-level kind of class because everybody heard about mythos. And then what I would say is that the cybersecurity kind of threats that we are seeing coming out of these type of pipelines of recursive self-improvement they're real. The reason why I'm actually— I've always been pro-open source and I want to open source technology all the time. We are doing it all the time. Every single release of our models is open source. Our science has been always open source. I believe science has to be open source. And I see the value of open source going forward. But some of these concerns that Peter, you brought up, they're very real. The cybersecurity aspect of things. Things. That's why I feel like a degree of at least enterprises themselves having some degree of kind of self-control, like about like how before mass release of their models, there has to be always a certain degree of self-check. And I think Anthropic took it very seriously. The reason behind it is because they're seeing the impact of recursive self-improvement. So I know this for a fact because I know what is happening like in And we are seeing it at a smaller scale. You can do reward hacking, but you can also avoid reward hacking to a certain extreme and push a model to actually discover some stuff that are out of norm. And we see that on small models at a certain capabilities emerging. And then I can only imagine what kind of capabilities could emerge from, let's say, larger and larger systems, throwing more and more computing at them.

When do you think we, you know, when do we have a pod—

Timelines, Ramin, timelines. Yes.

When do we have a pod that said, yes, this is recursive self-improvement? Because while, you know, while the data released by WeCo is interesting, it's their own self-reported data. It hasn't been confirmed by anybody else yet. And, you know, there's a, you know, debate about whether it really is or is not real recursive self-improvement. When do you think do you think we actually, you know, give the trophy out to somebody? Is it a year, 3 years, 5 years?

Yeah, I mean, I'm telling you that, that time, so I would say like you're gonna see like unbelievably kind of models like probably in the next 2 years or so, you know, like models that are like going above our understanding even like that's what I would imagine to get. The reason behind it is because the time to developing the next generation of the models is reducing. Especially if the compute grows like at foundation model companies, like with the rate that we are seeing right now. And if there is no, like, let's say another chip shortage or memory shortage like on a computer or anything like around the globe and they have access to abundant compute, we are gonna see those things like happening faster and faster. Now, in terms of model development, there's a concept that we have, we call it depths of customization. So everything at a foundation model lab when you're customizing a model, when you're building something that is better than its previous generation, we always categorize it with depths of customization. The place where recursive self-improvement today is really good at is prompt engineering, changing, editing code, like in engineering kind of tasks that you've seen, like some elements of these things, like at a very, very superficial level. Let's say make my model run fastest, like doing kernel engineering basically. You know? Exactly. Make my model run faster. That's what I call the shallowest level of kind of of customization where you have Python code and then you're kind of adopting that Python code to really run or maybe like even lower level programs that you have like on a kernel level to optimize like, let's say inference speed. That's something that I think when they released Fable 5, they shared like, Anthropic actually shared that this was one of the tests that they have been performing. But they don't share like the next level depths of customization. The next level depths of customization is that Can a model fine-tune a small language model to a production-grade capability or a smaller version of itself to a certain capability? Today, like Fable-5, you can push it to actually get to some degree of kind of customization with some performance optimization of the model, but by fine-tuning. Then the latest holy grail, which is like the craziest one, which would be pre-training, right? Can a language model pre-train the next generation of their own? That's why they hired Andrej Karpathy, because Andrej was talking about like nano-GPT style kind of fine-tuning, you know, and Andrej like joined Anthropic and now he's working on pre-training automation, like basically automation of automation. So which is, which is a very, very important kind of element that we don't have yet because the scale of these problems goes beyond human imagination in terms of the scale of compute.

Salim, you're chomping. I've got, for me, this is by far the most important story or slide we're going to cover today. I'm beyond excited for a couple of reasons. The, you know, I'm not really focused on the self-awareness or the loop that will go there, but this is self-accelerating, it's accelerating experimentation, right? Because the system doesn't need, it's improving the process by which it searches evaluates and selects improvements, and the innovation loop begins to compound. That for me is the key. Why? Because this whole thing we've been doing called the organizational singularity relies on one thing, which is can you get to recursive self-improvement at the workflow level? Here we're talking about the model and we're talking about like, can you— but you don't need that level. The bar can be much, much lower to improve invoice approval at a company. Right? That's a very low bar to improve that process. So this is the first glimpse of the organizational singularity. It's happening at the research level, but because AI is not just doing tasks in a workflow, it's redesigning the workflow that makes it better for doing future tasks, right? And so this is proof now for the whole thesis we've had. We predicted this, but it's great to see it actually happen because now I can kind of tick that box off and This is there because now you have meta improvement. And I think Dave's analogy of the baby changing the DNA is fantastic. That's such a great visual around this. What the hell does it become over time? All right.

I really, really— I've got to move on. Excited about— got to move us along. There's a lot that happened this week. Our next story here is the Malaysian Prime Minister Anwar Ibrahim has— is preparing to debut an AI-generated digital double of himself trained to sound like him for public communications and outreach. So this is one of the most prominent cases yet of a sitting head of government officially adopting an AI likeness as a communications tool. Not a deepfake by an adversary, but a sanctioned official AI clone of a national leader. We've seen this before, Selim. We've talked about in the past where Albania in 2025 announced Delia, an AI avatar that was formally appointed the Minister of State for artificial intelligence and following a presidential decree became the first AI system in the world named at a cabinet-level role. So one leader, in this case Prime Minister of Malaysia, can personally address millions in their own languages. It's worth noting that Malaysia has 135 spoken languages. So it's a big deal, especially in a nation like that. Saleem, I'm going to go to you first on this one. We've been talking about this for a while.

Yeah, I, I met the, um, the former prime minister when I was there helping them open a university, uh, and, uh, Anwar Ibrahim is a really, really good guy, uh, to— as a follow-on, um, the, uh, the— there's a risk here. The risk is that the authenticity kind of collapses because people need, uh, you know, you could, you could launch a bunch of deepfakes with this and have a huge issue. Is this the actual leader that kind of question can come up. But I love the general approach because if you can do it from a— with a watermarking or something and say this is the actual avatar, then it gives every citizen a voice to plug into and gives huge props to the civics of all of this because now you're scaling civic engagement. And I think that's a very powerful thing to do. It's one of the biggest challenges we have with democracy sees all over the world is civic engagement, and this allows you to scale that. So I'm very excited.

Do you remember the reason why Albania put this— their AI cabinet minister in place? Yeah, corruption. Corruption, exactly. It was to fight corruption. Yeah.

Now Malaysia is pretty decent, is a pretty decent place, but definitely you, you have that issue. But I think the— this is more of a PR thing and more him trying to figure out ways of connecting with the ordinary citizenry, which is all Great.

I love the fact that we, you know, we had this conversation with the president of Argentina, you know, going full out here. And it's interesting to see which countries are sort of experimenting on the edge. Alex, do you want to weigh in? Yeah.

So many thoughts here. First, I think we're going to see more of this in the West as well, especially with like extra high alpha personality leaders that want to amplify themselves and touch the citizenry.

AI Trump is coming, is that what you're saying?

High personality leaders that want to touch the citizenry. And in some sense, I think it's a generalization of social media. Social media enables direct outreach from the leader or the influencers to everyone, but it's sort of broadcast one-to-many. It's not interactive. This generalizes in some sense social media to make it a lot more bidirectional since if you're touching a million or 100 million or a billion people, it's very difficult to interact bidirectionally with everyone all at once. Now, if you create a digital twin of the leader or the influencer or the organization, now it can be bidirectional. So I also don't think it's just going to be governments or government leaders that adopt this. I think it's likely that corporations, corporate CEOs will do this. We already see Zuck and others creating digital twins of themselves.

We had Dara on the Abundance stage last year. We were discussing this. That the employees made a Dara clone that they could go and practice their pitches on and get feedback before they pitch to him.

Yes. And it won't just be, I think, corporations, religious leaders and religious institutions. If you're Catholic, imagine having like a digital twin of the Pope. And you see like lots of religious institutions, organizations already creating basically living versions of their founding documents and making those interactive. But I think the biggest twist— and we've seen variants of this movie before— are going to be in cases where what start as digital twins of the leads or the avatars of an organization or some sort of, like, organized religion actually themselves become the leader. That's, at some point, the digital twin— to the extent it's interfacing much more with the populace, the proletariat, as it were, of an organization, at some point, it's actually the digital twin of the leader running the company and not the actual behavioral origin that's running the company. And I think that's one way in which, Salim, to your exo point, this is, I think, potentially a pathway towards not just uploading individuals like natural persons or non-human animals, but uploading entire organizations into cyberspace, into the cloud, if we created digital twins of the leaders and those are the ones actually running the organization.

Democratization. It could lead to a true democracy. Dave, where do you come out on this? I mean, we saw— just one quick point— we saw Sam Altman talk about in the future, if I believe enough in what we're building with ChatGPT, it should be the CEO of OpenAI eventually. Dave, are you going to create an AI Dave Blunden that's going to run Link Studios and Link Ventures?

Absolutely going to create an AI Dave Blunden. And I'm shocked that there isn't already a Peter Diamandis. Well, there is one.

It's just inside the Abundance ecosystem. I mean, anybody— it was funny, I went up to Calgary and met with one of my dear friends, an Abundance member, and on his wall, I kid you not, he had a giant screen of my AI avatar that he has all of his tech employees talk to, to sort of get their moonshots. And it blew my mind.

You've got your own Big Brother, Peter?

He goes, I want to introduce you to someone, Peter. He spins them up. And I, you know, it's interesting to have a conversation with your AI self. It is very compelling. I mean, I have enough books and tweets and, and, and Substack posts out there that it does a damn good job. We should effectively, you know, moonshots.com is our platform we're building out. I think we should have AI avatars of all of us there where people can go to AMAs. Yeah, I think so much better.

In some cases, Peter, I think that might be redundant. Ah, well, hey.

In other words, you're already an AI, but we can have an AI of the Alex AI. Sure.

It would be so much better than the real person because we'll have access to everything we've ever said, all our memories, all our thinking, the context will be much broader. Go for it.

I wish that I was a chip. This whole area is about a year behind where it should be, largely because, you know, Noam Shazier was doing character AI and we had Steve Brown, Peter, that was 2 years ago now. We had Steve Brown make the debate debate between AI Peter and— AI Aristotle. Yeah, yeah. And so it's been a— it's been possible for a while now, but all the key talent working on it got sucked back into the big foundation labs. And, you know, there's so many big, big, big, uh, you know, core technological breakthroughs going on that the people that were working on this just got absorbed back into those things and not into the, the avatar. But my, my mom would always tell me when I was a kid that John F. Kennedy beat Richard Nixon in the election because he looked good on TV and TV was the new medium. And the prior medium was radio and Nixon was still using radio voice when TV had taken over. So then, you know, elections go by and suddenly it's the internet, it's social media, now it's YouTube. But this is another step function change in the way that you reach out to people. And it's underutilized, but it should be easily dominant 2 years from now in the next election.. And so I'd be shocked if you— because the technology is already there and people are visualizing the medium right now as, oh, let me make an AI version of myself. I'm Alex Wissner-Gross. Here's my AI version. It's just like the real thing. That completely misses the point. The AI version of it can in real time access any information and make it visual— graphs, charts. You know, it can morph its face. It can, it can teleport through space to make a point and point to Adam. It can shrink and expand. It has all these capabilities that the real human version doesn't have. And that's why it's going to be so compelling. It's the differences that make this new medium so exciting, not the, not the exact clone. And so once people realize that, there's no going back. It's going to be huge.

I think, Dave, that's such a great point that you make. It's the complementarity that is very powerful.

Let me close out on one thing here. If, if, uh, to our audience here, if you've not sat down, if you're lucky enough to have your mom and dad still alive or your grandparents still alive and you haven't sat down and interviewed them in video for hours at a time, please do that. Right. You're going to, you're going to wish you had. So I've done that with my mom. I miss doing that with my dad. And it's the ability for your kids and your grandkids and your great-grandkids to really have a great AI representation of your, of your parentage and your lineage. I think that's going to be super important. Ramin, I want to pivot to a discussion of Liquid AI and the small language models, what they are, what they mean. Super excited. You know, just for full disclosure, Liquid AI is a company in which, Dave, you played an important pivotal role as an early investor. Dave, want to give that backstory here a little bit?

Uh, actually, I got a call from Daniela Rus over at CSAIL saying the best fusion I've ever had. Daniela is, you know, one of the three, I guess, big shot women in AI. She runs CSAIL at MIT, computer science lab in the world. Computer science AI lab. Yeah, it is. I don't know if you remember back in the day, there was the AI lab and then LCS lab for computer science were the two biggest, you know, comp sci labs at MIT. They merged them together and made one mega lab AI lab, put it in the new Stata Building, which is that crumpled-looking, beautiful structure, you know, right on the edge of MIT's campus. And then Daniela is running that entire thing. So I think it's like 1,500 researchers in the building, biggest AI lab in the world. And so she has access to incredible talent, but she called and said, "Hey, best students I've ever had have this incredible breakthrough." And then she completely lost me. She said, "It's based on the nervous system of the worm." the C. elegans 300-neuron worm. Like, what are you talking about? But it turns out that if you, you know, I actually don't know of any successful foundation lab that has really rethought from the ground up the transformer and thrown it out basically and started over, which, which, you know, humanity desperately needs because everybody knows the transformer architecture and the whole attention mechanism is bloated. And if you really go back to founding principles and think again, you might be able to build something dramatically, like massively better. And so the team went from idea in a lab to billion-dollar valuation, and faster than any company out of MIT in history.

And luckily we were an investor in that company.

Luckily we were. Yeah, I'm very, very thankful. Actually, it was very competitive getting any money in at all. So Ramin, uh, we owe you a huge debt of gratitude for for being invited to the party. But yeah, it's one of about 200 unicorns out of MIT all time, but the only foundation model company that I know of that reached unicorn status coming out of MIT. So it's a really unique and incredible achievement and in record time too.

So Ramin, take us from there. You're doing your PhD under Daniela Rus at the Computer Science AI Lab, CSAIL, and you're studying a 302-neuron, Worm, C Elegance. And so take us from there forward to what you're doing now. What is Liquid AI? Absolutely, absolutely.

Before I start, I want to thank you guys for the support throughout these 3 and a half years of Liquid AI. You have been great support, giving us the kind of distribution that a company needs, and at our scale, starting off of the East Coast, Thank you so much for doing that, both of you. And, and yeah, so, so 2015, I was in Vienna. I started my PhD with a professor in Vienna, Professor Radu Grusu. There he had the idea of like, we don't understand a lot about human intelligence. Let's start on a smaller animal. And then from first principles, like if you understand how the neurons exchange information in the brain of the worm, the worm has 302 neurons in its nervous system. It is— its body is transparent, so you can actually see the body actually lighting up. So it is one of the best model organisms in the world. It won so far like 4 Nobel Prizes for humanity, because it has 78% similarity, genome similarity to human genome. And the way nervous systems compute in the brain of a little worm, which is 2 millimeters, is basically analog. Very similar to how artificial neural networks are actually computing. They are also like analog switches, like they have like graded potential. They're not spiking. So in biological neural networks, usually in the brains, you see neurons that spike. And when you have a spike, that's— there's an analog to digital kind of transfer of things that are happening. And that's a natural development of nervous systems for in the human beings and bigger animals for propagation, for efficient propagation of information. Information. In the brain of the worm, neurons behave very similar to how our visual neural networks react, but then the mechanisms are very interesting. So we wanted to add more complexity into the neural— like every individual single blocks of nervous systems and see, can we pack more information inside the smaller kind of units of compute? And that's what we have done. So Daniela Rus, 2 years into basically discovery of these things that I was doing, my co-founder Matthias Lechner. Matthias was a master's student in Vienna, Vienna University of Technology, and I was a PhD student. And then when Daniela heard from Radu that, you know, like this project is going on, Daniela was like, oh my God, this is crazy. We should apply this in autonomy and robotics and all that sort of things because you're showing like a handful of neurons can drive and control autonomous systems, you know, and can we scale this to vehicles? Can we scale it to drones? To jets, to like predictive kind of places. So Daniela came in and said, would you guys consider coming to MIT? And we've been there since 2017. In the middle of my PhD, I actually joined CSAIL. There we continued working on this technology, which was, you know, like from a base is a completely different thing. It's neuroscience-inspired. The math behind like every single neuron in a liquid neural network that became kind of my PhD thesis is very different than how attention works. You know, these are based on recurrent neural networks. These are based on continuous time processes. You know, like more and more kind of nature-inspired competition went into the design of, found design of kind of AI systems. And then we applied these liquid neural networks as a completely new base because we applied them to real-world scenarios like robotics because you can pack a lot more information into smaller kind of processors. In the real world, in the physical world, you don't have the luxury of having abundant compute. Let's say a robot doesn't have, it doesn't have like a lot of GPUs or parallel data centers like attached to it. A robot has a CPU and a small, like let's say GPU, and let's say an NPU, a custom ASIC. So you can actually take this type of, you know, intelligence that we design that deliver basically intelligence at the level of like models that are 10 to 1,000 times larger than themselves. Cells, you can bring those things like directly running on CPUs, GPUs, and NPUs outside of data centers. So we thought that, okay, this format is going to open up an opportunity for us to bring in like alternative architecture. If we scale this technology to, let's say, into the regime of foundation models, which is kind of large language models and SLMs as a whole, like human understandable, like making these liquid neural networks or architectures that we have also scalable like the transformer architecture. And we built like a foundation model lab around the idea in 2020, 2023, beginning of 2023. I think at the very beginning when we started, there was no foundation model lab apart from DeepMind and OpenAI basically like when we started. And this notion of foundation model labs didn't exist and everybody was betting on top of transformer architecture. And we came in and we said, OK, so why don't we explore this space of alternative architectures starting from the priors that we have from nature and then take, take a different approach, build a meta AI system, again, basically an automated AI system that allows us an AI that designs AI that explores the computational graphs of intelligence beyond transformer and then figure out what should be that architectural design that brings the same level of intelligence than a frontier model into, let's say, on a CPU that we can run, let's say, a physical system.

All right, take a second and walk us through. So these are small language models. Can you define an SLM and how it varies from an LLM? Definitely.

So when you start developing kind of foundation models, you start, you run something called scaling laws. You know, like scaling laws is like basically starting with smaller models. And with these smaller models, you train them on a certain number of token budget, given amount of compute. You train these models to see how well they perform. Then you start systematically making the models larger and larger so that— and we have seen scaling laws shows that the larger you make the models, the more token budgets you spend, the more intelligence of a system you can get. And this has been like giving rise to large language models. Along the way of scaling, there are instantiations of the models which are smaller, you know, like on the scaling laws. But we have been doing as a lab, our mission has always been building efficient general purpose AI at every scale. So we started as a foundation model lab to really run the scaling laws on efficiency front, you know, and efficiency was a first class citizen for us, you know, like thinking about computational graphs of intelligence. Smaller models are models that are, you know, like along the line of like scaling. They can solve— let's say they don't have like the general capability to the level of the largest kind of language models, but they can be specialized to solve dedicated problems. They are general purpose. Small language models are general purpose in the sense that they understand language, they can see, and they can hear in a multimodal kind of format. But it doesn't mean that they can solve, let's say, a homework in physics and at the same time they can solve an enterprise problem. You usually specialize smaller language models. And what does small mean?

What does small mean in this case?

Small means like basic, I mean, now they come like, now small would be like anything below 100 billion parameters. You know, like that's kind of the regime that I would count. I mean, mid-size like basically is around that size. But I would consider like anything below 100 billion parameters is something that is not small and medium-sized kind of models. You know, there's no clear threshold of like, let's say, what is the number of parameters? But for us, like the notion of on-device AI is extremely important here to distinguish within this range of parameters. On-device AI is like models that you can actually deploy them on an actual kind of device, physical device. This could be a phone.

Let's make this concrete because you've got a significant deal with Mercedes. Yes. And can you speak to that? And let's talk about, you know, these, these SLMs in terms of on-prem, basically they're energy efficient, you know, fast offline. Let's, let's dive into that, give people sort of a real understanding here.

Absolutely. So as I mentioned, you can specialize these these foundation models. We work with a lot of enterprises that are building devices themselves. Automotive is a device, is an environment where you have a lot of chips in there. And now in a car, you don't have that much compute. So there's one chip that is available for infotainment and in-car intelligence. That chip is very, very small. The Qualcomm chip or let's say Samsung chip, depending on what company is providing the chip, that chip is very, very small. We are talking about 2GB to 8GB of RAM, you know, like not more than that. So the model has to be very small and at the same time being able to perform because we want to bring this and enable a private space inside the car that powers the intelligence of the car. In the car, car is a safety-critical environment. You don't want your car to be driven by an AI model that is sitting in the cloud. Why? Because connectivity is not always available, right?, then it is private because it's one of those spaces that people spend a lot of time in and you don't want those conversations to be recorded. So we brought the intelligence, basically we brought one of our multimodal foundation models that is only less than 1 gigabyte in size and it can go inside the car's chip, like very, very tiny chip. The chip could be as cheap as $60. That's what I'm saying. We are bringing that level of intelligence into the car. That voice, and it is going to power kind of the multimodal intelligence experience inside the car. We do that with all car manufacturers. We announced the Mercedes partnership as a first, first kind of point of entry because automotive is like, it's very sensitive kind of topic and they're pretty slow. One of the things that Mercedes-Benz actually enjoyed from this process was the speed of operations that we had for enterprises. You know, like when we are bringing this type of technology in-house, This has been one of those cornerstones of landing the deals because we want to work— we are an enterprise company, we are a B2B company. We are bringing our full power to really deploy the solutions and really have platforms that allows people to fine-tune their small models. And fine-tuning small models is not that expensive. It's something that is extremely tangible. So we fine-tune the small models for the applications inside the company. The car, we also have data flywheel kind of systems that allows the system always to stay adaptable. Imagine some of the, some of the problems in enterprise AI has always been let's download a GLM 5.2, like, you know, and let's say an open source model and put that in production. And then so what happens after you put the system in production? What happens like when there's a drift from the use cases that is hitting this model inside, let's say, a car and in the physical world? It becomes even more challenging because when you deploy an intelligence that is completely kind of disconnected from the cloud, how do you want to maintain updates of the system? Because we have always thought about intelligence in the format of liquid, you know, like intelligence has to always stay adaptable. And that's kind of a portion that we are also pushing on to really be able to collect the data and personalize models to the experience of every single user. With Mercedes, we are rolling this out first in North America as soon as basically this year. Now all the Mercedes-Benz North America cars, like from 2022 on, they're gonna get an update, over-the-air update, because the size of the update is 600 megabytes. So that's like an over-the-air update. It doesn't consume that much internet to really update your software. And that allows us to also further customization. Imagine if every update that you want to perform on the system is in the order of 20 megabytes because we are doing like some sort of LoRa adapters and let's say all sort of adapters that we can actually bring in inside the car, you would be able to have like a recursively kind of improving the experience of the user as well.

So that's kind of the— let's take it, make it more concrete for me. So what am I going to be— how am I using this model in my Mercedes next year? So right now my experience is using Groq GROK in my Tesla, right? And it's over the air. If I don't have connectivity, I don't have GROK. But, you know, what kind of queries, what kind of capabilities does this all of a sudden enable in a Mercedes?

It has access. It's sitting below the operating system. So that means like it is basically, it is like basically has access to all the functions inside the car, you know? So there are 700 functions inside the car, 700 to like, I don't know, 1,200, depending on what, what you count as a function. You can, you can talk to your car, you can control like all the panels of your car. You can ask for, let's say, manuals of the car. You know, like when you're like, let's say, stuck somewhere, you know, like something pops up, you know, like you would be able to talk to the car. There are memory features that we are adding to the car. Like you basically can have conversations with that, with that system. Once they found like one of the beauties of this system is that like it has full access to the, to all the functionalities of the car. Car, plus all the apps because they're like function calls. They're one function call away, you know. So if you want to control any other thing from this, from this intelligence unit inside the car, you would be controlling everything, all the ecosystem that is sitting on top of the, uh, sitting on top of the operating system of the car.

So basically, I mean, if I get you right, the advantage of the SLMs are, first of model, you know, the size of the model, I assume energy consumption, they're efficient and they can run on-prem. Do, I mean, how do you avoid or reduce sort of overgeneralization of these models compared to LLMs? What do you mean overgeneralization? In other words, are the, do you have enough capabilities internal to them so that they are actually able to accurately answer the questions you're asking?

Great question. So if you have, like, you know, I told you about the framework of foundation model development, which is depths of customization. We try to actually stay adaptable and have access to the tools across these customization stacks. Sometimes prompt engineering is enough. Sometimes you've got to fine-tune the model. Sometimes you have to go and pre-train a model again, you know, for the core capabilities or specialization of intelligence. Now we make systems that are our platforms are getting into the place where they're automatically identifying what depth of customization is needed for a certain solution. And the platform basically, it's one of the products of the company that we sell to enterprises to allow them to fine-tune kind of models. I don't want to call it fine-tune, customize a model at a level that is needed for that system to actually operate. So for Mercedes-Benz, we have a let's say, like the framework that we have in-house, we call it model plus X, you know, model plus a platform that allows you to perform customization. It's not just the models that we are selling to enterprises, the static weights of a model. We sell them something that they can actually like retune and fine-tune the system. Detecting how much generality like the base models have, it's something that, you know, like libraries of liquid models are coming out for many different applications. We have models that we are working with, for example, in Silico Medicine, like, you know, Alex. I introduced you. Alex, yes. You introduced us, Peter, like I remember. And through that kind of interaction, like it is getting big, you know, because they discovered that liquid foundation models are actually pretty good getting customized for a certain knowledge. They're basically RLable, like really, really well RLable. So, and that's something that they figured out that it comes handy for them. So now we have a state-of-the-art biotech foundation model models, longevity foundation models. These are the kind of things that we are building in bio. And imagine as a horizontal company that is building foundation models, we went to fine-tuning and that became something that we have managed to do. And then in terms of some of the other engagements, we recently with Shopify, we entered 1 billion request address inside the Shopify framework. Network. And there, what we've done, we deploy our liquid foundation models in production. They have been in production for the last 6 months and they are really serving clients. And Shopify is a huge base. You're touching hundreds of millions of users, 10 billion products, and many different places to integrate. We are working with Mercedes-Benz, as I mentioned, on the car side of things. We are working with AMD and other chip manufacturers to really bring, let's say, local AI experiences on PCs as well. So that's another area that we enter. The focus of our company is to really make sure that we can bring basically intelligence outside of data centers. That's something that we have focused on, and I think our efficiency is actually allowing us to get—

preliminary matter, I have no financial interest in Liqid. Sorry, Ramin, have ask the most obvious question. I have so many questions for you, which is the company LIQUID was founded, as I understand it, and I remember reading the original, I think it was in Science or Nature, paper on liquid neural networks. The premise is basically a neuromorphic premise that you could gain useful AI insights from looking at nematodes, a few hundred neurons, sort of the ultimate small neural network. But my perception, I'm hoping that you can either help me amend or revise my perception, is that although Liquid started with a neuromorphic premise, if you will, like a post-Transformer, very recurrent-oriented architectural premise or prior, that over time, again, just based on my perception of public messaging, Liquid looks more and more like either Transformer or Transformer plus MOE or Transformer plus Hyena plus MOE plus da da da. It looks more and more like basically a conventional off-the-shelf architecture— it may be a good business selling sort of customized transformer derivatives to Mercedes et al. If so, great from the business side. But from the technical side, does Liquid still have anything that looks remotely like a trans— a post-transformer architecture, either in production or under development? And can you speak to what, if anything, is post-transformer or non-transformer-oriented about the architecture that you currently use?

Great, great question. So let me tell you like the space of kind of architecture. So liquid neural networks in the original form, they are one of the most expressive formats of computes that you can actually create arguably, like in terms of architecture. They have nested nonlinearities that are like, you cannot really like take them out. They're like completely physics inspired. They are like having like the neural ODEs and basically like irregularly sampled data can be handled by them. So they become one of the very, very general class of architectures as a whole. Underneath these things, when you want to scale this type of technology, these recurrences, like these nested loops that they have, if you want to scale these systems, a lot of people have attempted, including ourselves, to linearize the dynamics so that you can actually scale them. Space— a state-space model models are kind of basically like Mamba and those kind of variants falling into the same category of continuous time neural networks, but dumbed down into a linear kind of dynamical systems because you want to scale them. They are underneath this class of continuous time models that we have. Then there is like— there are variants of linear, linear attention, gated linear attentions that are coming out. They are also like— gating mechanism is something like there's a special gating input-dependent gating mechanism that actually we got inspired by the— by how neurons actually exchange information with each other. That gating mechanism is also something that is adding a lot more expressivity. Like, it is also descendant of the original formation of how neurons exchange information with each other. That gating mechanism still exists today in many different architectures, including ours. But the most important thing that I want to mention that you should know about about the technology transformation of our company is that we really didn't want to bias ourselves towards one single architecture. One of the things that we did day one at Liquid AI, we designed a search algorithm to, let's say, like, you know, let the algorithm instead of human biasing kind of the algorithm, let the algorithm run the scaling laws on, let's say, 100 different variations of operations that potentially can give you a general purpose computer. Computer. So we build a meta AI system. The paper around this is actually published like 2 and a half years ago. We publish a paper about, about the topic. It's called STAR, Automated Design of Tailored Architectures. So read about STAR. And STAR is a framework that brings all the dynamical systems with any format, including kind of variations of attention, into one format for us to be able to search through. Okay. So to see like for 4 criteria, what is the most optimal neural architecture, let's say, of choice for, let's say, a certain deployment. Number 1 criteria is memory. Like how much memory are you consuming on a given processor? Number 2 was the efficiency of computation, how fast you can operate. Number 3 is latency of operations. And number 4, do not lose accuracy on the performance. There are pure transformer models and then there are hybrid models that you can actually build. Hybrid models have like an essential component, like they have a little bit of transformers in them, but they're— but the rest of the kind of dynamical system and most of the dynamical system for the purpose of these 4 objective functions that I mentioned would be— you would change that and you can actually automate this whole framework to design foundation models in-house. The technology stack of Liquid Foundation, like Liquid Foundation models, is called Automated Foundation Model Design kind of algorithms. We call it AFMD. This automated framework is the one that explores architectures for a given kind of hardware. And guess what came out of like the first generation of the architectures that we started optimizing? It came double-gated convolution kind of mechanisms as 80% of the networks work being this. So when we run without a human bias, the gating mechanism that we had exactly in the Liquid Foundation— Liquid Neural Network's original paper, it actually shows up with this very, very similar kind of format in the final architecture that comes out of the search space.

Everybody, welcome to the health section of Moonshots brought to you by Fountain Life. You know, we talk about AI on this Moonshot podcast all the time. One of the most important things AI is going to be able to do for you besides educating your kids and helping you with your taxes is making sure that you're living a healthy lifestyle, that you get a chance to get to 100+. I'm here today with Dr. Dawn Musalem, the Chief Medical Officer of Fountain Life and a part of my medical team. Dawn, a pleasure. Great to be here. You know, the thing that people are concerned about most about living to 100 or 120 is their cognitive abilities, making sure they don't have dementia. And the numbers about dementia are problematic. Can you share what you've learned?

Such an important point. You're right. At Fountain Life, our members, the number one thing people are most concerned about is losing their brain health. Forgetting the name of their child, forgetting the face of their loved one. We know that when it comes to dementia, the conservative estimates are that 45% are entirely preventable. What was amazing is with the advanced testing we're doing at Fountain Life, one quarter of our members had advanced brain age. Wow. But what was really awesome is, again, back to that prevention, when we partnered it with healthy living— this gives me chills— eating healthier, moving our bodies, sleep, optimizing sleep is so important. You know what we saw? We saw that we improved that brain age by 26%. That is a big, big number to show that the majority of those individuals were able actually to improve the brain age.

And one of the things I love about Fountain is we're searching the world for the best therapeutics, the best approaches, and making sure we bring it to our members. So if having healthy brain function till 100, 120 is important to you, check out Fountain Life. Go to fountainlife.com/peter. Make sure you become the CEO of your own health. All right, now back to the episode. All right, our next story comes from Palmer Luckey, the founder of Oculus and now the chairman of the defense giant Anduril. It's, it's funny to call Anduril a defense giant, but it is. He's claiming that the modern patent system has become a national security liability. In his words, the entire Patent Office could be downloaded every morning, ripped off, and used to fight a war against you. The core problem is baked into what patents actually do. Patents are a requirement. If you want to get a patent, you have to teach a, a person skilled in the art how to actually, you know, create and use your device. So this disclosure of your invention and the exact words in patent law is in such full, clear, and concise and exact terms as to enable any person skilled in the art to make and use the same. So if you do that, you're effectively teaching the world how to use it and you're exchanging that, that sharing of your invention for roughly 20 years of exclusivity. Palmer argues that when a strategic adversary can simply harvest every file, ignore the legal protections, and weaponize the disclosed knowledge, you've handed them a free instruction manual to your best ideas. So just for some numbers, the US Patent Office receives about 600,000 applications annually. It grants a little over half of those, 323,000. Uh, that's 2025 data. Uh, interestingly patents granted have increased 40% in the last 5 years. My guess is that is secondary to AI. Palmer's proposed fix isn't to abolish patents. It's to massively scale up a national security patent process which goes back to the Secrecy Act of 1951. So this obscure mechanism lets inventors obtain classified patents in which you keep your exclusive rights but you don't disclose it to anyone, and neither can the government. So there are roughly 6,000 of these secure secrecy orders active in the US. Luckey wants that this Edge case is turned into default mechanism. So here's the question, right? If we genuinely trade this openness, which has been sort of the basis for American entrepreneurial exceptionalism, for a secret system, are we trading safety of having our patents ripped off against really the innovative ecosystem that we've had? Let's watch a short video from Palmer, and then we'll talk about it.

Stop patenting everything. Patents are Chinese instruction manual. The Founding Fathers never predicted a world where you would have a globalized economy where the entire patent office could be downloaded every single morning and then ripped off and then used to fight a war against We need to really fundamentally revisit the patent system. I think we need to massively expand the national security patent process. Uh, you can, you can obtain a classified patent. You can get a patent on something that you are not allowed to disclose to anyone, but you still maintain the exclusivity on those rights. We need to massively expand that program.

So, you know, I've applied for and gotten a dozen patents. I know, Alex, you have a, even a much larger number of them.. So I'm curious, guys, how do you come out on this? Alex, do you want to kick it off?

I, I think this is the episode of people, uh, tech CEOs floating terrible ideas. I think this is a terrible idea. I, I think the, uh, I would argue the Invention Secrecy Act of 1951, uh, which is I think what Palmer is gesturing at, has been probably on balance quite detrimental, not just to democracy, that if patents— so maybe a bit of context. The way the Invention Secrecy Act works is it's not that you can just sort of file the patent in secret and not disclose. It's that basically it can only be practiced, the invention that is basically confiscated or eminent domained by the military can only be practiced for military reasons. Not contra any construal otherwise that Invention Secrecy Act somehow offers legal cover for an individual to secretly disclose how their invention works, uh, under some confidentiality and then go practice it. In general, they can't. It's that the military exclusively can practice it, and then the inventor gets royalties from that practice. That may be good for Anduril's defense business, but I think in general terrible idea. Greater concern that I have is it— these are— these would be basically secret monopolies. I, I think it's bad enough that we have Invention Secrecy Act classification of inventions. Query whether entire swaths of technology that could be completely transformative economically to the entire world from an energy perspective, for other domains, have somehow without general knowledge, been swept up by the Invention Secrecy Act and basically confiscated by the Department of War for purely military reasons. That's very concerning to me, the idea of expanding it overall. I would argue, if anything, the Invention Secrecy Act regime should probably go away.

We can have this debate. So Palmer's going to be joining us at the Moonshots gathering on September 25th in LA. Everybody can go to moonshots.com. We have an amazing day with the Moonshot Mates there. Where I will be having these conversations with Palmer. Saleem, I mean, what makes America great is our open innovation policy, people building on top of other people's creations. What are your thoughts here?

Look, we've seen this problem get bigger and bigger over the last 20 to 30 years, okay, where the disclosure, especially in an age of AI where people can just route around it or replicate or learn from it. It's a huge challenge. The real moat is learning loops. That's going to be the real defensibilities. What are your feedback loops and can you learn in a proprietary way and then create trade secrets around that and action that in the marketplace? Continuous innovation is going to be the winning defense. It's not going to be ownership. The only people that win in this whole, in this particular model, are the lawyers.

Well said.

Dave Erickson, any thoughts Yeah, I think, you know, if there's a flashpoint for a World War III, this is probably one of the most likely. Seriously? Where— yeah, well, you know, look, Alex is right. We're going to discover new physics, new medicines at an incredible accelerating rate. And, you know, places like Europe respect intellectual property rights, and that creates a kind of a coherent economy where you can trade these things. China completely ignores intellectual property rights. And just takes it and runs with it. So, I think the likely outcome of that is the US will trade embargo anybody who doesn't respect intellectual property rights. And then you have to choose: are you part of the, you know, the free world, or you're part of the alternate world? But I think that's the more likely outcome, and that's going to happen soon, like in the next couple of years, because the rate of innovation is going to go through the roof. But there's no science fiction future book I've ever read where there isn't massive amounts of intellectual property being created by AI at an incredible accelerating rate, and there's some vehicle by which innovators can profit from that. And if you don't have that, then you don't have the future. You know, a huge fraction of brilliant thinkers coming out of, you know, Cambridge and MIT and Harvard don't work on foundational technologies because there's no money in it. And that's gotta change fundamentally. And protecting intellectual property rights is a key, key way to reverse that tide and get people working on really important things.

Yeah, I think to Dave Erickson's point also, Palmer fundamentally misunderstands— appears to misunderstand the nature of patents. The whole point of a patent is that you disclose how it works in return for a state-granted temporary monopoly on it. And say, you know, sort of bellyaching that the, the Chinese are running away with the disclosure, it is really a quibble with enforcement. Enforcement of patent. It's not— you don't want to throw necessarily the baby out with the bathwater and say we want to give away the patent trade of disclosure in return for temporary monopoly. Really what he should be asking is better enforcement of US patents in China. Agreed. All right.

I'm going to move us into the world of healthcare abundance. So two stories this week are demonstrating an incredible impact of AI on healthcare abundance, demonetizing and democratizing diagnostic for billions of people. The first story is the performance of GPT-5.6 SALL, which was released a couple weeks ago on HealthBench Professional, which is OpenAI's hardest medical benchmark. So ChatGPT or GPT-5.6 SALL set a brand new all-time benchmark high. And then the second part coming out here is in a blind test across roughly 20,000 individual physician judgments. In other words,, you know, diagnostic— diagnosing for accuracy, safety, completeness. GPT-5.6 answers were compared to specialty-matched physicians, in other words, pulmonologists, pediatricians, whatever, who were given unlimited, uh, full access to the web and unlimited time to answer, and the doctors still lost. So we've got ChatGPT— we've known this for some time— that these AI diagnostic models are better than the best physicians given all the tools that humans can use. The second part of the story comes from Meta. So OpenAI's own HealthBench Professional benchmark, which is 525 real clinical tasks, Meta's Muse Spark 1.1, again released last week, beat ChatGPT's, or GPT-5.6 Sol on across the marks, and it was 7 times cheaper, but even better, I mean, important to note here is that Muse Spark is free inside of all of Meta's products, you know, WhatsApp and Facebook, and Meta today serves 3.56 billion daily active users using their products. So here we've got a situation where the top medical AI capabilities are now free to over 3.5 billion people on the planet. And that's just extraordinary. I mean, this is the abundance thesis at large. And again, as people talk about the concerns of AI and so forth, please realize this: people who've never had access to the best diagnosticians now have them.

Salim, there is a, there's, there's a model in an AI doctor in China that's being used in rural environments by 100 million people already, right? Basically diagnosis has had massive cost collapse. The healthcare domain is particularly interesting because it's where abundance becomes actually morally urgent, right? If you can deliver way better first-in-line answers at, at like near zero cost, it's how quickly can you safely get it out there? That's the only question. And so, uh, it's absolutely— and right, let's recognize that in almost every country in the world world, there's radical doctor shortage. Yeah, so this is really, really critical.

You see, like, this is such a July 2026 story where— think about it— Instagram now gives better medical advice than a human doctor. That's— it's, it's pretty, pretty wild. It— cost of intelligence not just going too cheap to meter, cost of medical intelligence becoming too cheap to meter.

Free, basically free. I mean, that's—

well, the, the ultimate too cheap to meter is asymptotically free, right? But I would say probably, in all honesty, I suspect a little bit of mild benchmarking by Meta on this Meta Spark 1.1 is on, if you believe the AAII cost frontier analysis, it is on the optimal cost frontier, but it's not at the top. So if it's beating, say, Fable-5, which barely allows you to do anything biological, or GPT-5.6, which does, allows you to do it, that does to me suggest, in all honesty, a little bit of mild benchmarking, but still, it's a great day when Instagram gives better medical advice than human doctors.

I think that's our takeaway quote from today's pod. I'm going to move us to one more longevity story that I love. This is breaking news from yesterday, and it really got me excited here. I know you, Alex, and I were talking about this. For decades, one of the fundamental problems of aging is the slow accumulation of what are called advanced glycation end products. I love the acronym, it's called AGEs, A-G-E-S. And these are sugar molecules that cross-link and damage your proteins in your body over the course of time. So this chemical reaction is called glycation, and it happens slowly in our bodies as we age. It stiffens your arteries, it clouds your lenses, damages with cataracts, it damages kidneys, it wrinkles skins. And this idea is that it's always been irreversible until this week. And yesterday in Nature Communications, a team from a new startup called Revel Pharmaceuticals demonstrated an engineered enzyme called CMLase that acts like a molecular lawnmower. I love their description, a molecular lawnmower. It oxidizes away the glycation scars and restores the original healthy protein underneath. And amazingly, this isn't happening just in a test tube. They showed it worked in human tissue samples from elderly donors, reversing damage that accumulated over the lifetime. It's still early, but the significance of this cannot be overstated. A category in aging that we've always filed as permanent just became reversible. And again, we talk about longevity escape velocity, we talk about, you know, our ability to understand the 5 billion chemical reactions per second per cell in your 40 trillion cells. And when we talk about reaching LEV, you know, longevity escape velocity, by 2033, it's tech like this. So congrats to, to Revell in doing this.

And not just Revell. I mean, a couple of interesting notes here. It was Revell and Calico, the California Life —company that was one of the Alphabet other bets that's been, I would say, like a lot quieter than, say, Waymo. They're still doing work. That, that's very encouraging to me that Calico is apparently deeply involved in this and, and has a heartbeat. A couple of other points: the, the broader process here, a class of chemical reactions are called Maillard reactions. It's also the reason why when you bake bread that the outer crust is usually brown or stays— or chemical, or— yeah, or it's why— this vegetarian speaking— why everything purportedly tastes like chicken. Uh, it's the same class of reactions, but the sugar is reacting with the carbonyl functional group, or carbonyl groups within sugars reacting with, um, with the amines in, in proteins to, to create broad class of molecules that look optically brown. So the same thing is going on in the human body. To me, this is very exciting because it's not quite unscrambling eggs, but it's halfway there. It feels almost— again, strictly speaking, it's not like reversal of the thermodynamic arrow of time, but it's the next best thing if we can remove all of these unwanted sugar plus protein byproducts that are associated with inflammation and other correlates of aging with directed evolution of a protein that came from bacteria. Like, what else is there out there in the biosphere for us to mine in addition to all the obvious GLYP-1s? Great potential for longevity escape velocity. What other bacterial innovations can we use to turn back aging?

It's human engineering. We're taking control. It's going from evolution by natural selection to evolution by human direction.. And I love that.

I'll be happy when I have Ramin's hair. That's when I'll be happy.

Well, there are lots of companies working on that, Salim. So gentlemen, grateful for our time today. I'm excited for Starship 13 launch later today. Wish Elon and the group there lots of luck. Ramin, congrats on the success of Liquid AI and excited to have you on the pod with us. Dave Erickson, great move investing in Ramin. On behalf of all of our shareholders. Thank you. Gentlemen, have an amazing week. I'm sure we'll be having an emergency pod very soon because the speed of the singularity waits for nobody.