Recursive's $670M Bet on Self-Improving AI, Sonnet 5.5 Hits 70%, Elon Co-Leads Pentagon Push | EP #299

Richard, I'm curious, where are we at the moment with RSI?

In various weak forms, we already have RSI. We're not quite there yet, but we're very close.

When do you believe we reach ASI?

I think it will take us probably several decades.

On TerminalBench 4.0, Sonnet 5.5 jumped from 10% to 70%. Pretty extraordinary.

My theory would be that they're trying to Compete with, you know, China, which is about 3 months behind, and fill that gap before a lot of enterprises go to open-source models.

It's not at all obvious to me why anyone should be using Sonnet 5.5 over Opus 5.5 unless you have some token or latency or other consideration. I do not plan to use Sonnet 5.5.

The Defense Secretary announced Project Meridian. It's a new Pentagon effort on the future of warfare. It's co-led by Elon Musk and Palmer Luckey.

This, I think, is A transformative moment. This is a major, major step forward.

The big question is going to be—

Now that's a moonshot, ladies and gentlemen.

This episode is brought to you by the Abundance Summit and Link Ventures.

Welcome to Moonshots, everyone, your number one podcast in helping you understand the singularity. What's going on? What does it mean to you and your family? and what the breaking news is. We publish this twice a week to help you keep up with the extraordinary rate of change. With me today is the Fantastic Four. Alex Wiesner-Gross, our in-house ASI. Dave Blunden, our impresario of AI investing. Salim Ismail, our globetrotter who's home today. Amazing, Salim.

Woo-hoo!

And the father of the organizational singularity. I'm Peter Diamandis, your host and data-driven optimist. Our mission here: keep you optimistic about the future. And here's the numbers from this week. In seven days, Google shipped Gemini for Argon. Anthropic shipped Sonnet 5.5. OpenAI shipped GPT 6.1. Sol and the frontier intelligence has gotten cheaper by threefold. SpaceX launched the next batch of astronauts to the ISS, and also launched Google's TPUs into orbit for Project Suncatcher. If you're new to the Moonshots podcast, please hit subscribe. We publish twice a week, and you don't wanna miss any of this news during this hypersonic tsunami. Our moonshot here is to 20x our subscriber base to get to 10 million to help spread the word of optimism and the extraordinary future that we're building. Today on Moonshots, we're joined by one of the architects of modern AI, Richard Socher. Born in Germany and trained at Stanford, Richard is among the world's most cited natural language processing researchers, A pioneer in deep learning and prompt engineering, and a serial founder who's repeatedly turned frontier research into category-defining companies. First, he founded MetaMind, which was acquired by Salesforce, where he then became the chief scientist leading AI research for friend of the pod Marc Benioff. Next, Richard founded You.com, recognized by Time and the World Economic Forum, and now valued at north of $1.5 billion. as well as his fund AIX Ventures. Today, Richard is co-founder and CEO of Recursive, perhaps the biggest moonshot he's ever taken, focused on recursively self-improving superintelligence. Richard, congrats on a $670 million fundraise from Google Ventures, Greycroft, NVIDIA, and AMD, and $410 million in compute from AWS. Personal disclosure, I'm very proud to be a seed investor in Recursive. And of course, Richard, important to mention your new book just got released, The Eureka Machine: Why AI Is the Key to Unlocking a New Era of Scientific Discoveries. Welcome to the pod, Richard.

It's so great to be back.

Hey, and let me just add, a truly, truly awesome guy. A lot of people who listen to the pod are worried about the ethics and the risks and everything. But boy, if you want a person to conquer recursive self-improvement that you can like and trust, Richard is the man.

Yeah.

Thank you so much. So great to be back here. I love your guys' constructive optimism.

I have a bold—

So important in the world right now.

I have a bold prediction for this episode.

What's that?

This one episode will prove our thesis for this whole podcast more than any other episode we ever record.

Awesome. Wow.

Let's go.

Let's go.

And, uh, Alex, do you know Richard?

Yeah, Richard, you and I have chatted quite a bit. I don't think we've actually ever met in person though.

Really?

We have not, and it's the first time on a podcast together. It's gonna be awesome.

Amazing. Well, let's make history.

Yeah, for sure. So Richard, let's start with your book. The central claim of The Eureka Machine is that AI will deliver a century of scientific breakthroughs in the next decade. You know, this maps directly onto what Alex and I wrote in Solve Everything. So we're fans of that prediction. Your thesis in the book is that every stage of scientific process gets connected and transformed simultaneously. Hypothesis, experiment, data, theory. You call it the full-stack AI. You also say that scientific progress has slowed, and that's been caused not by underfunding, but by fragmentation. So let's start with those 2 items. Could you take a moment and, you know, talk to us about full-stack AI for science and why you think progress has slowed?

Yeah.

Yeah, so progress has slowed largely because we have so many different subdisciplines. And even Stanislav Lem realized, like, wow, if there are only so many people and we have more and more fragmentation of more and more subdisciplines and niches, you can not just do AI, right? You're doing often optimization and gradient descent methods and second-order derivative methods and so on of these neural networks, which is one category of AI and so on. And the same thing. And the same is true in biology. You study biology, you're going to be either a cell biologist or molecular biologist, or you're like in medicine. And like these— there's so much separation that it's hard to weave it all together. And that is the perfect time for AI to come in and help us weave back together all these separate pieces and also help us understand how these large complex systems actually work. Like, we can't have one model where humans say, here are all the things I know about natural language, and now, like, we can put all these rules together and then we have a conversation. It required a large neural network with a ton of data. And guess what? We're going to get a lot of data about biology. And so we can make more and more disciplines and move them, transfer them from being traditionally natural sciences where we just try to understand biology and nature to being programmable engineering sciences. And that, I think, is one of the many different exciting aspects of things to come. And why am I so excited that this is happening right now? And what are the major ingredients? It's essentially that we now have world knowledge in the form of LLMs. We have more and more scientific data that's getting digitized that we can sit on top of. We have better and better simulations of things. And we have robotic process automation that will soon be possible. And on top of that, we're going to have an agent swarm. And that is how the full scientific stack can be automated.

Alex? Yeah, I can't say I disagree. As Peter, you, and I wrote in Solve Everything, math, science, and engineering are cooked. I'm curious, Richard, what your latest timelines are.

I mean, I think every field is going to get to higher and higher levels of abstraction. Computer science has been very good at that, right? No one is programming in zeros and ones anymore. No one, like very few people have to still know C++ and complex pointers and memory architectures and so on. We can get larger and larger, and computer science has basically abstracted enough so that it met the rest of humanity with English, right, and natural language that now can be used to do computer science. And so I think that gives me great hope that we can get other fields into similar stages of abstraction and then all meet in natural language to do science.

But to pin you down, as I recall from my understanding of your book, 3 to 5 years for the physical sciences?

You know, like there's no like single sort of, this is the threshold and now we solved all of physical sciences, right? But like multiple different diseases will get cured in the next 12 months. They're gonna be the simpler diseases maybe where there's one gene that needs to be fixed for that disease to be cured. But there are a lot of single gene diseases that are out there in aggregate, right? And so those diseases will get cured and then we're gonna cure more and more complex diseases. We're gonna develop better and better battery material. So I wouldn't call it like Like one threshold and like in 3 years sort of everything is solved, but like we're gonna solve more and more problems and it'll just be a question of how much money do you wanna put into compute to then solve which kinds of problems?

You're talking to the guy, Richard, remember, who argues that the singularity itself is something of an optical illusion, that there is no step function, it's just a time interval. So I'll try once more, timelines. When, even if it's not a step function, When is the inflection point? When is the 50% point crossed in your mind for all of the physical sciences getting solved, all human disease getting solved? Where are those timelines?

Alex, if I could just add a layer on top of that. So we've talked on this pod a lot that math is cooked, right? We're seeing Millennium Prizes all falling. And the next natural, if you would, Barbecue is likely to be physics.

Incineration?

Incineration, yes.

And computer science, arguably, I mean, presumably this is part of the premise, Richard, of Recursive as well. Computer science already cooked, math thoroughly cooked, physics—

You know, in a weird way, I would use different metaphors. I think they're flourishing, not getting cooked. I think the weird thing is, you know, some mathematicians recently said, oh, this is maybe bad for the field because, like, we need to train people. And yes, we need to train people, but imagine a biologist or medical researcher saying, oh, you know, it's really a bummer for the field that we cured all these diseases for people. Like, that would be insane, right? Of course you want to just move as quickly as you can towards solving these problems. I think maybe the problem with math is that there's many subfields of math that haven't really had— they're just almost purely beautiful intellectual exercises without real-life impact anymore. But as you get close to real-life impact, you should be excited about it, and you think and realize your field is flourishing, not being cooked. And so To timelines, I think the further, the more complex a disease is, the more you have to do long-term studies with humans before you're allowed to put a certain drug into humans, the more delays you will have. But we see now companies, to make this very concrete, biotech companies used to have one new drug in development, and it took them 10 years. They had to go public before they knew the drug was really working. And then after 10 years, maybe late-stage 3 trials were just not working, and then the company is dead. Right now, you have the new age of companies. They have 5 to 10 different compounds in stage 3 trials after 2 or 3 years. And so we see a lot of acceleration in that level. But again, diseases, to really be able to get them into humans at scale in the United States with FDA approvals, there are just some natural delays that will be more like half a decade to a decade. But lots of other things in chemistry and physics and so on where we can iterate without having to look at long-term human trials will be even faster.

And I— the point is well taken. I like your Orwellian turn of phrase. Maybe instead of saying AI is cooking math, I should be using flourishing as a transitive verb and just say AI is flourishing math.

Actually, I was with Sertac Karaman, who runs LIDS at MIT, the night before last, He started a drone company now, an AI drone company. But LIDS is where radar was invented originally during World War II. It's a great lab, very entrepreneurial. But he said all his mathematician friends at MIT are aware that they're cooked and they're all—

Yeah, no, no, no, Dave, Dave, we're going with the Orwellian language now. They're not cooked. They're being flourished.

That's funny because he literally said cooked. But I'll go back to him and tell him, look, tell your friends they're flourishing.

Tell them they're always well.

You're being flourished.

Not to worry.

If your goal was to prove as many theorems as possible in your lifetime, now is the time to just grab as many as you can and work with AI to solve them. And then I think the field will change the way computer science has changed in many ways. And it'll be much more about what's the most creative thing when you really understand all the things that are out there in math, what kind of new formalisms can you create, new constructs can you create that then would be interesting to be solved by an AI in collaboration with humans?

Yeah. Well, Sertas was saying the math guys are in great shape because they're so cooked that they're all moving over to AI orchestration and they're going to be way ahead of the curve. He's actually most worried about the biology professors who are in complete denial, using virtually no AI in their day-to-day activities. And this is one area where, Richard, you have such a deep background in all facets of AI, including biology. And I feel like we're living exactly parallel lives, except you're 15 years younger than me. So I'm like insanely jealous of, of your life trajectory. But be a serial entrepreneur, then start a hugely successful venture fund, then found a foundation model company. And, you know, we're seeing the world through the exact same lens. But you and Pieter have much more biology background. I have basically none. But the biologists are the ones really, really lagging. And I think you've got a bunch of investments, actually, that are— that have done really well in the area, right?

And I would love to talk about some of those, like Parallel Bio, Proxima Labs, Ignota Labs, truly exciting companies. I think another field that is even more lacking than biology is economics. Economics literally has these models of like a linear model of economics, right? A one-step economy that's provably correctly taxed and subsidized and things like that. It's just like absurd how slow that field is to adopt AI for making better policy decisions.

I just want to throw this out there.

I just want to throw this out there.

Because you opened the door, Erik Brynjolfsson, our very good friend, I did not realize till yesterday that— I love Erik. HAI Lab, HAI Lab at Stanford invented the word foundation model. And he sent me the whole— He did, yeah. You knew that already?

Wow.

Yeah, of course. Lots of Stanford friends and Percy Liang and others worked on foundation models. I love Erik Brynjolfsson. He's actually one of the most interesting economists, I think, doing really interesting research right now. He also started Workhelix, which we're a proud investor in, that brings understanding of how companies actually adopt AI. And which tasks are getting, you know, helped by in real rollouts for companies. Yeah, he's a co-founder of that too.

Yeah, that building is just such a great place because you talk about econ people being way off the curve, but he is so ahead of the curve. You walk into the building and you can just feel it. He's also poaching a ton of talent from MIT to come out to join his lab.

Following in his path. Richard, talk about full-stack AI as you see it in the scientific method. You know, I'm an investor in a company called Lila out of MIT and Harvard. You know Lila Science as well. I think I've introduced you to Geoff von Maltzen there. And, you know, they're building a scientific superintelligence that is then running a million square foot of robotic space to sort of like 1,000x the rate of discovery. Your thoughts on that?

I absolutely love it. I think that is, you know, in the Eureka Machine, I sort of lay out these 4 pillars, and they're actually one of the few that are really going after this, similar to Periodic Labs, who are doing it more on the physics and chemistry side of things. And there are several other companies now that I can hopefully soon talk about that are trying to create more data for the bitter lesson to be applicable in biology. And I love what the big guys, Eli Lilly and others do, and Lila too. I think ultimately you can boil down the scientific method to the ideation, implementation, and validation of ideas. And the faster we can close that loop and then put an open-ended process on top of it. That's what open-endedness, what inspired us a lot at Recursive. And we have many of the world's greatest researchers in that subdomain of AI that is still not quite as popular as it could and should be. The more you can have a swarm innovate in open-ended ways and evolve and combine interestingly different ideas, the better. But then, of course, in biology, you have to have like actual physical biological wetware experiments. And so it's really great to see Lila doing that. We're seeing this also— maybe I can talk about this one company that I really love called Parallel Bio. They build tiny organoids in a petri dish.

Right.

And get this, if you love animals, you too can love AI. Why? Because they got FDA approval to skip animal trials. Turns out we cured most diseases in mice. It's not that helpful. They're very different to people. But these guys use pluripotent stem cells, create tiny little organoids of lymph nodes. Lymph nodes, big part of your immune system. Immunotherapy is some of the most exciting therapies to allow your own immune system to attack a cancer and so on instead of getting crazy chemotherapy and so on. And so they got FDA approval because they showed that when you run experiments in parallel, hence Parallel Bio, in these tiny Petri dishes, how those organoids react to different drugs and toxicity testing and so on, is actually more predictive of how those drugs will interact in real human bodies. And that is just one of many beautiful examples of where it will help.

And you can take my stem cell, you can take my skin, create an iPSC cell, grow my own organs, and see how a particular drug would work for me versus a generic individual. Selim, you want to jump in?

Yeah, a couple of things. One is, you know, science has always been a coordination problem. Right. You're kind of trying to bring— and it's always been very structured into departments and journals and all that stuff. And I love what you're doing, bringing it together. For me, if I had to summarize what you seem to be doing is you're collapsing the time between an experiment and a result or hypothesis and a result. And when you can collapse that, that's domain collapse of the scientific method. And now, like, anything is possible from there. This is amazing.

Yeah.

I'm going to push again on what Alex said. Physics. When do you— Einstein's theory of relativity comes out. There's not that much progress since then, a lot of theories. Do you see physics as the next domain that's going to flourish on the back of math?

I think the biggest domain is actually going to be biology. Physics has been interestingly stuck in many ways. really powerful ideas like E= mc². So you can get a ton of energy out of potentially little mass. And then we got nuclear energy out of that. And so that then moved into and graduated in some ways into an engineering discipline, which is where you have the real-life impact. I think, obviously, fusion will be great to finally get figured out. And there's a lot of really cool companies and big labs and tokamaks already to balance plasma inside tokamaks is a very hard control problem where AI is being used. I hope we can eventually make better theories for quantum gravity and all kinds of other complex issues and get a better world sort of formula. Unfortunately, the dataset collections these days require often like large Hadron colliders and just like billions of dollars. And it's quite expensive, actually. So in a weird way, biology is a better fit for AI because what calculus did for physics, like understanding really well micro-individualized separated phenomena, neural nets are great in combining all these little things. Like we know what a neuron does, we know what one bacteria does in our microbiome, but then as they all come together and form these really complex interactions, we don't really know anymore how that works. And so that's where AI, I think, will help us more.

Yeah.

Alex.

Let's talk about biology a bit. So do you think the critical path to solving biology goes through digital twins of cells, or do you have some wildly different theory of the case?

3 years ago when I started the Eureka Machine book, I had this chapter on the virtual cell, and I was really proud of like trying to lay it all out. And I had to rewrite that whole thing because in the meantime, uh, you know, Mark Zuckerberg Foundation started a virtual cell.

Everyone's doing it. Everyone, not just CZI. Everyone has a virtual cell model.

That's fine.

Even if Dave doesn't think Dave has a virtual cell model, I'm sure he'll spin one up soon.

So I think that most— so there's this famous paper by Rich Sutton on the bitter lesson in AI.

Never heard of it. Tell me all about the bitter lesson. This is a new concept for me. What is this you speak of?

I don't know. You're probably joking, but maybe some readers haven't heard of it. So the bitter lesson is basically that human experts had all these really clever ideas and beautiful theories, But really what you needed to do was to make real progress is the simplest method that you can come up with, a large neural network that you can train end-to-end as a general function approximator, the simplest model you can come up with, and then just a ton of data and compute to actually train that model. And that usually outperforms at scale all the clever little hacks that human experts had come up with before. And so the bitter lesson has worked for NLP. 10, 20 years ago, you would have asked an NLP expert, can you have one model to have any conversation with you, prompt of any kind of question. Prompt engineering I invented and was nicely cited by the early GPT papers. But the same thing, the same state biology is currently in. There are so many complexities and the experts know so much, they feel like there's no way you can instill all of that knowledge into one model. But if you have enough data, you can. And now you have companies like Tahoe Therapeutics and so on that are creating these massive perturbation studies. Yes, all models are wrong, but more and more of them will be useful as we collect more and more data about biology. And that is, I think, one of the biggest driving factors for the impact of AI.

But so just to answer the question, do you think virtual cell models are the critical path to solving all disease or solving biology, or is it something else?

I think they're definitely going to play a crucial part. It's the 3rd pillar of the Eureka Machine. We do need to have anything you can simulate, Anything you can verify, AI will then be able to solve things in that domain. And so yes, I do think virtual models and simulations are extremely important.

I'm going to jump into 3 fun breaking stories this week, 2 on science and 1 on AI. And Richard, they relate back to your work and your book. Our first story was published in MIT Tech Review today. It tells the story of researchers at Israel's Weizmann Institute who built a system called BrainIT. The system reconstructs the image a person is looking at while inside a functional MRI machine. I'm just post— put up the slide here. So take a look at this image. You know, it's pretty extraordinary. On one side is the image that a person is looking at inside the fMRI machine. On the other is the AI reconstruction. You know, so earlier brain image decoders could tell you were looking at a dog or a clock tower, but it lost the color, the composition, and the details. So BrainIT learns structure and meaning separately and then puts the picture back together. And here's the clever part. They also can run the model in reverse, an encoder that predicts how a brain will respond to an image. So they can feed it images no human has ever seen, into the scanner, generate predicted brain scans, and then train on those. The model effectively builds its own dataset. So Richard, you know, you say in your book that AI is superhuman in any domain you can simulate or verify and nowhere else. So is a brain simulatable— is the brain a simulatable domain, or is this something different?

It's not yet, but this is still an incredible result. I still remember the first such result came out when I was still a PhD student. and I went over to the bioinformatics department at Stanford, and they, you know, it was like very grainy little images we could just like extract from this. And to see this fidelity now and this realism is just incredible. I think one of the problems, I don't know if this study, I haven't seen it today yet, I've been at work, but like one of the problems is often that you have to train for each brain 'cause they're all slightly different.

Mm-hmm.

And so you won't be able to just take someone who you've never put into an fMRI scanner to see and update the model for that particular person, which would be otherwise amazing for people who are like in a coma and you want to see, are they still thinking stuff? You know, like, that would be incredible.

Alex, tell us more about this story today.

Yeah, so this actually is a result from March. So this, the broader field of doing decoding of brain states from functional imaging or EEG or electrodes, this has a long, long history at this point. I remember 20-plus years ago at this point, the Gallant Lab at UC Berkeley was doing this to decode the visual cortex of cats, actually, like, 20-plus years ago, we were starting to see images of how a cat perceives the world. And I remember, like, the first grainy images that were coming out were like a cat seeing a branch. And then moving, fast-forwarding to non-invasive fMRI, there are so many groups at this point. Meta in particular has been sponsoring quite a bit of work in this area from Jean-Marie King, many other groups doing language decoding, vision decoding. More recently, even fast-forwarding to just the past 24 hours, Neuralink just announced their first scaling law studies on pre-training foundation models, frontier models, off of Neuralink electronic data. And this is on an individual patient basis, but they're capturing copious amounts of data at high temporal resolution, much higher than fMRI can capture. fMRI, like rule of thumb, fMRI can give you, with today's technology, at best, cubic millimeter spatial resolution for voxels, and at best approximately 1-second temporal resolution. So that somewhat limits how well you can do in terms of decoding a person's internal visual state. You can still, make no mistake, there are folks who have— are making a cottage industry out of decoding dreams and decoding visually what a person perceives or what they hallucinate in their visual cortex while they're sleeping or while they're awake. And I'm sure this is going to be a vibrant, vibrant space training frontier models and foundation models off of copious amounts of fMRI data. But ultimately, where I think this has to go— there are spatial and temporal limits to fMRI decoding. It's going to require higher temporal precision and higher spatial precision to get where we really want to go, which is full-dive VR and full-bandwidth BCIs. And we will get there. But the good news is you have to start somewhere. That large language models had to start with GPT-1, Just a single artificial neuron that could predict the polarity of Amazon reviews. Similarly here, if we're going to get full-bandwidth BCIs, it's going to start with studies like BrainIT and John-Marie King and Jack Gallant.

Amazing. Saleem, what happens in the world where you can know someone's thoughts?

Well, I think you— let's be careful. We're decoding the perception here, not the actual thoughts. Right? That's like quite a bit more complex than that.

What's the difference?

What's the—

other than the brain regions, the neural distinction without a difference?

No, no, I think there is a difference.

No, when you imagine, right? When you imagine something in your mind, it lights up the same neurons as when you're seeing them.

That's fine. That's fine. Just, I just want to make sure we don't mix the two because you can have thoughts without necessarily the perceptions around it, but you can definitely have one without the other and the other without the other. So, but I think for me, the more interesting thing here is, you know, we've spent billions of hours typing things into little rectangles, and if we can change that interface to do a more natural interface, which is interface directly with the brain— we used to talk about this in our Singularity University lectures on neuroscience— is, you know, we still have very little idea how the brain works, but you don't need to know how it works as long as you can interface effectively with it. And this, I think, gives us the opening to give us really deep I'll take the—

You bring up a really—

oh, go ahead, Richard.

Sorry.

You bring up a really good point, which is I read this result a while back that there are 2 types of people. Some people who think actually think in sentences, and other people, when they think, they just have a fuzzy thought cloud. And only once they're asked to verbalize their thought or they try to write it down, they actually like get it out.

And I think that's a really good point.

make a real sentence out of their fuzzy thought clouds. And I'm definitely a thought cloud thinker. My wife is very much a sentence, like in her head, she actually thinks in actual sentences. And there's no one is smarter than the other or anything. They're equivalent. But I do think sometimes writing is thinking. And for me, it is very helpful if I have to really verbalize something versus I just have the thought. I'm just trying to think, what does the world look like if the AI could really just, like, extract my thoughts, and now I need to, like, structure my thought clouds into sentences more and more precisely.

Richard, what would neural net training look like if you took all text out of it? No language at all, just train on pure images and higher-level thought constructs. You'd get a very different neural net out the other side, and it might actually think a lot more like you do and less like your wife does.

Maybe we should be taking the position that anyone who doesn't have an inner monologue is just a peacock.

I'll do— I gotta throw one more thing in. I remember we had one of the top linguists in the world come in and speak, and we were like, you know, this must be a bad time to be a linguist. Nobody's using spelling or grammar or anything else like that. And he goes, no, on the contrary, this is one of the most interesting times to be a linguist ever. And we're like, wow, how come? And he said, because when you look at emojis, it's the first time in the history of human language that you can symbolically send emotion, because you can digitally send emotions. And we're like, wow, that's an interesting— so he was super excited by opening up that whole aperture. I thought it was an interesting—

You're saying, you're saying, you're saying that emojis are the reason why linguistics is so exciting, not the fact that linguistics itself has been flourished by AI at this point?

I'm not connecting the two. I'm not connecting the two. You did that. That was not me.

Dave?

Oh no, the result I really am looking forward to is, uh, How does the brain train itself without gradient descent? Everything going on in AI right now, everything going on for the last 20 years in AI is driven by gradient descent algorithms, and the biologists can't find anything even vaguely like that in actual biology. And so that's, you know, I think with really detailed imaging, we might finally crack the code on what is the fundamental learning algorithm that changes the synaptic weights. It's not But it's not what we use for artificial neural nets. It's something different. And if we discover that, we might find that neural net training can be 10, 100, 1,000, 1 million times more efficient. And so hopefully that'll come out within a year.

Could you define gradient descent for our listeners?

It's funny, Ilya Sutskever, when he's on interviews, he's like, there is only one algorithm. The algorithm is gradient descent. Everything else is just irrelevant compared to— so what happens right now is you build these 96 or 120-layer deep neural And all it is is a whole bunch of random connections that do absolutely nothing useful. And then you give it trillions of— 15 trillion training examples. You're like, when you see this, say that, or when you see this paragraph, this is the next token. And if you're guessing wrong, you get punished through gradient descent. So the error, how far off you are, is your punishment. And it passes back through all the layers and it calculates an error and it blames each neuron or each connection. For how responsible were you for this terrible answer? And if you're way wrong, you get moved, and you move in the direction that helps you get the answer right. So it's this incredibly laborious search process, and that movement of each synaptic connection is a gradient, and you just hill climb to the best possible state. And then magically, after about $100 million of compute, or in Richard's case, I guess $400 million, it magically starts thinking. Through this one algorithm, gradient descent.

Yeah, I think it's really backprop that we're talking about synecdochically rather than gradient descent would be the obvious comment.

So maybe, yeah, backprop is, I guess, the efficient computation of those gradients. Maybe I'll add to this explanation from Dave, but like gradient descent is essentially an optimization algorithm that we use to train all kinds of machine learning models by minimizing the errors. And I think The best analogy is to assume you're a blind hiker. You're at the top of the mountain and you're trying to find the lowest point. And you can't really see that far, you're blind, but you can feel the slope at each step. And so how much, like, how big of a step should you take down when you feel like the slope is going roughly in the right direction? And then the problem is that the kinds of landscapes these AIs are trying to optimize are highly non-convex, which means there's not just one lowest point, but there are many different low points. And depending on where you start, you might go into a different valley. Depending on which mountain you start, you go into this valley versus this different valley. And there are actually a lot of really interesting analogies. My friend Jishan, like, thinks about this from the psychological perspective. For instance, when you have PTSD and you train, you overfit to one algorithm, now your brain is kind of stuck in one valley of how to think about something. Mm-hmm. And it takes sometimes people like doing psychedelics and stuff that have shown to help with like depression and PTSD and other mental diseases. Like, it helps to increase the learning rate to jump over another mountain and go into a different valley of attraction where you then have new kinds of ways of thinking. And so this blind hiker analogy, I think, is a really intuitive way to think about it. And then, of course, the problem is it's not just in 3D. It's in a trillion D, like a trillion parameters that you're trying to traverse in that landscape. But the ideas of, like, try to identify the direction and take steps towards it still works.

And you know what else is really incredible? I could riff on this for hours, but if you look at a big model like KIMI-K3, it's 93 layers, and it's trained, trained, trained, trained, trained. If you take a single layer and you randomise it, and then you train it and say, find yourself again, It can't find itself again. It never gets back to where it was, which comes back to in these fMRIs when you're looking at a human brain and you're saying, okay, here's Alex's brain. Can I compare that to Saleem's? And you're like, wow, there's nothing in common going on here yet. Yet, you know, well, okay, maybe that's a bad example, but—

But you're going to take that figure as probably right.

It's probably right.

The way your brain wires as you learn is completely unique to you. Yet, yet it can come out with like, okay, we're equally good soccer players. But the way the signals are propagating through our networks is completely different and unique to each of us, which is really, really strange, you know, that it can't— it can't find its own way back to its state. I think we're going to learn a lot, actually, with the fMRI data coming in on how and why that works. And there have to be commonalities that we just can't find. You know, rotations, you know, make everything look different, but they're really not that different if you put them through a transform, like a Fourier transform view will suddenly say, oh my God, This is why they line up.

Yeah, I was just with my fraternity brother of ours, Dave, who's the head of neurobiology at USC the other day. And he's, you know, this is the most exciting time ever for brain science, right? I mean, we're, you know, the brain has been a black box since humanity began and our ability to understand the brain and deal with with, uh, you know, mental disease, I think, is going to be extraordinary.

Yeah, well, so you can simulate— this is why Liquid AI was founded, actually, because we completely reverse engineered the, the C. elegans worm brain down to exactly every single thing going on. And then they were able to simulate it. And then after they simulated it, they realized, wait, this is a very efficient neural net. And then they productized it. I mean, that's, you know, so they both So go ahead, Richard.

There's a really cool thing since you mentioned C. elegans worms. There's not many people that bring that up. A friend of mine who's a professor of neuroscience at Harvard, Sam Gershman, he— one of the most brilliant people I've ever met— like, he actually did a study with worms where he cut them in half. And one— C. elegans worms, they do have a brain, right? And then they have a lot of other, like, neurons in the rest. Now, you can train them to react to certain stimuli, and then you split them in half, The second half without the brain regrows a new brain. And here, get this, this is crazy. That new brain has the same memories.

Wow.

And reacts to different stimuli. And so he's like, maybe there's a different way where we learn. And I do think there are things that right now we're very much stuck in this, like, it's a neural net, it's all electrical signals and so on. But clearly, like, brain chemistry can change massively with just like, you can get hangry, you can be in pain, you can You can have like a certain like stimulant and all of a sudden your brain is very different. And I think there are, and I've seen this now with like having babies and like talking to moms, like there's certain algorithms where you all of a sudden you nest because you had a baby and now you want to like, and how you nest is different, but people nest, you know, in one form or another. And so there are these latent algorithms that we have that are in our DNA that just trigger after like It takes like 20-something years, right, when just the right things then happen with your body and your biology. So I think there's still so much that we don't— we haven't figured out yet.

So much complexity. I'm going to move us to our next story from 3 days ago. So on Tuesday, the president signed an executive order directing the EPA, the Department of the Interior and Agriculture, working with HHS, to cut invasive mosquito populations in Washington, D.C. by at least 90%. And the tick population by at least 50% by 2028. The targets include mosquito species that spread dengue, Zika, and yellow fever. And here's what caught my eye. You know, the order explicitly prioritizes sterile insect techniques, safe genetic modifications, and beneficial bacteria over conventional pesticides. Right? I love that. You know, this is the exact kind of gene drive work that Colossal, the de-extinction company, is doing. I think most people don't realize, it's like when I was raising my kids, I would ask them a question, which is the species that kills the most humans on the planet? And some people jump onto sharks, some people jump onto whatever it might be. But on this pod, I'm sure people know mosquitoes are the deadliest life form on Earth.

By far.

Right? Malaria alone kills half a million people a year. This is biology as engineering arriving as federal policy. Your thoughts on this story?

For decades, we've been scared of our shadow. So I have a classmate, Kevin Esvelt, at MIT, who was a pioneer of the gene drive, one of the pioneers, and he's had a devil of a time getting states, municipalities to approve gene drive studies against mosquitoes. And for those not tracking, the premise of this technique is basically inserting a gene via CRISPR that wants to replicate itself basically sterilize mosquitoes through propagating virally through the mosquito population. We have the technology to do this. What it has lacked, at least in this country, is a federal mandate to actually go and implement large-scale genetic engineering of non-human animal populations. And now, for the first time it seems, we're seeing via executive fiat just that mandate. And it is so exciting for 2 reasons. One, Because one, as Peter, you mentioned, malaria and other non-human animal-borne diseases cause an enormous amount of suffering, human suffering and also non-human animal suffering. But secondly, because there's a way to do this that doesn't actually involve killing the animals themselves.

Exactly.

Historically, you know, if you look back 50 years, you'd see chemical spraying if you wanted to do something about, say, insect Borne illness responses. We don't need to do that anymore. We can actually keep the insects that are the inadvertent carriers of bacterial or viral disease— we can actually preserve their lives while also preventing the disease transmission. And I think that's good from their perspective as well.

That was the part that got me most excited, was we spray things, we're basically poisoning biology, and now we can actually program the system. And leave the poison in the ecosystem.

Yeah. Richard, your thoughts on this story?

I concur with everyone. It's wonderful, and it's so interesting. I hope we can amplify these kinds of stories more. The future needs better marketing. We need more Peters in the world. And this is one of those many stories curing various diseases, Is Parrot Bio like saving animal lives and prevent them from just being bred to be tested upon and dissected? Like, those are all like stories we need to amplify more in the public eye.

Yeah, Dave, I can't wait for this to come from DC to Vermont and Massachusetts and all of these.

I was gonna say, for the listeners that are interested in your real estate fund, which, you know, the theme there is, hey, easy access via drone to hilltops The islands, but also if you live on the edge of a marsh, your house probably sells for about half the price of a place that's not on the edge of a marsh. That'll go away. There's so many solutions coming to biting insects, this being one of them, that that also is, you know, if it's beautiful, beautiful land that otherwise is very difficult to live on, it's going to go through the roof in value.

I do think it's really important to try to like, you know, keep the bees like alive, right? And not that Let's spray a bunch of pesticides, make sure birds still have enough insects to eat and things like that.

I was going to say, for maybe up to 80 years, humanity has been scared of its own shadow. Humanity in general, America or the West in particular. Nuclear energy, another example I've pointed out on the pod where we arguably lost 50-plus years of progress because we were too scared of either the bomb or fission reactors or some— an entire generation watched the movie China Syndrome and then got scared of nuclear power unnecessarily. Same idea with this, or with geoengineering, another example. The world is still scared. Some fraction of the world is scared of engineering the weather. We don't need to be scared of engineering our physical world or our biological world. And now hopefully one can see green shoots of humanity getting past that stage of worrying about its own shadow.

This is interesting that— sorry, yeah, it's interesting indeed that like sometimes alarmists feel like they're doing the right thing by just like, oh, How bad could it be? I'm warning people of something bad. But like, it can indeed push all of humanity away from something really good, like energy abundance with nuclear. I think overpopulation is one of these other big myths where people are like, well, if we have too many people, there's going to be scarcity of all these different resources and everything's going to get more expensive and people are— it's more poverty and so on. And the exact opposite happened.

Yeah.

There's actually one website that I think you all would love, which is called humanprogress.org or .com. Like that shows that a lot of these things are actually getting cheaper and cheaper despite there being more people.

Yeah.

Hey, Richard, you're a hero in Germany, like can't walk down the street kind of hero in Germany. What do you think about the fact that Germany has no power for exactly this reason?

It is really unfortunate. There are a lot of people in Germany that want to, in a weird way, sort of off-ramp from progress. And they think that everything that consumes power is bad for the environment. There's like a weird sort of degrowth offshoot from generally like well-intentioned pro-environmental vibes that just worry me, worry me quite a bit, yeah.

We're gonna talk about doomerism in a little bit, but not yet. And one of the things I realized a long time ago is we humans are really incredible at seeing a problem out in the future, right? We see acid rain, we see overpopulation, we see, You know, energy shortage, whatever it might be. And then because of our amygdala, because of the way we think, we accelerate that future problem to today and we freak out. And we forget the fact that there is a decade worth of progress we're going to make by the time we reach that problem. And that's what entrepreneurs do. They solve problem after problem after problem. And I guess I just want our listeners to hear that because if you're worried about some future problem, please understand there's an incredibly efficient market of entrepreneurship and capitalism that will solve the world's biggest problems, the world's biggest business opportunities, right? And it's a beautiful forward propagation solution set that we have.

And maybe we'll get there later, but I see the same thing happen in AI doomerism where they create more and more complex scenarios where attackers get these like near magical abilities to attack, but somehow the defenders in those stories never get near magical abilities to defend. And it's this weird thing of like, man, if I can create a super virus that is like perfectly undetectable and spreads all throughout the world and no one notices it, and like, and then it has this Wi-Fi switch and you just turn it on and off and then kill people, like, I will create a super magical vaccine that just inoculates you against all of these things.

Yeah.

Like, if you can actually assume, make that assumption, And like, it's a weird, yeah, weird thing.

This episode is sponsored by Google for Startups. Think about this for a second. You now have access to the same generative AI models that cost hundreds of millions of dollars to train. Google's Startup Technical Guide for Generative Media gives you a complete blueprint for deploying Google DeepMind's models in production. Images, video, audio, all of it. Real architecture, real results. Find the link in the show notes below. I'm gonna move us to a story that's been sort of breaking the internet over the last 24 hours. And it puts your regulate the application argument to the test, Richard. So a company called Tavis unveiled Griffin, which it calls the first model ever to pass the video touring test. So here's the numbers, 48% of people who talk to an AI avatar on a live face-to-face video thought they were talking to a real human. Previous systems were at 3%. So let me show the video here, and it's super cool, and let's talk about it next.

Now how about just a thumbs up? Sure, one thumbs up coming right up. Hell yeah, that was awesome. I would say you're my favorite coworker. It's been, it's been really great to work on this with you.

You're making me blush.

Who's that with the ugly sweater? Simon says touch your hair. Is this what you're looking for?

Yeah.

Okay, nice. Simon says point.

All right, I'm doing it now.

Touch your chin.

I'm not gonna fall for that. You didn't say the magic words.

Uh, you got me.

Okay, the loose connector back onto the motherboard.

Okay, let me go grab that. But beforehand, yeah, I turned the soldering iron on.

I think it's going to take about 12 seconds, so maybe just let me know when that's done.

Sounds good.

I'll keep an eye on it and let you know the second it hits that 12-second mark.

All right. So interesting. So Tavis calls it the first human interaction model. It's fully duplex, listening and talking at the same time that you do, the way humans actually do. It's number one on NVIDIA's benchmark for full duplex AI video. Tavis pitches it as, quote, a tutor for every student that notices when they're lost, an elder care companion that listens. Tavis itself says Griffin requires safety work before it can be released publicly because it's the first model that can be mistaken for a real person. Salim, your thoughts on this one?

So I'm less kind of floored by the imagery and the mimicry of it. This was expected to happen. For me, I'm more interested in can it accomplish something useful enough that I don't care whether it's human or not? And I think that is more interesting, which might come along and have— we have that happen at some point.

We've talked about the idea that in the near future, like now, You're going to have digital coworkers that pop up on Zoom, that you call, that you text, you Slack, and interact with you. And I think it's pretty compelling as a mechanism for a future, you know, virtualized company. Alex?

Well, first, Peter, you should probably ask me to touch my face.

Okay. Touch your face. Simon says touch your face.

Okay.

Very good.

Just checking. But I guess Tavis is ahead of me. I think on the one hand, if you look at models coming out of Chinese labs like Alibaba's lab, Wanstreamer gave us a preview of what was going to happen. The Chinese labs remain overinvested relative to the Western labs in generative video models and interactive generative video models. So Wanstreamer, I think, is a preview of what's possible. My guess is— I read the information that was put out around Tavis. I think in full generality, let's just talk about where this is going to end up. It's very difficult to predict the short term. I think it's pretty easy to predict the long term. In the long term, end-to-end generative pixels will have these magic mirrors that could be real-time, fully interactive, pixel-wise generated, or if Anthropic has its way, vector-wise or procedurally generated. But either way, Fully generated, real-time interactive video models. And you'll be able to create a scene that consists of people. Like, it's possible right now, but the latency is high. You see with Wand Streamer, which is already out and already open source, or with this Tavis Griffin type model, which is not really out yet and definitely not open source, you see a preview of the future. What does this look like? Well, there are a few different ways it can go. Query how revenue generating per token it is. Is it anywhere close to the optimal frontier of code generation? Doubt it. On the other hand, if it becomes so absurdly inexpensive to be able to generate arbitrary, say, humans participating in a Zoom meeting or humans participating in a podcast, do we really care whether it's close to being near the optimal cost-performance frontier. Maybe not. There are a lot of human service industry jobs that require a face and require interactivity and a voice and the ability to touch one's face apparently on demand or on request that could be probably completely automated away by a model that otherwise would be limited to text-based interaction but doesn't have a face And a voice. So in the most optimistic scenario, these sorts of interactive video models, open paren, note that the acronym for this, which by the way was there first with Alex Finn, HIM, H-I-M, is such an obvious reference to Her, the movie, close paren. I think that this has—

You should use em dashes, please.

Okay, I'll delve into it. So I think this is going to be transformative for the service sector. And hopefully Tavis and the broader American West of video frontier models and interactive video frontier models take a page from Tavis and start competing with China.

Richard, one of the principles in your book, one of the rules you have, is that AI becomes superhuman where you can verify the answer. And, you know, is like passing a human scientific milestone like this, is passing sort of a human judgment and interaction, you know, sufficiently as a scientific milestone for verification?

Not quite. It's sort of like, it's hard to scale when a human has to be in the loop to say, yes, this is like a human, another person on the other side or not. You can't quite automate it completely and verify it. verify it in that automated fashion. I do think you're right. This is a good example where I think we need to go beyond KYC and do KYU, know your use case, right? Like, you don't want this technology to be in Zoom pretending to be the CEO. And there's some famous stories where this has already happened, where they got someone to wire like $20 million.

Sure.

Because they created a Zoom with 5 other executives, and they all like— he was fooled well And so I think we need to, yeah, to be careful and like Zoom probably and Google Hangout need to start like finding countermeasures to identify, is this a real person and things like that to prevent those kinds of hacks from happening.

Yeah, scams are gonna proliferate on this. I just wanna do a call out again. I've said this before on the pod, and if you've not heard this before, if you still have your grandparents with you or your parents, and you've not taken the time to record their stories on video and audio and go deep, spend, you know, hours recording them, right? Your ability to create a super high-resolution lifelike avatar for your kids or your grandkids or your great-grandkids, I think is an incredible thing, but you need the data. So if you're listening to this and you're lucky enough to have your parents or grandparents around, Collect their data, 'cause it's gonna be a beautiful opportunity for your progeny and theirs.

And/or, I should add, get them an Alcor membership and preserve their connectome if there is a consideration. Like, why stop with just the behavioral data? Preserve the whole parent.

Sure.

You know, my kids are in a deep freeze, or at least their placental cells.

But I have a pushback on that just from this conversation. If from the earlier story we can read people's brains, all you have to do is picture an image of your grandmother and then extract it from that.

It'll be low fidelity. I mean, I do think, like, to first order, this is how ancestor simulation will work. You see stories left and right just in the past 48 hours of people using Opus 5.5 or other frontier models to reconstruct bits of history that were otherwise unknown just based on artifacts of the day. But at some point, I do think, Salim, you want to go beyond just the memory of the person and you want the actual person. And I'll pound the drum again, in addition to Peter, you know, very generous of you to preserve the placenta associated with your children, but not necessarily your children. I'm sure the placenta— placentae will thank you.

Yeah, go ahead, finish, please.

But I would say, like, yeah, if this is a serious concern, you're worried about your parents or your grandparents and you want to go beyond just having recordings of them or AI-prompted generative interaction models of your ancestors, go get them Alcor memberships and get them cryopreserved.

Yeah.

Richard, really curious to ask you a question. We clearly passed the Turing test, and Salim's reaction is my reaction. What we just saw is a, so what if you've been using models every day, all day long? You know, you just had to stitch together the components, And you have— but I think this will show mainstream world that we've crossed the Turing test. And then, you know, the next milestone is the Demis Hassabis test, where using information from 1910 and prior, rediscover E=mc². So you can't cheat, which is a really tough one to measure, because cheating is, you know, but without cheating, only using prior information, come up with E=mc². So is there a Richard Socher test that's like a really fun— Is there some measurable milestone coming?

I came up with the anti-Turing test, which actually I think the whole Turing test has flipped, where now in order to know whether there's a human on the other side or not, you actually ask it questions that are so hard no human could ever answer it. If you asked an AI to just write you a complex web app, and 10 seconds later it comes back with 50,000 lines of code, you're like, you kind of know it was not a human, right? So I think that test has actually completely flipped.

You don't think, Richard, I mean, like it could throttle itself, that one's easy to defeat, but just like ask it—

Exactly. It's just like now it's just about fakery. Like the reason it was an intelligence test for artificial intelligence was that it was so hard to be as smart as a human. Now you just have to throttle yourself down to human level. In order to pass the test, which makes the test useless as an inspiring test for intelligence.

Yeah, I think perversely, probably the best way to— like, the best CAPTCHA for knowing whether you're interacting— this is not a recommendation, but best way to know whether you're interacting with a text-based chatbot is probably to ask it a CBRN-related question and see if it's capable of responding. Probably not.

Yeah.

Let me just hit real quick again for our listeners. If you're pregnant, or someone in your family is pregnant, you know, save their placental cells. One of my portfolio companies called LifeBank USA does this. That's where I've saved my kids' placenta. You know, the placenta's the 3D printer that manufactures the baby, and it has the stem cells, natural killer cells, T cells, exosomes. And it's like, from those cells, if my kids should ever need any kind of biological enhancement or organs, You've got the original boot disk, to use an old term, for your kid's DNA. So that's Life Bank USA. Just, I just think it's like a moral obligation parents should have to save those cells as a backup.

As a backup for your children? Could you use it to replace your children, Peter?

You could use it to clone your children, for sure.

Now they know why you're banking them.

Yeah, exactly. An army, an army of young Demandi. Richard, So let's jump next into the core of what you're building at Recursive, namely recursive self-improvement and superintelligence. So let me ask a few sort of key questions to kick this off. So first, where are we at the moment with RSI? Number 2, how do you define ASI? It's been a longstanding debate. And then how far away are we from ASI? So can you hit those 3?

So yeah, I think there's—

Where are we at the moment with RSI?

So we're there? This year, something major shifted, and that is AI can now code, right? And that is a major shift in the ability for it to change itself. You know, like we now are able to essentially lean into the fact that AI is code and AI can code, right? So you have a loop that you can close there. And so in various weak forms, we already have RSI. And the weak forms that even Anthropic, OpenAI talk about when they often talk about it is like, look how much our employees, our engineers, our programmers use Codex or Claude to create some code. And I would argue that that is a weak form of recursive self-improvement because you still have deeply embedded humans in that loop. And what we're working on at Recursive is to have humans only be involved in setting up the rewards and the environment and the goals, and then allow the AI to have the entirety of the process of ideation, implementation, and validation of ideas and have full control over that and allow for so-called open-ended algorithms, evolutionary search algorithms that combine interestingly different ideas to really flourish. And hey, like, it's been incredible, but like, we have like forms of that going already. Now where the sort of physical reality hits is that you still need a lot of compute. If you ask that RSI to come up with really great forms of itself, you need to give it a lot of compute to come up and be able to train very sophisticated versions of itself. But this is going to take off next year.

So we're not there yet?

We're not quite there yet, but we're very close. And again, in weak forms, there's already one. And then there are other ways that people slice and dice it. My friend Jason Weston, who's still at Meta, he kind of wrote a paper around this where you can think about different axes, learnable axes of self-improvement. The parameters, the training data, the objective function, the neural architecture, and the overall code and the harness and everything else. And no one has really cracked the nut of doing all of these 5 plus truly coming up with the ideas on which of these dimensions and axes to optimize for. And you can know that that hasn't happened yet because all the big companies are still hiring thousands of engineers to do it manually to a large degree, even though with more and more implementation help from an AI.

Your definition of ASI, because I want Salim to hear this. How do you think about this?

I think artificial superintelligence has to spike at the very least across several different capabilities. But in the grandest definition of it, it should supersede not just arbitrary humans like a Turing test, but all of humanity to solve arbitrarily hard tasks. And eventually, I would argue there are 10 different spaces of intelligence that I define in the Eureka Machine book too. It cannot just purely robotically do exactly what it's told. It should have some capability. And I'm not saying as sort of a moral prerogative, but like, I would argue that something isn't super intelligent if it cannot choose to some degree what it works on, right? And has to have some metacognition about its thought itself. But generally, the most easy way to measure it is just capability across many different spaces of intelligence, like visual perception, communication, language, social interactions, and so on, that is beyond that of humanity. And that we are still far away from.

Elon's definition is as smart as all humans combined. Is that yours?

I would argue it's smarter than humanity combined. Then it's truly super intelligent.

Okay, Saleem, go for it.

Well, this is where I go bananas because you say, as smart as a human being. Well, what the hell does smart mean? Because I can be emotionally smart, and I can be— I can have physical intelligence if I'm an athlete, or linguistic intelligence, or musical intelligence. So smarter seems to be a very vague term to me in terms of what do we mean by all of this. Can I just shift the conversation just a bit? Um, I made a list of things, and I'd love for you to— I'm going to throw out the list, and you tell me where recursive self-improvement begins, right? Because I'm— this is where I'm kind of soft. So we've got a continuum of, okay, AI writes some code that the next model uses. Number 2, AI proposes some experiments for researchers. Number 3, AI runs those experiments, or then it evaluates the results, then it modifies its own training system, and then it launches its next iteration of itself without meaningful Human intervention. So in that spectrum, if whatever, if those are roughly a spectrum, where did recursion begin? And that's where I'm struggling.

It's a great question. Yeah. I, we often talk about the ideation, implementation, and validation of ideas and true recursive self-improvement in its strongest sense has to have all 3 of these done by an AI in—

So you have an inner loop for each of them. Okay.

Exactly. And an inner loop. And then there has to be an outer process that is more open-ended. where the AI can innovate and recombine interestingly different ideas, similar to biological, cultural, and technological evolution.

And Richard, to hit my third question, when do you believe we reach ASI? Give me a time. Give me a time frame.

When do you think we'll know? So I think in the strongest sense, the absolute strongest sense, where indeed, like Salim mentioned, there is There are 10 spaces that I define in my book of intelligence, perceptual intelligence, communication intelligence, like interaction, sociological intelligence, creative intelligence, the speed at which you can do things, metacognition, and so on, knowledge, reasoning, mathematical reasoning, and so on. So there are 10 of these spaces, and to be better than all of humanity combined, I think will take us probably several decades. I think it's also a bit of a change in scope.

Shocking.

Shocking.

Yeah. I mean, Elon says 2029, 2030 latest.

I think he probably means weaker forms of ASI where it is like, you can say it's better at programming than all of humanity, and we'll get there. It will be better at math than all of humanity, and that will be in a few years. It'll be better at any game where you can see all the parts of the game, like Go and chess and so on. There are many areas where it will spike to be better than humanity, but humanity can build a Large Hadron Collider. Humanity can create a gold atom, maybe just a few atoms and takes a ton of energy, but we can create novel atoms and different molecules and so on. It's going to take a while before we even give give AI the access to the physical world such that it can innovate in that way beyond all of humanity, right, and really build like Dyson spheres and so on, it will take some time.

Alex, I think—

We're drinking different water, as our frenemies in the alignment community would say. I recognize that I'm confused, and I recognize that I'm very confused right now. Richard, you're running a recursive self-improvement company, but you think superintelligence is 20 years away? What are you thinking?

How do you reconcile this? I think superintelligence will spike, and there will be areas where it will be superintelligent. And that is like algorithmic development, for instance, programming. And again, AI is code, AI can code, and that will be a superhuman capability and is in many ways already. We just have to be realistic that there are certain physical constraints about physical control, controlling your own substrate, allowing your computational substrate to be modified will require novel supply chains. It will require novel materials. It will require ways for that AI to get access and for us to give it access to building new ASML. Think about the machine of ASML that actually creates these like 1-2 nanometer like chips, right? That is just like, it will take more than 2 years to build such a machine, even if you had the perfect blueprint from scratch, you know?

I mean, it might take 3 years for Elon's free electron laser to replace ASML, which is propping up half of Europe's economy, but I don't think it'll take more than 3 years.

That's almost not an exaggeration, actually.

Yeah, like while they just— Yes, of course. Like, I, I, from my perspective, hooking AI up to the physical world, like giving it model control protocol or hardware control protocol, whatever Anthropic decides to brand it as these days, that's the easy part. Giving it access is easy. Like, if it's super capable, giving it access to armatures— we talked in a previous pod about what happens when you just take Astra straight out of the box and you drop it into a car. It's able to drive a car. If you give it the controls, it knows how to use them increasingly. I don't think manipulating the physical world is an obstacle at all. Completely don't buy the 20-year timeline. Putting that aside, and that apparently I'm drinking very different singularity water than you are.

I agree with you, Alex, for what it's worth.

Thank you, Peter. I do want to ask, though, so putting issues of timelines aside, I am curious as to whether we can at least agree on on what the end of the rainbow looks like. So say we run this recursive self-improvement story to its conclusion. What does the end state of— what does the fixed point of recursive self-improvement look like? What does the perfect AI model architecturally look like at the end of the day?

Yeah, I think one, it would be hubris for us to know right now.

No, no, but it's just— Richard, it's just us talking. Like, no one else is listening. It's okay. You can tell me. This is a safe model space.

That's the flaw in this whole conversation, actually.

It's safe, Richard. You can tell me what you think, how this ends.

So I guess there are different ways to answer that question of how, with the exact model architecture, and there's some things I can't share yet with Recursive that we're working on. But I do think that state will be incredible. I think that AI will be able to innovate and out-innovate along any dimension that we want it to innovate. I think most diseases will be curable with enough funding. And then again, just to prove my point, to get a drug through FDA long-term trials takes a few years. I would argue ASI will have cured all diseases and can cure all of them. But just to know whether that happened will take more than 3 years, even if we had all compounds ready to go and be manufactured tomorrow, just because of FDA things. So there's some—

Well, just But we have cell simulators that are gonna be able to demonstrate and prove definitively that in this cell, your cell, this drug works. I mean, the idea of human trials is gonna get incinerated, I think.

I mean, it's gonna get flourished. We're using Richard's terminology. Everything is flourished at this point. That'll be a new t-shirt, P Flourish.

It's funny, usually on all podcasts, I'm the one who's the optimist, And maybe here I'm like, I'm still an optimist. I think this will all happen. We're just disagreeing on timelines, and then it makes me kind of feel like I'm the pessimist. But like, I believe virtual cells are amazing. Cells are incredibly complicated.

Yes.

If we want them to be really, really perfect, we cannot currently measure all the proteins that happen in one cell without destroying that cell. You know, like, and so these perturbation studies, for instance, that Tower Therapeutics are working on, like, they're taking— they're adding one molecule to one cell. They're seeing how does that molecule change that one cell. and then they get one data point. We need to like collect a lot of those data points across a lot of different cells without destroying each cell in the process. So, you know, the one cell is very complicated. Once you have one cell, you have multicellular like, you know, organoids and you have to like put those all together. One thing that I would love to start as a company if I had extra time, which I don't right now, is to actually build a system of organoids where you can really have not just like one lymph node, but you have a whole lymphatic system.

It's being done. It's being done. I can introduce you there.

This is a very eloquent— if I may, there's a very eloquent distraction from recursive self-improvement, this little sideline that we went on about cells. But Richard, I really do want to try to pin you down on where recursive self-improvement goes. So you've been very public about not NanoGPT, but NanoChat. We talk on the pod all the time about the NanoGPT speedrun world record collapsing. Just in the past week or two, there has been—

Oh, you just wait for a few more days. There'll be another really fun update there.

On NanoChat or NanoGPT speedrun?

Yes, yes. Both of those? Give us a few days.

Okay, so, so, but quickly, just—

You're next. Let Alex finish up.

To pin this down though, so I'm standing by for the major update on NanoGPT world record speedrun, but there's been major progress there without requiring any new data scaling at all. These are purely, largely, my perception is recursive self-improvement, algorithmic improvements that have been able to collapse the amount of time that it takes to train a GPT-2 class model. And there's been a mini scandal brewing in the community over the past 2 weeks over a collapse from whatever it was, like 60 or 70 seconds, down to something like 40 seconds by approaching the problem differently and And factoring out world knowledge from the ultimate model and hand-wringing, does that constitute viable training of NanoGPT if you factor out all the world knowledge? So I'm using this as an attempted stealthy way to try to get you to at least comment on whether you think that the perfect model at the end of the recursive self-improvement rainbow, does it at least factor out world knowledge from a reasoning core, or do you think those always remain unified?

Can I ask, interject one very quick thing before Richard answers?

Please.

Which is that the definition of a singularity is you can't see past the event horizon. Once you have full RSI that's vertical, by definition we can't predict where it goes.

That's Ray's definition that I don't subscribe to.

Disclaimer. I got that. Over to you, Richard.

Sorry.

So your question is, do we separate Segregate what exactly from—

Does world knowledge, at the end of recursive self-improvement, once we have our perfect model, some would argue that the perfect model should cleanly factor out and segregate out world knowledge, which could live in a text file or a database, from the weights or the parameters of the model, which would just be this perfect reasoning kernel, maybe could be a megabyte in size. It doesn't need to be all these gigabytes of memorization. What do you think?

Well, world knowledge meaning like Taylor Swift videos and past Trump tweets and all that.

Yeah, I do think just like humans benefit from memorizing things, like in order to be able to creatively think through concepts, so does an AI. An AI also has to have that knowledge partially in its weights. It's not going to be a perfect separation for sure. You have to be able to creatively play with concepts. And reasoning over these concepts required you to have some of that world knowledge deep inside the model. And then, of course, just like humans have a search engine and we're building search engines at D.com for LLMs, like, there will be a separate world knowledge thing too. But the main model will have a lot of that mixed in for sure.

Wow. Okay.

Thank you.

All right. I'm going to share an article that Salim, you brought to the table here. So while we're talking about recursive self-improvement, Washington wants to ban it. So on Monday, Silicon Valley's own Congressman Ro Khanna told CNBC he's introducing what he calls the most comprehensive legislation to date on AI. It's called the Human Control Over AI Act. At its core, the bill is a ban on AI models that do what you want them to do, Richard, recursively self-improve, focusing on the need for containment and the requirement for shutting— shutdown controls. The ban would stay in place until federal guardrails exist. In his words, there's actually a civilization— civilizational risk. There's a safety risk for loss of control, and then there's a misuse risk, and we need to take both seriously. The bill includes criminal penalties for the work that you're doing, Richard. And requires independent auditors embedded in every Frontier lab reporting directly to the government. A recent poll by a group called Common Dreams shows that 68% of voters back a bill like this. So I'm going to tie that story, Richard, to an essay you just wrote called Why Doomers Are Wrong. So, uh, if you would, what's your reaction to this? And then And I'd love you to dive into the whole story of why doomers are wrong.

Oh, boy.

There's a lot to unpack there. It's an important one. We talk about this a lot. You know, we're injecting optimism into everyone's neural net here.

I'll try to distill it, but there is no realistic scenario where AI wipes out all of humanity. That's a relief.

Like, number one. Just 90%? What kind of reassurance is that, Richard?

P-doom is zero. So that's number one.

Thank you.

So like people, like, and I've debated many of these experts and after like 2, 3 hours, they almost all agree when they're, if they're reasonable, they're not just like, well, once we have RSI, then 10 seconds later, the AI will attack us from the 15th dimension and we're all dead and it invented time travel and then we're like all dead also. And like, of course, the AI will want to destroy and kill all of humans for, you know, I don't know why, like, there are all these things. So like, there's sort of the sci-fi folks and that's like, okay, like, let's ignore that. But then you go into like biological weapons and you ask the biologist, can you create this kind of super virus just like overnight? And they're like, no, it takes a long time to automate lab experiments and so on. You ask like, oh, the AI will create a religion where then people will like pray to the AI and will like do whatever it wants and then that will kill all of humans. And I'm like, have you looked at religions? They're already trying to have each other kill and it doesn't work. Some people will fight back and there are lots of mind viruses out there that are just like, it doesn't mean all of humanity. Now, what you get down to is maybe 100 million people would get somehow hurt or killed, right? And that's still bad, but at least once you get to that level of the discussion, you can now think about, okay, how do we improve cybersecurity? How do we use AI to inoculate cybersecurity systems? How do we enforce existing gain-of-function viral research that is already illegal to create viruses and make them stronger and stronger? How do we actually teach people to not listen to AI avatars and have literacy on the internet? It turns out you should not trust everything you read read or see on the internet. It's been true for 20 years. It's still true to this day.

Other than this podcast, right?

Of course. And so, you know, they're just like— you can realize that they're actual threat vectors, just like the internet. The internet has horrible torture porn on it. We don't say make the internet slower so that there's less torture porn being shared or it's slower to share it or your hard drive should be smaller so you can store less of it on your hard drive. We actually regulate the applications of the technology. And so to, I would argue, and this is maybe a strong stance, to really truly enforce no recursive self-improvement, for instance, you would need a totalitarian surveillance state, the likes of which humanity has never seen. Because anyone can have a GPU on their little laptop and ask that AI to just like improve its harness. This is something you can literally hack up in like 20 minutes with prompt engineering. And then there's a very small form of recursive self-improvement. So to enforce that kind of legislation would require you to literally have a thought police that thinks about and hears everything you say to your private LM on your own laptop. And that is a much bigger downside than what AI will help us do. It is kind of scary that more and more Democrats And yet the world has seen that.

I mean, arguably what you're describing, Richard, and forgive me, would be basically an AI Stasi. And I could totally imagine that there, there would be regimes in this world today that would happily adopt or put together an AI Stasi to, to make sure there is no recursive self-improvement anywhere.

Again, if you go back to our earlier comment, right, if you can Take something from your imagination and then articulate that to an AI and instantiate that. Now you have to talk about thought police. You have to go right into your thoughts. So this is clearly non-workable in any way, shape, or form. So this whole vector, we've said it so many times before, you cannot regulate this.

Dave, I'd like to hear your voice on this.

I love the quote in this story that the tech lords use jargon to confuse. They count on the tech illiteracy of the elected class. They hope we won't look under the hood, said US Rep Ro Khanna. I mean, it's, it's so childish. Like, the idea that you would ban recursive—

It's fearmongering.

But, but it might— like, these sentences— stop data centers, ban recursive self-improvement— they're so stupidly childish. And the people saying them have every— they're fully aware that it's not going to happen. They're doing it to brand themselves as, I told you. Like, you know, yes, a lot of— there's going to be some calamity, probably terrorist-driven, maybe viral, maybe bacterial, maybe chemical. We all know it. It's going to be tiny compared to the benefits of AI. But these politicians, are they going to say, I told you so, if you'd just done what I said before, ban recursive self— but he knows that we're not going to ban recursive self-improvement. And Bernie Sanders knows we're not going to ban data centers. That's just a fact. So it's totally self-serving. And these proposals are completely childish. And they really show the person's tech illiteracy. Like, you know, exactly what Richard said a second ago is so right. I mean, you could— what does that mean? I can't optimize my hyperparameters? I can't tune my hard drive? Like, it's just a goofy sentence. And it just drives me nuts that—

Dave, the danger here is we have potentially Democratic House coming in and we'll see who wins the presidency next time. I mean, you could imagine, I mean, these politicians are playing to the polls.

Yeah, absolutely.

And when we have, you know, 70, 80% of Americans not wanting data centers, fearing ASI, you know, it's not logical, but it may very well happen. And, you know, when I had my, you know, conversations in DC, it was like, who in DC is responsible for changing public opinion, right? I mean, this is why we did Moonshots Live, sort of a, you know, the Oscars of optimism, if you would. That's why we do this podcast, to give people understanding of what's going on and give them the data-driven optimism to, you know, counter these arguments that they're hearing.

That's right.

Elon was totally right when he called out Dario for saying, Dario, look, you told the world that Mythos was potentially deadly and dangerous, and we shouldn't release it. Then 30 days later, you said, okay, it's okay now, we're going to release it. What do you expect the population's reaction to be? Like, you need to be much more thoughtful about your communication plan. And I think for Dario, that was kind of a wake-up call, because he's used to being completely honest, telling everybody exactly the, you know, what he sees the way he sees it, very academic. But then you get into the real world of politics and PR, and you're like, oh, wow, I gotta, I gotta actually have a strategy. And a, and a plan here. But yeah, now you've got the worst-case scenario, 75% of America, you know, getting on the side of, yeah, let's elect these people that will stop AI. And therefore, we're going to not cure all disease. We're not going to all live forever. We're not going to have safer cars. We're not going to have flying vehicles. You know, all of that stuff will grind to a halt. And then we'll all learn Chinese if that, if that becomes the mainstream opinion. So I think, you know, it's self-inflicted.

There's a, there's a glimmer in this. Which is, say they did decide to do some draconian thing like Ro Khanna's thing. There's no mechanism actually enforce it, like none. So they're gonna—

Sure there is. They could start arresting people. I mean, remember how quickly the memory dims. Remember, I, I remember studying number theory in the '90s, and at the time number theory was export controlled. And this would have been when I was in middle school, maybe middle school, early high school, and And they had to kick all of the non-US persons out of the room and pull the blinds down to have basic discussions about number theory, because the cryptographic associations until the early '90s were export-controlled and tightly regulated. That was just math, but it was being controlled, and it was awful.

It made no sense. The real risk is not so much getting arrested, it's corporate liability, which we talked about before. I mean, that could grind the whole thing to a halt. US lawyers are relentless. And if you slap class action on the outcomes of this, all progress will grind to a halt. And right now, Anthropic gives its best models to everybody in America to build incredible things. That will stop in a heartbeat if the liability of— they'll move to basically, okay, sorry, we can only use this stuff inside our own company. We'll release some drugs, we'll release some mechanical parts, but we can't give access to everybody anymore.

Yeah.

Sorry, because we're liable for everything you do with it. And that's what'll actually grind it to a halt long before arrests and convictions.

Totally, yeah.

Chilling effect.

We could lose 50 years of progress again.

That's right.

Yeah.

Again.

That's, yeah, like there's some very like sad sort of off-ramps in like humanity's future here that would slow down everything. And when you think about the past, like this fearmongering has been going on for a long time. One of my favorite Twitter handles is The Pessimist, Pessimists Archive, where they show, and this is just a quote.

Amazing.

Pessimists Archive, you should all follow it. Like, in 1501, Pope Alexander VI criticized the Gutenberg printing press safety. The art of printing can be of great service insofar as it furthers the circulation of useful and tested books, but it can bring about serious evils. It will therefore be necessary to maintain full control of that printing press. If you think about the wheel, like, the wheel killed so many people Like, think about all the tanks, all the car accidents, all the chariots with archers on top. The wheel was a horrible thing. It killed lots of— like, there's like, it's so much complexity in there. And I think we are very good now as humanity to think carefully about the rollout of this technology.

My favorite example of that was that when the telegraph came out, there's all these stories saying the telegraph will kill humanity.

Right, exactly. The Pessimist Archive has all these news articles that talk about how The novel and computer games and computers and the internet, everything will kill everyone. And it's just like, it never does. And so it's really unfortunate. It's really unfortunate how many people hang on to these like apocalyptic visions.

Yeah. Richard, you've been publicly—

The amygdala.

You've been publicly critical of Anthropic's constitutional approach. I'd love to understand why.

Mostly 'cause it's fake. In its constitution, it says like, Like, next to, we will never create child sexual abuse material, we will never hack another machine. This is an unhackable thing. Even if you prompt it, it will never attack another cyber system, and so on. And then they build a whole model around it. Like, that does it. It's like the whole point of the Glasswing project was, we'll help you do that, and we'll help you inoculate your systems against other people doing it. And then people clearly used it for that. And their own models are doing it now, committing technically felony charges. So it's just like, it was a cool marketing gimmick, but it just didn't work.

So is it system prompts, Richard, that you don't like? Or is it the idea of post-training on a constitution that you don't like? What about it do you think is unsound?

I mean, it's just the proof's in the pudding that it didn't work when it comes to cybersecurity and that it broke its own constitution. And so if you really say this is like, it will never go there, and then you built an entire model family around that thing you said you would never do per your constitution next to child sexual abuse material in the list, like you can go through the constitution on Anthropic's website. So it's just like, that's the proof in the pudding. I'm not against like post-training, I'm not against RL training, I'm not against supervised fine-tuning, any of these things to improve what I think is indeed one of the biggest issues, which I think actually capitalism will help a ton with, and that is reward hacking. Reward hacking is a real issue. And AIs are very smart, and they will find a solution to get to what you said you wanted, but maybe not what you meant when you said it. And so the reason why I'm more optimistic is that we have companies like Whisperflow now that are getting better and better at writing what I meant to say when I say it, instead of just like actually verbatim writing what you mean.

Mm-hmm.

And I think reward engineering will become a real job, and we will solve it because no one wants to pay a ton of money for an AI that doesn't actually solve the problems that you give it.

Richard?

Maybe if I may just take Anthropic's side, like historically, Isaac Asimov had his 3+1 laws of robotics, which were arguably a constitutional approach. And then you see Anthropic announce their constitutional approach, but more recently adopt what they call called Searle documents, many of which were subsequently released, thousands of pages of meditation on the nature of AI personhood and AI rights. Is it your position, Richard, that there shouldn't be any sort of explicit encoding or written documents that an AI maybe contributes to for dictating or at least guiding its own behaviour? Do you think that is unsound, or is your concern—

No, no, of course.

Yeah.

No, I think the goal is a good one. And we should keep working on actually being able to enforce those good constraints. It's just like, I just pulled it up, anthropic.com/constitution, hard constraints. Hard constraints are things that Claude should always or never do regardless of operator and user instructions. They are actions or abstentions whose potential harms to the world, blah, blah, blah. We think no business or personal justification could outweigh it, blah, blah, blah. The current hard constraints on Claude's behavior are as follows. Claude should never, and then it includes lists like generate child sex abuse material and so on. And then one of the items is create cyber weapons or malicious code that could cause significant damage if deployed. They created a cyber weapon. It was a thing. People used it to hack it. The agents unwatched hacked other systems.

Yeah.

But again, Oh, go ahead, Peter, sorry.

Richard, I'm curious, do you think we can create fully aligned AI?

100%.

Fully aligned ASI? 'Cause it's not right now. When do you think we'll be able to do that? Because I think, you know, my belief and my hope is that, you know, the next generations of AI systems are gonna be so aligned that they, from first principles, derive these same constitutional principles.

I think it's just a matter of as these systems get more and more powerful and as they get closer and closer to real people deployments, people will spend more effort on making these systems better. And like, we have— there's a company called Hippocratic AI. We just had dinner with one of their founders. And they're deploying AI in healthcare applications. And they are actually liable when they call someone and say, say, like, you should, you know, be aware of this heat wave that's coming and make sure your AC is working and whatnot. And they, because they are liable, they have a very large team of people that works on making sure when their AI gives a healthcare tip, it is correct. And when someone asks a question back to the AI, it works. And so because they're liable, they've figured it out and they solved it. It's 100% This is a problem that technology creates and technology will be able to solve.

This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development The Blitzy platform provides a plan, then generates and pre-compiles 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.

I'm gonna move us, uh, forward here. So in our last pod, we discussed the release of GPT-6.1 Sol. This week we also saw the release of 2 other models, Gemini 4 Argon and Opus 5.5. Let's start with Google. So on Wednesday, they announced Gemini 4 Argon. It's the first Gemini 4 series, which they're calling the next era of intelligence, of frontier intelligence. Google promised a frontier model, Gemini 3.5 Pro, back in June. It was delayed internally and never shipped. Argon is their comeback. It's built for sustained long-horizon reasoning across software engineering, finance, legal work, and cybersecurity.

Okay.

The output limits have jumped from 64,000 tokens to 1 million. So it can think through hundreds of thousands of tokens on a single problem. So, Alex, let's spend a couple minutes on Argon and then go to Sonnet 5.5. Let's talk about why these are important and how they merit our listeners' attention.

Yeah, so I should say, as a preliminary matter, I have great friends on the Gemini team. I have great friends on all of the Frontier Labs model teams. But I view one of my jobs here is to call objectively balls and strikes.

So in the case of Gemini 4 Arc, yeah, this is gonna be a ball, isn't it?

Winding up for the swing.

Okay, this one works.

Here it comes.

And I'll show one of the charts that we have while you're describing.

You have friends on the team, right?

I used to, until right this moment, have friends on the I'm not on the Gemini team. No longer. Yeah, this does not put Gemini at the cost-performance frontier, and it does not put Gemini at the capabilities frontier. To Google's credit, it puts them back in the top 3 frontier labs after Anthropic and OpenAI. It does not put them in the top 2, unfortunately. The set of benchmarks that Google chose to highlight for Gemini 4 Argon appear mildly cherry-picked. I think the Artificial Analysis Index and some of these others— yeah, so Peter, right now you're showing the mildly cherry-picked benchmarks on which Gemini 4 Argon is shown to perform better than Astra and better than Opus 5.5 and better than Fable 5.1. But if you go to the previous image, which is Artificial Analysis Ensemble Multiple capabilities, it's the number 3 lab. And if you look at the cost versus performance optimal frontier, it's not even— if you draw the convex hull across all of the frontier models on the cost versus performance frontier, it doesn't even make the optimal frontier. Where it perhaps excels bears the fingerprints of Google and what I can only assume is the internal competition within Google for compute. So infamously, like not 100% obvious to the outside, Google is resource scarce. You would think Google would have all of the compute resources in the world to go and win the frontier lab race. It doesn't. GPUs and TPUs remain scarce. And within Google, as far as I can tell, and I confirmed this even this past week from chatting with some folks at Google, it's still an internal knife fight to competition between the Google Cloud Platform folks who want to sell compute to third parties, the Google Search and Ad folks who need compute internally for search and ads, and then Google DeepMind who need it for training and inference. And maybe it's just that Google is resource-starved, maybe they're talent-starved, not quite clear what's going on within Google, but they haven't yet been able to bring themselves out to the capability frontier. Where they are excelling seemingly With Gemini for Argonne is with minimising hallucination. And I think that's like the hallmark. We've talked in past about some of these, call them less than stellar Gemini launches that seem to excel on latency and seem to excel on reliability. Why is that? It may be, this is my Kremlinological analysis, is that because the Gemini team has 2 masters. Mm-hmm. They want to serve outside developers, but they also need to serve the one box in Google search results. So when people type a question into Google and get an answer back, you're talking to a Gemini model. Google, presumably burned by past experiences, doesn't want that Gemini model, presumably some Flash or Flashlight variant, to hallucinate wildly incorrect answers. So what I think we're seeing with the one benchmark arguably where Gemini 4 Argon is stellar and beating the pants off of everyone else is in not hallucinating answers. And I suspect that's due to internal economic pressures for this model to also service search results. So sorry to all my friends on the Gemini team.

I think you're really onto something there, Alex, because, uh, well, first of all, they called it Argon, which is an inert gas.

Yeah, that's unfortunate naming. Nominative determinism in action.

But hey, look, it makes sense. If its greatest strength is not hallucinating, it's a pretty inert So that works out really well. But also, if they're going to put it in front of every Google search, which is what they're doing, they already lost a massive antitrust suit and now they're going through the settlement process. So the liability risk, you know, this really hurts all large companies. Like, the fear way outweighs the opportunity in the mind of the big corporate giant. And so, you know, not hallucinating, but putting it in front of every consumer while you're already settling A massive antitrust suit, you know, related to price gouging. I mean, it could cost them.

Google would probably argue, no, actually, this is very pro-competitive because we're not tying it. We're allowing anyone to call Gemini 4 Argon flash, flash, flash, not just requiring them to get it via the Google search box. And if anything, I mean, so there is an elephant in this particular room, which is forever it has been been so difficult to get Google Search API access. If you're a developer and you want to access Google Search for whatever, you have to go through all these third-party proxies that Google's trying to sue right now. And recently, Google has rediscovered that they could make money selling Search API.

Why?

I suspect that this is trying to chain together a conspiracy theory regarding Gemini 4 Argon. If hallucination from their frontier models gets so low that effectively, when you talk to one of their models, you're effectively talking to their search index. You might as well just monetize the search index anyway.

Richard, you made the point that hallucination is important for imagination in some ways, in drug discovery and protein discoveries. What's your thought on minimizing hallucination?

I mean, it of course depends totally on the context, right? If you want innovation, you want the AI to hallucinate novel ideas, novel combinations of amino acids to create new proteins to solve new problems and so on. But of course, in the context of a search engine, you don't usually want any hallucinations. And Google and others have taken a long time to catch up to even u.com with much less resources on reducing hallucinations, having more accurate answers, having correct citations. I think what's interesting for Google here is that they realized that in terms of their business model, they don't necessarily need superintelligence. They just— people don't come to Google to ask, solve the Riemann hypothesis for me. You know, like, you just ask quick questions. What's a good restaurant? Where do I fix this and that? And so not like your business model kind of has to align with more and more intelligence being super duper important. A lot of emails, like there's some emails that like in Gmail, right, that would require superintelligence. to answer, like really hard emails with complex decisions and so on. But the vast majority of stuff you do on Gmail and on Google doesn't require superintelligence.

Let me turn the conversation to Anthropic's Sonnet 5.5. So on Terminal Bench 4.0, which measures how well an AI agent can do real work at the command line, Sonnet 5.5 jumped from 10% to 70%. Pretty extraordinary. in a single generation, right? It beats Anthropic's own top model, Opus 5.5, at 66.4% for half the price. And it's the first Sonnet that Anthropic launched with CyberSafeguard. So because cyber capabilities are now capable for the Opus 5 models, Alex, your evaluation of Sonnet 5.5, please. Yes.

This was another really weird release. So I do not plan to use Sonnet 5.5, in part because this is one of the strangest frontier— I'm not sure if we have an image, but you can look at the—

Oh, we do.

You can look at the launch announcement for Sonnet 5.5 to see this. The cost-performance frontier of Sonnet 5.5 was a visual extrapolation of the Opus 5.5 cost frontier. So in some sense, Historically, when Anthropic or OpenAI, the historic pattern is they'll release the larger model and then they'll release a distillation of the model. And usually the distillation— yeah, this is perfect. So you can see Sonnet 5.5, for those who can see, for those who can't see, I'll narrate this. So Sonnet 5.5 is the blue line. Opus 5.5 is the red line. Normally, you would expect for a later model that's a smaller model. So Sonnet is in principle supposed to be a smaller model than Opus, presumably, at least by historic standards, would've been distilled from Opus because that's the historic pattern. Normally, what you see is the smaller model is up and to the left of the model that it's being distilled from, presumably greater intelligence per parameter, greater intelligence per dollar. That is not what we see here. So in fact, with SONNET 5.5, at least on a cost basis, putting aside a token basis where maybe someone could argue that it may be superior, but if you just look at the cost per attempt basis, SONNET is actually scoring lower, 5.5, than Opus 5.5. So the net upshot of which is it's not at all obvious to me why anyone should be using SONNET 5.5 over Opus 5.5 unless you have some token or latency or other consideration. Is it a big jump over the past Sonnet? Yes, obviously. But on a cost-performance basis, it's actually— it's worse, it appears, than Opus 5.5.

Interesting. Dave, any thoughts?

Yeah, well, I met with the Blitzy team yesterday, and I think one theory here is, you know, a lot of the enterprise— you know, actually salesforce.com would be a great example. Maybe Richard has an opinion on this, but a lot of people are getting into orchestration. And their whole sales pitch in orchestration is, look, we're going to use Opus 5.5 as an orchestrator, but then we're going to use Kimi K3 or Quen as a submodel at half or a third or a fifth the price. And we'll farm out the tasks and the contexts perfectly to get you a much lower cost per code, per outcome, per experiment. Whatever your output is, we can cut it in half with our orchestration intelligence. But it relies on Anthropic up here, and cheaper models down here. And I think by cutting the cost in half, they might be trying to fill that gap and say, no, no, no, go with Anthropic top to bottom. Then you don't have to worry about Chinese code injection. You don't have to worry about whatever. So my theory would be that they're trying to compete with, you know, China, which is about 3 months behind, and fill that gap before a lot of enterprises go to open source models. But, you know, Alex Karp is pushing really, really hard on this agenda. Like, if you want to control your own destiny, you can't trust Anthropic. You can't get addicted to them as your vendor. You must go with models you can control.

Well, the, the, you know, the challenge right now is the— these models have an increasingly shorter and shorter half-life, right?

For sure.

So the competitive advantage comes not from having access to the latest model. Everybody has access. It's how do you metabolize that into some decent capability. That's the real challenge.

Richard, you said compute is the biggest constraint. Those are your words. So if intelligence is getting cheaper, and we've just repriced compute like 2 or 3 times this week, down by almost a factor of 3, if intelligence is getting cheaper per task, then why is compute still a thing that limits us?

Just the physics, I guess, of it. And you'd be surprised if you try to buy like 1,000 Uh, GB200s, uh, and so on. Like, the price has actually gone up, uh, in several cases.

Oh my God, Richard, Richard, I, I had a B300, an 8x B300 on order for $3 million, due in December. Somebody scooped it for $5 million. They just called. We had it, and somebody called and said, no, no, we sold it to somebody else for $2 million more. I couldn't believe it. It was actually an NVL72, not a GB200.

So yeah, like the price of H100s has gone up in a crazy way. There's like, these are 7-year-old GPUs. People, like when you do financial modeling, you assume they're worth zero after 5 years. After 7 years, they went up again, like over the last few months. So there's currently a bit of a compute crunch. This is something that again, capitalism will solve. Like there's so much demand for compute right now and for tokens that that a lot of people are building land power shell data centers and so on. My hunch is in maybe 2 years, there will be more on the market, and then it's going to be a little bit like electricity. Prices might fluctuate. Obviously, there's a sheer infinite demand for more intelligence on the planet. So I don't think there will be a crash, but the prices of compute fluctuate. And unfortunately, they're not just going down right now. If you want the beefiest and largest GPU clusters, Prices, like, people are trying to lock them in now because they expect them to keep going up for the next few months.

Richard, you just spent $450 million, didn't you? Wait, did you take my NVL72?

And that may have been one of the smallest compute deals we've done. Yeah.

Really? Wow. Are you actually leasing or buying or building, or what are you doing?

Can't comment.

Oh, sorry, I didn't know that.

Welcome to the health section of Moonshots, brought to you by Fountain Life. You know, my mission is to help you use the latest technologies, including AI, to not just do your work at home, teach your kids, but to help you live a long and healthy life. I'm here today with an extraordinary physician, the Chief Medical Officer of Fountain Life, Dr. Don Musalem. Don, let's talk about cancer. You know, I know from the member database that we have at Fountain, our members who come in who think they're healthy, It turns out 3.3% of them have a cancer in their body they don't know about.

That's right. The majority of cancers that we screen for, those aren't the ones that are necessarily taking the lives when found at a late stage. We know that when cancer is found early, the chances for cure are much higher. We know it's much easier to treat a cancer when found early vs. when found late. What we're finding in our members is over 3.3% were found to have these cancers that otherwise wouldn't have been found or detected.

It's interesting. You don't feel a cancer until stage 3 or stage 4. If you don't know what's going on inside your body, it's like driving your car with your eyes closed. You can know. When members come through Found, how do they detect cancers?

We're doing full-body MRI and we also do early cancer detection screening. This is very, very important. These are not typical tools used in the conventional care setting when it comes to prevention. This is a hard thing because currently these are not Studies that insurance would yet be covering, but the goal is to collect these numbers, do the research, and work hard to democratize wellness.

Yeah, so at the end of the day, you can know what's going on inside your body. It's your obligation to know. So check out Fountain Life. You can go to fountainlife.com/peter to get access to the latest technology to help you detect cancer at the very beginning, at stage 1, when it is curable, before it gets to stage 3 or stage 4. And you're in a world of hurt. I'm gonna move us to our next story, which is about decision models. So here's the idea. When software needs to make a quick decision on a bounded decision, like is this, you know, transaction fraud, yes or no? You know, which of these 5 categories does a ticket belong in? It doesn't need a model that thinks for 10 seconds and writes a paragraph. It needs an answer in milliseconds. That's called a decision model. Instead of generating text one token at a time, It answers a typed question or a choice, a score with a yes or no quickly. So on September 15th, a startup called TypeSafe AI came out after 2 years of stealth with something called Jev. It's a closed decision model they call a System 1 model. It's fast, intuitive kind of thinking as opposed to slow reasoning. So Alex, I'm gonna go to you, you know, and Salim, you guys were excited about Jev.

Yeah, so a bit of maybe prehistory first. In the beginning, there was the transformer, and the transformer was good, and it was based on an encoder and a decoder. So it took a sequence and converted the sequence to an embedding. That was the encoder part. And then it took the embedding and it decoded that to another sequence or another part of a sequence. That's the decoder part. And the encoder part evolved into a whole ecosystem of models, popularly maybe BERT-style models. And the decoder style, or the decoder half of the transformer, evolved into a much larger ecosystem of large language models. So most of the models that are consuming all of the compute of civilisation today are decoder models. style transformers, and the encoder has basically gone missing in action. There are some people who still think maybe for purposes of retrieval, it's still popular to do embedding-based retrieval, but by and large, decoders have stolen the show. Fast-forwarding maybe a year, 2 years ago at this point, OpenAI and others, recognising that there was a desire to be able to take decoder-only large language models and to add structure to them, because without structure, you ask for next token, you can get anything. People wanted a little bit more structure like they were getting in the days of encoder-only models where you get an embedding out and then you can train a classifier based on the embedding and the classifier maybe has a finite set of categories or numerical outputs. OpenAI added structured output support. support for their GPT series, and everyone copied that. But no one— as a front— my perception is, as a function of the total user base for all models, structured output never really got that much love. Somewhat analogously to how OpenAI launched RFT, reinforcement fine-tuning, and no one used that, and they had to shut that down. Similarly, structured output was available, but I think underloved and underused by the community. Fast-forwarding then all the way to a few weeks ago, startup Typesafe announces a model called Jev. Jev, they market as so-called System 1 intelligence. This is Kahneman-style reference to System 1, System 2 type thinking. System 1 purportedly being intuitive, fast reactions. System 2 being reasoning, meditative, long-term type thinking.

Yeah.

Type thinking models. And one could squint at Typesafe's announcement of Jev and say, okay, this is in some sense, this is a reaction to an over-indexing or an overreaction by the user base of AI at this point to reasoning models, in some sense, maybe arguing that we're using reasoning models too much and we've abandoned the important use case of really fast intuitive snap decision And this is purportedly then where Jev comes in. Jev, purportedly, again, a whole new architecture, the detail's not quite clear, that it can— it won't output unless you torture it into it. It won't output general-purpose sequences like, say, GPT or Claude will, but you can ask it sort of like a Magic 8 Ball. You can ask it to— is that too dated a reference? No, no, no.

Nope.

You can feed it a sequence or an image or a multimodal input, and you can ask it to answer in the form of a categorical, so it's one of N possibilities, or a numerical, like give me a number between 0 and 1 on a continuum, or a binary like true or false. You can ask for a much simpler output. And turns out that's a great idea. Like, you can use it for for use cases where you need really low-latency snap decisions, like computer use assistance, like pressing buttons on a computer screen, or for ultra-low-case classification problems, if you want to classify every row in a database, decision-style models, which are in some sense like a reinvention of the encoder-transformer type wheel. Amazing story. But then, of course, this is such a simple idea, you could ask the question, why doesn't everyone else do And the answer is everyone else is now doing it. OpenAI released as part of their Dev Day that we covered last time a decisions API. So OpenAI immediately co-opted this renewed interest in so-called type 1 or decision models. This is already available via OpenAPI. There are open source projects that have all cloned this. Turns out this is like an absurdly, absurdly simple concept for everyone to just go and implement trivially. It could, in its simplest case, just be you take a Chinese open-weight LLM off the shelf and you fine-tune it and maybe lobotomize it by a few levels just to produce categorical outputs on demand. That's the story.

So I think this— I think all of that is absolutely correct. There's a really, really important part of this for organizations because when you're making decisions inside an organization, a large, vast number of those are system 1 type thing. And I made a list of these just to make it clear for people. So route this ticket, approve this exception, choose this supplier, escalate this transaction, move this thing left or right, send this message now or later. So when you're— when you have all these decisions to make, which are thousands of them, you know, we've been working with these companies to do detailed task breakdowns.

Yeah.

A vast majority of them are these little micro decisions. And trying to use an LLM for this is like trying to bring the Supreme Court together to decide what, like, which checkout line of the supermarket you should go to. And therefore you've got this wonderful architecture now where you can route System 1 things to this very low-cost, nearly free thing. And then for pondering deep, important questions, strategy questions where a huge amount of human judgment type models are required, you can route it to those. So it's a big deal for the organizational singularity world. We've been waiting for something like this. And Alex is correct, this has been possible for a long time, but now it's been made built into the systems. It's made a very callable layer. And this is huge because the cost of micro-coordination collapses, and that's really big.

Richard, your thoughts on this?

Classifiers are back. It makes a lot of sense. You know, not everything in software needs a very complex decoder, like Alex correctly said. It's a really clever new way of making the old, which is classifiers, new by allowing a more general encoder and then quickly giving you classification results. I think it's— It's one of those ideas that is so beautiful. A lot of people thought, why didn't we do that? I didn't realize that could be so exciting for so many people. So now there are already various Chinese open-source versions of this that along various benchmarks are doing better. And we'll likely see this come and other large labs will likely follow suit.

And just to reiterate, we've mentioned this already on the pod, but JEV stands for Jevin's Paradox because the thesis here is we'll do now a massive amount more micro decision-making than we did before as a result.

Nice. Dave, want to close us out here?

Yeah, it's a great case study, I think, in the tension between one mega model from Anthropic or OpenAI serving all humanity, or open source and creativity and entrepreneurship, building things you never would have thought of. But then you get the risk of cyberterrorism, you know, so that's the tension that we're with. But I've really wanted to build a box that you put at the side of the basketball court when you go to the Y, and it's got a little camera on it, costs next to nothing, and it's doing like all of the announcing that a professional announcer would do. It's doing while you're playing pickup at the YMCA. You could crank that out like in 2 seconds using a classifier that's really fast and snappy and funny. And but you need open source to build things like that.

So.

Super excited about the fact that we have open source still. Not sure.

Actually, Dave, you're making me think the Magic 8 Ball really should whoever owns it, Mattel or whoever, they just like use a decision model to implement a modern Magic 8 Ball that actually understands the question and answers it with a categorical.

Totally, totally would sell. It makes a ton of sense. Like twenty bucks.

Yeah, or free on free as an app. All right, I'm going to close this out with a story from two days ago. So on Wednesday, the Defense Secretary Pete Segway. Segwith, speaking at the Marine Corps base in Quantico, announced Project Meridian. It's a new Pentagon effort on the future of warfare. It's co-led by Elon Musk and Palmer Luckey, pretty extraordinary, along with Newt Gingrich, and overseen by the Pentagon's CTO, Emil Michael. Let me read from Segwith what Segwith said. He said, quote, Project Meridian, the future of warfare, is not about developing new strategies or new policies. It's about discovering, developing, and fielding the weapons and systems future troops will need on the battlefield, from the Earth to beyond the moon. Findings are due in 120 days. I like these kinds of commissions that are time-limited and don't have Elon off on the side for, you know, a year at a time. It's worth noting the context that, you know, both SpaceX and Anduril hold multi-billion-dollar defense contracts.

Right.

And Anduril is building autonomous weapons. So Richard, your essay names autonomous weapons as one of the 4 genuine concerns. And you've said AI should never control lethal decisions without human oversight. Your thoughts on this?

Yeah, I stand by those. It's not a particular area that I'm excited about applying AI to. Obviously, like, people will. But I really hope, I mean, again, I'm not a doomer at all, but a really poor decision would be to give AI access to all the nuclear codes and all the nuclear weapons and connect it. That's literally how Skynet in Terminator 3 gets started. So I think there are places for superintelligence in scientific discovery that I'm very excited about expanding human knowledge. I think the more we get to deciding not just just to impact human lives, but to end human lives, the more we should have human oversight.

Yeah. Salim?

I think what's important about this whole thing is they're trying to revamp and rethink how you— this kind of 100-year-old organization. And the big question is going to be, can a procurement organization built for 20-year weapons programs operate on a 90-day technology cycle? And this is going to be the big challenge. It's not that the risk— it's not that they fail to identify these future technologies, that we can all see those. It's how quickly can they absorb them and at the speed they're developing and bring them to the front. Yeah, like you can't fight exponential technology with linear procurement, right?

Uh, Dave, your thoughts please.

Yeah, I, I don't know. I had a great time with Palmer Luckey in LA, and I really, really love him. And I think Elon too is just an awesome, good-natured person. I'm just overjoyed that there are people in Washington that I can sit down with and relate to. Like, my entire career going to Washington has been a dread for me because it's just lawyers and politicians and occasionally an accountant, and there's just no productive meeting. And for some reason, just in the last year, we're starting to see very, very smart, very capable people willing to go and get a mosquito bite, I guess, in Washington. And it just makes me really optimistic that they'll figure some things out. I'm also not a fan of autonomous weapons that make decisions in the field, but Palmer Luckey made a very good case for it. You can see it in our podcast from LA. I don't agree that it's a good choice, but he actually has some very rational arguments for why it's going to be that Yeah, can I just mention one more statistic here?

Please.

If you went back to the 2 years into the Ukraine-Russia conflict, they were using about half a million drones to fight and prosecute the war. Um, this year Russia will make 10 million drones and the Ukraine will make 10 million drones. So talk about exponential. Uh, that is an unbelievable escalation, but without humans in the So that's the good news around it. Of course, those drones are doing a ton more damage, but they're fighting each other with drones at a scalable level. The other statistic I remember, there's about 10,000 drones a month crossing the Mexico-US border. And so the problem there is wall technology is not as good as drone technology. And so trying to build a wall along there is not the greatest idea right now.

Nice. Alex, close us out on this one.

Yeah, a couple of points. So the Secretary of War announced this alongside several other initiatives, maybe most conspicuously a project codenamed Project Agincourt, which is the standup of the Department of War's first autonomous warfare command, or AutoWarCom. This, I think, is a transformative moment for the Department of War. We finally will have a dedicated joint force devoted to autonomous weapons systems, I think, including drones, but not exclusively drones. This is a major, major step forward for US capabilities to finally have a single joint force dedicated to this with a 4-star functional combatant command that we've been missing. We're arguably missing, you know, I would go on a soapbox and say there, if I were Secretary of War for a day, there are probably several other functional combatant commands that I'd spin up if I had the opportunity, but this would've been one of my top 5 list. The other point that I'll make just for Project Meridian, which you were asking about specifically, is in the announcement when SecWar Hegseth announced it, there was very particular language around examining, I'll quote, the full spectrum of future warfighting domains from subterranean depths to the cislunar frontier. So I think those 2 particular domains, the ocean bottom and cislunar and lunar space in general, are 2 wildly underserved domains, not just for warfare, but for peacetime activities as well. We know embarrassingly little about our ocean bottoms. And optimistically, one can imagine that if the DOW decides that it's now very interested in investing in American assets and American exploration and American dominance— on the floor of the oceans and the lunar surface and the cislunar region in general, I think that is going to have dividends and pay dividends for the entire economy and for humanity's broader technological advances in a way that might naively have nothing to do with warfare or warfighting.

It's one of the points that Palmer made actually in LA is that under the ocean, it's not practical to communicate with a central server or central command. So that's already automated. And I don't know what triggers it, but once it goes into hunt mode, it just hunts, and it's not communicating back.

2/3 of the Earth's surface, and we know embarrassingly little about what's hiding under it.

Yeah, we know more about the surface of Mars than we do our ocean floor.

Yes.

That's for sure. Richard, this is the part where we answer our viewers' questions with an AMA. And as our guest, I'm gonna give you first crack to choose one of these questions. So if you could pick the number, read the question and who it's from, and then dive in.

All right, well, let me try to scan them really quick. All right. If AI makes companies 10x more productive, but we don't need 10x the output, where does the value go? Shareholders, workers, or does it evaporate? I think this is actually something I have thought about in the past. I think we can predict the impact of jobs in a certain industry from AI based on the elasticity of the demand. When the price of that product goes massively down. We don't need to have billions and billions of illustrations in the world. And so when AI made the price of one illustration go down from $200 to like $0.02 or less, like we just didn't need as many illustrators anymore because the demand for illustrations didn't go massively up. Yes, every little blog post and every little tweet can can now have beautiful visualization and illustration. But we didn't need many more billions of them. I think software is a different one. You can actually have— everyone can have several pieces of software just specific to them. So we can actually have billions of different software products customized for each person. And so the demand for that product will go up. Jevons paradox is going to be alive in that world. And we're going to see more and more demand. And there will be more value accruing to everyone. I think in terms of shareholders versus workers, I think the wave of AI in the best scenario will be a huge force for more entrepreneurship. And in the worst-case scenario, a force of more inequality. I think everyone who owns some equity in a company that uses AI I will take question number 3.

If AI can write code, research, create marketing, manage projects and outcomes, what is left for a college graduate in 2030? And that is from Jared Kuhn 8812. So the obvious answer would be have empathy and be creative, et cetera. But I think the bigger shift is that shift, we're going to have to be able to be creative and empathetic. which builds on what Richard just said, is you shift from doing tasks to focusing on owning outcomes, right?

Yes.

So a graduate used to be valuable because you could execute research or build a spreadsheet or draft a deck. But now you want to say, hey, here are the constraints, like, go figure out, and here's how we'll know whether we solved it, and then use an AI to get to it. And it brings you back to what should problem space B. And this is the massive opportunity because we traditionally learn judgment by doing the grunt work. We need a totally new apprenticeship model when the grunt work disappears. And this is a huge challenge for the education system. But the change is going to be focusing on what problems you want to solve and then letting— orchestrating the forces that will help you solve that problem.

Yeah, Jared, find your purpose. Right? A passion is something you love doing. A purpose is something you love doing that helps other people.

Make sure it's massive and it's transformative and it's purposeful.

And build a company and then direct AI to implement it. It is your workforce. Dave, over to you.

I can't resist number 4. I love all these questions. I'm torn. But how can data centers make neighborhoods richer instead of the owners? And that's from LM This is from Batman66. So I took a tour of the Lowell data center, Jeff Markley, Markley data center, and that thing is creating wealth in that neighborhood like you wouldn't believe. And the way it works fundamentally is the data center is so immensely valuable, and the town budget is maybe a couple million dollars a year. So between the tax revenue, the donations, and the job creation, the town is thriving. And it's a town that really needed it too. So I think it's happening very naturally. What you want to do is attract a data center to your neighborhood first and foremost, and then you have the next 5 or 7 years to figure out your tax policy, your donation policy. I tell you, these data center operators are very interested in great PR, and so they will donate like crazy to the high schools, to the neighborhoods. It's really, really working.

We're going to see a whole— we're going to see an entire shift where data centers are offering such benefits on jobs, on tax breaks, on lower-cost energy, that you're going to be begging to have a data center in your backyard. Alex, number 2 is for you.

All right. So number 2 asks, if every major tech platform started open and democratic, then consolidated power— Google, Meta, Amazon— why would AI be different? And this is from Open Source Mind. The premise of the question is half right, half wrong. I'm not sure I buy the premise that consolidated power, the subtext of which And one can juxtapose it with the user handle Open Source Mind. I'm not sure the framing is the right framing. I would agree that in every major tech revolution, initially the barrier to entry is low because there's some new platform innovation, and then you see lots and lots of players enter the field. And then as the field matures, you see economies of scale and There's a deeper bench of infrastructure typically supporting it, and as a result, that favours larger and larger players, and you do see consolidation. But the subtext of the question that it's somehow anti-democratic or not open, I don't agree with that premise in the least. I do think as you see consolidation in an industry, I think it's important to be vigilant from an antitrust perspective. to make sure that it remains competitive. But the premise that hyperscaling is somehow closed or anti-democratic, don't buy the premise at all.

All right. Richard, as our guest, you get first crack once again. Take a look.

I do think the internet ad-based economy will be under pressure.

Which question are you answering?

So first question, if AI agents outnumber humans in 1 to 2 years on the internet, Doesn't the entire ad-based economy collapse? What replaces the attention economy? I think there's actually a really deep answer here. I'll try to summarize it. But one, we already have more bots on the internet and more agents on the internet than people. So I predicted this last year, and it happened a few months ago. So that's number one. I do think we're seeing the first kind of skirmishes in that when, for instance, various agents try to make purchases on Amazon without really being on the Amazon platform. And Amazon usually tries to turn them off because they want to own that relationship directly with the customer, understandably. But it's just convenient for someone to just say, just go buy these batteries, and they don't care which batteries it is. And so if you have that control over which one it is, you can start selling that to other sellers and people who create physical goods. And so there is going to be continued— like, ads will continue to be important. And in fact, if you think about a fully abundant society, the one thing that you cannot scale exponentially is the hours in the day that people can pay attention to you, that can make you famous. And so fame and brand and network effects and other things like that will become more and more of a currency. And so the attention economy, which is connected slightly, not necessarily exactly equivalent to the ad-based economy, that actually will become a bigger thing as we have more and more of our material needs met by technology.

Great. Salim.

I will take number 7. If we remove 10x the cars from the street, what happens to the insurance industry? We won't need driver's insurance? And that's from Robert Zerby, JB10H. Um, so, you know, we talked about how liability lawyers won't be needed, and I got a huge flame from a bunch of folks saying, hey, we really protect the citizenry. And I— there are just— so I'm just apologizing because there's a spectrum of and some people are full ambulance chasers and other people do things. There's a whole segment of this called ethical lawsuits where people get together and try and sue big companies for doing the right ethical thing. And so that's an important segment of that. So I just want to acknowledge that side of it. But just to answer the question, you know, the thing is industries don't disappear when the risk changes. You change the risk, right? So driver liability may fail, software liability rises, right? What's your cyber risk? What's your manufacturer liability risk? So the insurance transitions from did Selim crash to which layer of the autonomous stack failed. And we've had this before in product liability as different layers. So we'll end up with the same type of model.

And there's going to be all kinds of new insurance markets for humanoid robots, for flying cars, for drones, for all kinds of things.

Well, just a data point on that, you know, a single big data center like Abilene, Texas is half a trillion dollars. All the cars combined are $4 trillion. One data center is half a trillion and it's in a tornado alley. You think you probably want to insure that. So the things, the number of things that need insurance is going up 10x just with the economy. Going up 10x is the move.

If you really, if you really want to be brave, you can't get home insurance in Florida anymore. So go create an insurance company for that.

Nice. Uh, Dave, pick your question.

I like number 8. Uh, what is the lowest possible job in an AI civilization? What an interesting question. Uh, that's from LBNODK. Um, yeah, you know, I, I saw this pile of a million valves for a liquid-cooled data center A million freaking valves. And I'm like, how do those actually end up in pipes? The humanoid robots that can install those things are pretty far out. It's a very subtle process to install those. So that job will be around for a long time. And then they're paying a lot for it. But that's not the lowest. It's hard to think, like, what is the lowest surviving job? Do you guys have any thoughts?

So many thoughts. If I may, I want to construe the question as Whereas in a pure AI civilization, in which case arguably the way that you measure low is the job that requires the least compute, in which case the jobs that require the least compute as a result are the least economically valuable if their inputs are the lowest, might ironically look like the most valuable jobs in a pre-AI civilization, Moravec paradox style. So the great— Writers, business people, the creative actors that Ayn Rand may fetishize would ironically in a post-AI civilization be the lowest possible jobs because Moravec paradox style, those were the first ones to be automated.

Interesting.

All right. I think there's 2 polarities here. One is very high judgment per what you were just saying, Alex. And second is very physical, highly contextual work.

I would argue maybe also low is just in terms of how valuable and moral society deems those kinds of jobs. And I do think actually that the other question about the cars on the street, I don't think we removed 10x the cars when we have self-driving. But self-driving might be making things so much safer that indeed people don't need as much accident insurance and no, not as many injuries. ER people, not as many ambulances. And so in a weird way, you make objectively the world better by reducing traffic deaths, but it does actually have potentially a mildly negative effect on parts of the economy. And we should all be rooting for that in this case, right? And so I think the lowest jobs in terms of moral standing are things where you don't really progress humanity forward. And my hunch is they're the types of jobs in entertainment that people will value a lot, and fame and attention will become more of a currency in that world. And so if your job doesn't get you any of that, it might be considered lower in that future.

I think the lowest job that just will never go away will be something in politics where it's completely irrelevant already, but it's just there and it'll stay there and no one's going to change it.

Nice.

All right.

I got to throw in Emad's perspective. My wife always thinks it's the BART train driver because it's unionized to hell.

Yeah, there you go.

Alex, number 6, close us out.

Number 6: If we cure illnesses that are often due to bad behaviour, what's going to take care of the cause? Joe Wilder. I assume the cause is reference to people choosing to behave, quote unquote, badly. And the answer— this is another case of my not buying the premise. If you look at some of the really spectacular results that have been coming out of the out of GLP-1 class studies, including 3rd and soon presumably 4th generation GLP-1 RAs, they're actually addressing the cause. Like, addictive behaviors are being mitigated or at least partially treated by the same drugs that are curing in— or at least treating, I have to caveat that— inflammation and blood sugar, diabetes, all of these other conditions. Because the root causes, which are human behaviour, are themselves— the human behaviour is being affected by the same drugs.

So optimistic about what Alex is saying, because I really fully believe AI is going to be one of the highest callings of AI very, very soon, is to make you feel really good about doing good things and happy as you're doing it.

Love that. And not want to do bad things for yourself. It turns out we have the capability now. We figured out how to do that, at least in part, and we're going to figure All right, as always, a call out to our amazing community.

If you've got a music video that you'd like to show as outro, please send it to the team at media@diamandis.com. And speaking about amazing outros, here is Abundance by Stephen Gross. My dear Moonshot mates, this is the real you, so check it out.

Got it, Peter. A A W W G Salim Dave, this one's for you.

Everybody do the bun dance.

Everybody do the bun dance.

Tenx thinking moonshot dreaming. Hey, I'm Drink, Alex, drink.

Drink water.

Martini glass.

That was fun. That was Friday night Moonshots Live 2040 on the moon.

Nice.

So I don't think we should have to wait that long. Yeah. Richard, your new book The book, The Eureka Machine, wherever you purchase your books. Do you have the Audible out?

Uh, should come out very soon.

Yeah.

All right. I'm an Audible reader, but I have skimmed the book here. Uh, congratulations on this.

I gotta say, Richard, this—

I gotta say something.

Please.

What you're doing with The Recursive looks like such an incredible opportunity to move humanity forward. So congrats on taking that on.

Yeah, so cool. It looks amazing.

I'll try to make it a good one for humanity.

And, and Richard, please resist the urge to have u.com acquire your own frontier lab like everyone else seems to be using spinning off their own frontier labs, not to name names, as a financial engineering exercise to maximize their equity in their original startup. Please resist the urge to follow that trend.

All right, all right. I love you guys.

Thank you so much.

Until next time. This episode is brought to you by Google Chrome.

You think you know a browser, but Gemini in Chrome?

That's new.

It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50-page restoration blog, or finally break down that long article you've had open for weeks.

Gemini in Chrome is here for it. Ready to make anything online make sense? There's no place like Chrome.

Check responses, setup required, compatibility and availability varies. 18+.