AI:AM — Human Tissue Models, Physical AI, and the Future of Testing · September 25, 2026

There's several levels to the AI safety thing. And the product safety question is, I think, for the existential risk people, really something that they do not care that much about at all. How drugs will affect the human body.

Yeah. So these are not what most people think of as conventional organoids, right, these round kind of clusters of cells. They are like, you know, large pieces you might, you know, biopsy out of someone's lung or liver or kidneys. And the responses that they have to all the different perturbations that we apply to them align exactly with what happens naturally in the human body. And so we see both in cases of like disease and in health, right? If you have a healthy tissue or a diseased one, and we perturb it with some drug, that the responses align exactly with what we've seen in clinical trials, for example, in people. And so this ability to just have human testing at your fingertips on all these different types of human organ tissue, we think, is the kind of critical path requirement to testing on humans at scale before actually testing in people.

How do you ensure, for example, I imagine if you have a single organoid, some human cells, I imagine flooding them with the dose appropriate for a human being as a whole is probably not what you want to do. How do you actually, like, figure out what the dosage is and whether or not, you know, that is actually what the cell would encounter in the body if it's not, you know, if it doesn't have a chemical reaction before that, before it actually encounters the cell? Like, how do you make sure that— you are actually studying exactly the conditions that the cell itself would have in the body.

Yeah, so that's a super good point. So, you know, I think for basically everything that's been done preclinically to date, right, outside of testing in animals, you are just dunking, you know, some substrate, whether it be an organoid or you have cells in a Petri dish, just basting them with this test compound that want to test. And that's exactly the problem is that's not how transport happens in the human body, right? And I think it's a perfect example. A lot of cell therapies that are designed for solid tumors in a dish will completely kill that tumor, right? In a day or two, that thing is reduced to little blebs and pieces. And then those same CAR-T cells, when they go into a patient, just flow right by the tumor through the blood vessels. They don't even know it's there. And thus is the challenge, is modeling the transport of that therapy into the tissue. Do they recognize the tumor? Do they concentrate there? And so, in Vivodyne tissues, even the blood vessels, the kind of capillary bed in that tissue is self-assembled. So, these little blood vessels grow within the tissues, and we can perfuse directly through them. And whenever we dose a drug, it is always through the blood vessels that are native to that tissue. And so the transport through the interface of the blood vessels and then the meat of the tissue, all of its parenchymal, its functional subparts, is just like you would find in a normal tissue. The density is the same. And so—

—or customization for each piece of equipment?

Yeah, yeah. So the big problem we're solving is with physical AI, most people, when you say that term, they might think of maybe humanoid drones or maybe self-driving cars and things like this here, where maybe the modalities for sensors are typically cameras or maybe a LiDAR or something like this. Maybe they have motion sensors. And typically the output control is some form of actuation of the humanoid. It's you move the joint fingers. It's maybe the drone, you steer the drone. Or if for the car, it's, you know, you're going to park the car, you're going to drive down the highway. We are taking that idea and expanding it much broader than that, really, to every single type of sensor in the world and every single type of control surface that you might want to plug that up to. So the most obvious use cases for that are today already in large industries, in manufacturing, in energy, in agriculture, in all these areas, there's already billions of sensors deployed in the world. There's already very sophisticated control systems And they've been there for decades, actually, typically driven by more classic signal processing or heuristic algorithms. If this signal is higher than this, then trigger the alarm and so forth, right? And really what those systems are missing is the intelligence layer to be able to take in all those sensors and actually be able to understand what's happening in the world, adapt as things are changing because the physical world never stays stationary, right? And then be able to map that to control surface. So One of the biggest problems we, we have to solve is evidence to the contrary.

Yeah, I don't know. I mean, the whole biologist panning this thing, it feels like missing the forest for the trees, really, again. And that's the other pattern that, you know, it's kind of hard to mistake. The other thing, too, is like on the useful front, I think we're going to see probably a reversal of the order of utility and Millennium Prize solutions in biology relative to what we've seen in math. The issue with math is that we already had a tonne of math that was interesting to mathematicians but not super useful. And so to push the frontiers of math, you kind of had to get into the not useful domain of math. In biology, we have tons of useful stuff that is still not even that well understood, right? I mean, it's just very— it's a far more empirical than theoretical domain. And I would expect a tremendous amount of utility to emerge before we have anything approaching a full mechanistic or fully causal understanding of what's going on in biology. And to some degree, obviously we'd love to have that, but we don't necessarily care, right? At the level of like, does this cancer drug work on this patient's tissue? If it works, you don't have to have a mechanism, right? That's not part of the clinical trial approval process. It's nice to have, not required. So, yeah, the other thing I would just Maybe my last comment on this bio story is this is exactly what Dario set out in Machines of Love and Grace. His core argument there around bio is it advances in kind of step function form when these new general purpose tools that make biology more, let's say, just programmable, for lack of a better term, When they are discovered, they unlock just huge new potential in experiments you can run that you couldn't run before. I think that is not yet, as far as I know, demonstrated for this new DNA editing system that they've found. Like, does it do more than other things? Does it do different things than those other things have done? I don't think that's been mapped out. But the general strategy that he outlines is We want to find these foundational technologies, and then we'll run experiments like hell on top of them to map out the space. And so this may or may not be one that really moves the needle as to what experiments can be run. I don't know if that's clear yet, but the fact that they're starting to find even good candidates for such systems, I think it's extremely telling as to where we're headed in the not too distant future?

I wonder how not too distant is not too distant, because I think they have to show something within the next 12 months. And I think biologists, bio people are very hard to convince. So it has to be something significant. I don't think an AlphaFold 2 level discovery is enough, is sufficient at this point, because AlphaFold 2— AlphaFold took maybe 5 years plus to kind of filter through and for people to start using it. And it kind of displaced a lot of chemists, actually. It displaced a lot of people who are doing on structure, because typically one protein folding structure is like one PhD, a couple of years of work. It displaced a bunch of these people. And I'm not also certain that there has been or there is clearly benefit yet from AlphaFold. Yes, it has made some things easier, but those things were not perhaps the bottlenecks to getting drugs out there. So it has made some costs cheaper, but the costs have shifted and the delays have shifted to other parts of the stack. Right. So I think bio is going to be pretty tough. And I wonder what the discovery is going to be that is going to be so interesting. So I have a list of— so I tried to build a tech tree of the next 10 years. So I have a list of like 60 different fields and what experts think are the big developments that they're going to need or they're hoping for in the next 10 years. So kind of millennium challenge for each field for 60 different fields, including pharmacology, et cetera, et cetera. And that's where I pulled up the— Right.

stuff from.

And I think some of the people that we've spoken to on the show are actually on path to kind of solve some of these. Like, the Vivodyne guy is on path to solve, I think, the organoid simulation challenge. So I wonder to what extent many of these actually do get solved in this kind of 5 to 10 year period, and perhaps even sooner than that, and we'll see the results. Which will be very exciting, I think. Very, very exciting.

Yeah, I think the other comment from Dario recently that, if nothing else, suggests that they get it is— I forget exactly how he phrased it recently, but it was like, talking about curing cancer isn't going to help. What will actually help is curing cancer. And I do think they have clarity there on how high the bar is in terms of, I mean, certainly convincing biologists, right? But again, that's not even necessarily the most relevant bar. You could have your holdouts who claim it's not impressive somehow, even as you go on saving people's lives. But it is clear that they do understand that life-saving breakthroughs are the—

It is Friday, September 25th, 9:01 AM. Nathan, good morning to you.

Good morning, Prakash. How are you today?

I am very good. We have had somewhat like not, I would say, uneventful, but no new releases, but a lot of new news. I think we'll kick off with Jensen. So Jensen Huang is going to be on the Ezra Klein Show This evening, I think, and they have started to put out the promotional clips, and the promotional clips are a doozy. Uh, we have—

Yeah, it's out, by the way. The full episode is out.

I listened to the full episode. The full episode is out, and, uh, 2-hour—

this is, uh, as somebody who can't discipline myself in my podcast length, uh, department, Ezra is usually very disciplined with a pretty tight hour. Jensen runs pretty much to 2, so that gives you a sense of, uh, just how much time he had to try to make his case that we have nothing to worry about.

Let's see if we can share the audio. Uh, no, I don't think we're going to be able to see the— have the audio. But anyway, so I, I will just summarize very quickly. Um, I'll just summarize very quickly. Jensen goes out and he says, if the companies cannot control their products even during the build phase, then we should shut them down. So simply, very kind of simplistic viewpoint. I think it shocked a lot of people, a lot of people across the AI safety space. are still trying to come to terms with exactly what he's saying. And some are saying he's being disingenuous. And Zvi comes out and says he doesn't believe in AI. He doesn't believe in superintelligence. So what do you think? I'm sure you've watched the clip. What do you think?

Yeah, I think it's an interesting artifact on several levels for sure. I do have a lot of sympathy for the crowd that says, like, what the hell is going on in this guy's brain? Because he's clearly not uninformed. He has a front row seat to everything. And it is increasingly hard for me to understand at this point how somebody can look at everything that's happened just this summer And say, oh, there's nothing to worry about here. It's just software. It's just software. The it's just software analysis is extremely strange to me. So that is weird. I have a very hard time wrapping my head around that. I have a hard time steelmanning that case. It is strange to see somebody at the centre of AI call it just software. I think that's definitely true. And I think Zvi's case that, yeah, if you're talking about superintelligence, you've got to be prepared for this thing to be unwieldy and surprising and potentially outwitting you. And we've started to see glimpses of that. So again, I have no idea how somebody looks at all the evidence that we've accumulated recently and says it's just software. At the same time, I did find myself more sympathetic to his point of view than many of the AI safety people did, in the sense that I did think he had a certain moral clarity in his stance, which was basically the shut the labs down clip is kind of not really what he's arguing for, but it is sort of him jumping to the end state and basically throwing the gauntlet down back at the leadership at OpenAI and Anthropic. So what his argument is, basically leaving the it's just software part aside for a second, he basically says, look, these guys have been putting everything they have for years into making their models capable and to making their products useful. And we really just got there in just the last however many months. Many people, of course, would date that to Opus 4 or 5, right around the turn of the end of '25 into '26. Wherever you want to date it, it clearly hasn't been that long that agents have really worked, that people have been able to get a lot of useful work out of AIs without themselves having to do a lot of work to set things up and make it useful. So I think on that point, he's right. That's a fairly, I would say, uncontroversial take on the industry. They've put everything they have, for the most part, with some budget for safety research. They've put pretty much everything they have into making the models capable and to making the products useful. And now they've achieved it. And then he says, well, now's the time to grow up, basically. You're now shipping products worldwide. You've got Millions of users, the stakes are going up. And he said, we at Nvidia, we spend something like, I forget exactly what numbers he used, but it was like, we put 10 to 20% of our effort into designing the chip. But designing the chip is just the beginning. We put 80% of our resources into validating the design, verifying, testing, simulating everything, making sure it's going to be reliable, making sure it's going to last a long time. That's part of the reason, as we talked about on an earlier episode this week, that the chips are lasting longer than their financial depreciation schedule called for, because they are doing just an absolute ton of stuff to make sure that the quality and the reliability of the product is high. And so he says to Sam and Dario, like, look, congratulations, you made it. You now have a product that really does work. Now it's time to maybe flip that ratio and be more like us instead of putting— remember, the super alignment team was promised 20% of compute secured to date 2 years ago when compute was maybe 10 to 20% of what it is today, and they couldn't even get that, right? So he's saying it's time to actually write those checks and expect to actually pay that overhead to make these products safe and reliable. And if you can't do that, or if you haven't done it yet at any given point in time, then you shouldn't ship the product. Simple as that. No big company that has all these users, that has all these people counting on them, that's trying to serve the enterprise should, nor would they find it in their long-term interests, in his argument, to ship a product that isn't safe and reliable, and that's it. Boom. I do think we should have some new rules around AI, but I did appreciate his moral clarity of saying, look, it is up to you guys. You're coming out with these sort of, we need— and I am in favor, by the way, of at a minimum some clarity on antitrust enforcement. I do think it would be great for the administration to come out and say, look, We're not going to come down on you guys if you collaborate on safety programs. But Jensen's saying, don't put it on the government. Don't put it on everybody else. It's on you. It's your product. You made it. If your ratio of capabilities advancing investments to safety and reliability advancing investments is out of whack and your products are not safe and reliable, then fix it. Simply put, don't blame anyone else. Don't ask anybody else to come in and be the hero. You don't need new laws, you don't need special exemptions. And again, I don't agree with that conclusion, but I did find the moral clarity quite refreshing because he just said, you guys are the ones doing it. And this is what I say all the time with respect to US and China. We are the ones doing it, right? It's our companies that are calling this a race, that are dead set on this RSI phenomenon. And he's just like, just don't do it. So the last kind of bit around closing the labs was basically him just saying, essentially, I think, trying to call their bluff in public and say, because then Ezra came back and said, well, these products weren't even released. These were just models in testing, right? They didn't release them. This just happened in the R&D process. And he basically said, first they got to fix their sandboxes, blah, blah, blah, blah, blah. A lot of mundane stuff. But then ultimately, I think this was kind of a, you know, in the poker game of what should be done here, this was re-raising over the top on them and saying, well, if your R&D is so dangerous that it can't be controlled, then we might need to shut you down. So basically, again, the implied message there is it's on you to fix it. And he did go on to say, and this is not as clippable, but he did go on to say, They're going to fix it. He said they're going to mature. They're in the maturation phase as companies now. They've got their hit, they've made their thing useful, demand is off the charts, but now is the time when they've got to fix their investment ratios, they've got to take ownership. And that happens throughout not only the deployed phase, but the R&D phase too. If they can't, I guess we'll have to shut them down. But his Idea is not that they will be shut down. It's just that they need to take ownership and they need to quit, you know, acting like the world is forcing them to do this when it's not.

I want to note that we had a, um, at about the same time, we have Zuck who went on another, uh, podcast and he said, look, you know, We launched Muse. We could have launched Muse a few months ago, 2 to 3 months ago. We didn't. We didn't because we did a lot of testing for security and privacy because we didn't want to put a product out there which people would be afraid to give their details to, which would mischarge people or hallucinate or whatever. And Zuck is deploying I think Muse is probably at the 100 million user phase right now. Kind of 100 million people have downloaded the app because it got to the top of the App Store. Perhaps daily user base is probably around 10 million or so, I think, probably 10 to 20% of that, and climbing quickly, right? It's been in the New York Times. It's been reviewed in the New York Times as the best AI app anyone has used. Ben Thompson is like, it's the Best AI app I've used. So when you think about in terms of the number of people using it, the number of transactions going on in Muse right now, you are probably at that level where, you know, if you even have like 0.1%, you're looking at hundreds of thousands of, you know, cases of issues where someone loses something or Muse like hallucinates something or Gets hacked or whatever. So I think what is clear, and which I have been also kind of pushing forward for now, for a while now, is that if you look at deploying widely to people, it really is a safety issue that is— it's on the onus of the company deploying the model. So I think that's quite clear in terms of product safety. On the other hand, the Eliezers of the world, they don't regard this kind of safety as AI safety. Existential risk is the risk of everyone getting wiped out. It's not the risk of having your credit card credentials hacked. It's not the risk of the AI telling someone to commit suicide. It's not even the risk of helping someone build a nuclear bomb. Most nuclear bombs cannot decimate humanity. You need the infrastructure of hundreds of thousands or thousands of missiles to achieve mutually assured destruction. So even a single nuclear bomb is not able to existentially destroy humanity. So I feel like there's several levels to the AI safety thing. And the product safety question is, I think, for the existential risk people, really something that they do not care that much about at all. But meanwhile, to consumers, it really matters a lot, right? And I think that is a big difference between, I feel like, the labs and the people inside the labs and the people outside the labs who are like, okay, it's a product, you can just deploy. And I don't really know to what extent what the labs are, what the AI safety or AI doomer people inside the labs, I don't understand what they are philosophically trying to achieve. Noah Smith has a good thing. Do you want obedience or do you want benevolence? Because you cannot have both. Obedience implies that someone can tell it to do something bad and it will do it. Benevolence implies that it will look out for your best interests, even if you tell it to do something bad to yourself. Right. So those 2 things are philosophically like kind of opposed. And I think the labs are being like very confused on what they're trying to achieve. Because in especially the American context, it's always been, I think, more individualistic and more driven by kind of like your tools should obey you. It's very different from, let's say, a Singapore context where they'll be like, look, we have intelligent, benevolent dictator and that works. That's the way that society was run since independence. So I think the— and then that goes to a little bit of the left-right split. That we're starting to see. Again, you know, people on the left believe in kind of like this benevolent, you know, overarching system that kind of guides the community. It's a, you know, that's what socialism is like, this benevolent system, right? And people on the right are like, no, right? Like that, we do not believe in that. So I, so I feel like these are philosophical differences that may arise from the fact that the labs themselves are confused on what they're trying to achieve.

Yeah, this is, I mean, I guess first of all, I would say I pity someone who downloaded Muse and said that's their best AI experience to date. Because I did download Muse and it was far from my best AI experience to date. So I think that just suggests a failure to get to the frontier on the part of the speaker. It's not nothing, but I was initially like, what do I want to use this thing for? And then I thought, well, it comes from Facebook. I happen to have an old history teacher from my time in high school who was a legend at our high school and is now running for school board. And another teacher just told me about that. And I was like, I wonder if I could get Muse to go on Facebook and identify a bunch of old classmates that I had in her class and see if we can't drum something up. And so Facebook has all this information about me, of course, right? What groups I'm in, my connections, et cetera. And so I asked Muse, can you go find groups associated with my high school and identify them and kind of help me plan this out? Now, it has somewhat limited access, which makes sense. It's not supposed to be Spamming, obviously, on your behalf. So there's some like just hard constraints, which I understand why they put in there. But then in terms of like going and finding these groups, my high school is called Chippewa Valley High School, CVHS. And I set it to go look for CVHS groups, you know, my high school, and it came back with like all sorts of CVHSs from all over the country that had nothing to do with my school. That, you know, most of them were like in California. There was like Chula Vista and whatever else. And it was your classic, like, wait a second, like, you're supposed to know this. You know, you have my biography. And also, like, I have no shared connections with people in these groups. You know, that's a red flag right there. You go to a group that you're in, you're gonna know some of the people, you're gonna be friends with them. And just like didn't do basic checks on does this seem like it would plausibly be the right answer for this person, given what we know about them? And so I was kind of like, I don't know, dude. I think both Claude and GPT would do a better job of just using the browser with me logged in, frankly, than Muse did, even given its kind of privileged access to Facebook data. So I wasn't super impressed, in all honesty. That's my initial review of Muse. Like, you have one big advantage and that is you're directly built on top of this network and you didn't really use it at all.

You had one job.

Yeah. Not super impressive. So was it— who did you say it was that— was it Ben Thompson? We gotta get Ben Thompson a tutorial. Let's see if we can package up some skills and help Ben Thompson raise his game.

So what I found, the first thing that I found was good about Muse is that Muse handled 2 Gmail accounts without issue. And this is something that I've had an issue with, you know, Codex for a long, for the longest time, because I have like so many Gmail accounts from different periods, et cetera, et cetera. And It just, it just, you know, stutters. It just cannot handle, cannot like conceive of 2 different Chrome profiles.

That's really weird. Claude has no problem with that, by the way. I have not really used GPT for that. But yeah, no issue on the Claude side.

And, and so, so it's been, it's been a hassle, like, so I haven't been able to hand over a lot of things to Codex because of that. And Muse also helped me file an FTC complaint. You know, I wrote letters. What are you complaining about? So, you know, my daughter, my daughter, like all kids, she gets attracted by things on YouTube, right? And so YouTube has like some like vidIQ or whatever, which is like a YouTube site that helps you make better videos and gives analytics, et cetera. And it has like a $0.98 subscription. So my daughter has an Apple Pay account and she clicked on the $0.98 subscription and subscribed to it. The next bill is $49.99. So they do this kind of like 7-day $0.98 subscription and then they go up to $49.99. And so my daughter doesn't have that much money in the Apple Pay account, but sure enough, they hit the account. And then the next month they kept hitting it. They kept hitting the account even after it was drained. They kept hitting the account. And I called Apple and the number that Apple gives, Apple doesn't administer its Apple Pay. It's outsourced to Green Dot Bank. Green Dot Bank is like, we can't help you because we don't know, we don't handle this. We can turn off the account, But the vendor keeping hitting it, you have to go back to the vendor. And I'm like, all right, so let's go back to the vendor. The vendor, because my daughter is using the Apple relay email, the vendor doesn't have my daughter's email address or any email address associated with the account. So then you have to go through and dig up like the relay email. And then, you know, the relay email is very shady, right? It's like Rover 3, Papaya, like blah, blah, whatever. And then finally you get to the vendor and you hit their customer service and they're like, oh, why don't you go to the website and you try to unsubscribe? And I'm like, she signed up through the app. How do I go to the website? Oh my gosh. Amazing. Right? Like, have to have— Alex telling us that he loves vidIQ. Anyway.

Plus one for vidIQ in the comments.

That's one for vidIQ. So anyway, I got very upset, right? Because like, it's such a hassle. And plus I felt like they were screwing on my daughter. So, and this whole like customer service thing, like, why don't you go do it? Yeah, I was like, Muse, why don't you do something about this? This is very annoying. So Muse filed a complaint to the FTC and complied— filed a complaint to the California Attorney General. And Muse also dug up all of the statutes. And the statutes, you know, my daughter is under the age of 12. The statutes are incredible for under the age of 12. Like, no, they can't enter into contract. So, and these people cannot, like, you know, have any of their information or they go to prison.

Basically, right?

It's like, this is like criminal offenses, right? So, uh, and then they have to assure you, like, if you cancel the account, they have to give you back all the money, and they have to, like, you know, uh, erase all the data, and they have to be upfront about it. And, you know, these guys are not even responding to emails, right? Off goes Muse, files both complaints. It doesn't press enter. I have to do the last, the last step of the way. I have to press enter, but it takes screenshots of all of the things that, you know, You know, email trail, the screenshots, et cetera, all of this stuff that you need to put together for like a legal case that you would, that eventually the, you know, an attorney general would go after. It put together and with timestamps, et cetera, and packaged everything together. And then after that, it sent a demand email to both companies, to the company. I get a response immediately from like the senior community support manager. Like detailing what they've done, what data they have, all the steps they've taken, like all the charges they removed, everything, like a day. So this is the kind of thing that I think, you know, after Muse went out, all of these companies which are listed, which have this kind of like nuisance subscription kind of, you know, ethic where they kind of get you to subscribe and then you kind of forget about the subscription. Gyms being one of them. So famously, gyms sell more memberships than their facilities can support because people sign up for memberships in January because they have a resolution for the new year, and then they forget about it, and then they feel guilty about canceling it. They're always like, oh, I'm going to go, I'm going to use it, I'm not going to cancel it. And so Equinox, I think, is a listed company. Share price dipped dramatically. So this is the kind of thing I think that will be a way that you see kind of consumers disciplining the corporate world, because consumers actually have a lot of legal power, but it's very hard to exercise and very expensive. And so you depend— you can't really exercise it for like $5 or $10. So you end up like, consumers end up taking it on the chin until You have a class action, and the class action goes after bigger actors once enough damages have accumulated across the ecosystem. And so this is one of those collective action problems that I think AI is going to fix. It's going to make single consumers able to follow the entire thread all the way through. You don't need the class action, the costs of maintaining the lawsuit. is tiny. It's basically, basically maybe like, you know, a dollar to, you know, for like several months of action because it's all like electronic emailing anyway. Right. And so I think that is going to be a pretty significant way in which like the entire commercial ecosystem changes.

Yeah. It reminds me of an episode of the podcast I did once with the CEO of DoNotPay. Which was— Browder. Joshua Browder. Yeah. A little bit ahead of their time in the sense that they were trying to get these agents to do things before agents really worked. And I had sort of a mixed experience with the product, although I did really like the concept of putting AI in the hands of, you know, just retail consumers to try to fight all this nonsense. So I do think that is a great use case. And that is a good anecdote on Muse actually being reasonably generally capable, which doesn't surprise me. I mean, all the kind of benchmarks and stories suggest that it's at least pretty good.

Alex says Muse Grok bot with Opus 5.5 intelligence. That would help you with that.

Well, it's coming before too long, I'm sure. I guess the most important point you raised here probably, though, is the philosophical question at the labs and them not really knowing what they're doing. And I do think, first of all, that's a really important point, that we're really just figuring it out very much kind of as we go. There's some pretty deep philosophical commitment, I think, on the Anthropic side, less so on the OpenAI side. And we'll see who's right. I think I told this story once before. We'll see if I'll keep it brief because we should have a guest joining us in just a minute. I attended this event Recursive back, I think, in May, where the subject was, hey, recursive self-improvement is coming soon. Is it real? What are we going to do about it? How fast is the FOMO, et cetera? One of the most memorable moments— there were a few memorable moments that weekend. One of the most memorable moments was this panel, Corgibility Versus Character, which I think is a pretty good title to kind of encapsulate the debate that goes on right now between OpenAI and Anthropic. This was Chatham House Rules, but it's so obvious which company is which. People can obviously infer who was kind of representing what positions. But I think this is one way in general, I'm like, man, we're doing a depth-first search in AI space. We're racing on just the very first thing that worked? Wouldn't it be better to explore more things? And in so many ways, I think OpenAI and Anthropic are basically the same company. And part of me is kind of like, maybe we should activate that merge and assist clause in the OpenAI charter so that these 2 companies don't have to race each other to the brink. The one factor— there might be a few more that aren't so apparent from the outside, but the one factor that really pushes against that, I think, is the fact that they are pursuing such different directions on this forgeability versus character or obedience versus benevolence, as you put it. And I do, I do kind of want to see how that plays out. You know, there's— I think there's good philosophical reasons to think that pure obedience can't work with enough power, that it will just go haywire. But there's also a lot of evidence of Claude getting a little big for its digital britches, too. So I think both are very unproven. And in my sort of dream hypothetical world where maybe these companies do activate the OpenAI charter clause and start working together much more closely, I would hope that they would maintain those 2 schools so we can kind of see how they Compare and contrast and hopefully eventually get the best. It's not clear to me that either one really has the right solution, but I am glad that there are at least 2 schools of thought on this super critical question. And in that conversation, I mean, the big thing that jumped out at me was they recognized that they were both like, yeah, we both don't know if our approach is right. Maybe yours is right. We kind of believe more in ours, but we're really not that sure. And then the other thing was, it's all academic philosophical talk if you can't translate it into model behavior. And as of that time, neither company, and still today, obviously, neither company has really translated their spec or their constitution into reliable behavior. And this is where I think, as the capabilities have gotten more advanced, I do think there's kind of a convergence now between practical product safety, the sort of thing that Jensen says you have to own as a company, and the more like big picture doom type stuff. Remember, Ajeya said Hugging Face incident was, in her mind, halfway to total AI takeover. So if that is true, or even if it's 10% true and it was only 5% of the way to total AI takeover, then these issues are starting to converge now. where I think it's less of a, like, avoid offending versus existential doom, and it's more like, hey, these guys are starting to climb the existential doom hill. We better take care of it, both for practical and for big picture reasons. So, yeah, watch that space. Obviously, few questions bigger right now than the obedience versus benevolence or the corrigibility Versus character, and it couldn't be more wide open, I don't think, really, in terms of which way is ultimately going to be better. And maybe there'd be some hybrid. With that, I'll shut up and let you introduce our first guest.

All right. Let's see. Yep, there we go. Our next guest is Andrei Georgescu. He is the co-founder and CEO of Vivodyne, a biotechnology company that grows human tissues for experiments outside the body. Its researchers combine living cells, tiny fluid channels, robotics, and machine learning to study how treatments affect human biology. The company develops models of more than 20 tissue types, including lung, liver, intestine, and bone marrow. These are experimental representations of parts of organs with features such as blood vessels and immune responses, rather than complete replacement organs. Andre studied at Cornell and earned his doctorate in bioengineering at the University of Pennsylvania, working in Dan Hu's organ-on-a-chip research group. During his graduate research, he helped prepare lung and bone marrow experiments for the International Space Station. That work involved an unusual engineering problem: keeping living tissue experiments running without someone standing beside them to replenish fluids or adjust equipment. He co-founded Vivodyne with He in 2020. His published research includes work on reshaping the environment in which organoids grow, allowing them to develop for longer and become more mature. Vivodyne has taken that combination of tissue engineering and automation into pharmaceutical research. Its robotic laboratories cultivate tissues, administer treatments, and collect measurements that researchers can use to investigate disease and choose which medicines to develop. In May 2025, the company announced a $40 million Series A to expand that work.

Andrei, welcome to the show.

Hey, thanks so much for having me.

Let's dive right into it. Organoids on chips, how effective, how useful are they in actually predicting how drugs will affect the human body?

Yeah, so these are not what most people think of as conventional organoids, right? These round kind of clusters of cells. They are like large pieces you might biopsy out of someone's lung or liver or kidneys. And the responses that they have to all the different perturbations that we apply to them align exactly with what happens naturally in the human body. And so we see both in cases of like disease and in health, right? If you have a healthy tissue or a diseased one and we perturb it with some drug, that the responses align exactly with what we've seen in clinical trials, for example, in people. And so this ability to just have, you know, human testing at your fingertips on all these different types of human organ tissue, we think is the kind of critical path requirement to testing on humans at scale before actually testing in people.

How do you ensure, for example, I imagine if you have a single organoid, you know, some human cells, I imagine flooding them with the dose appropriate for a human being as a whole is probably not what you want to do. How do you actually, like, figure out what the dosage is, and whether or not, you know, that is actually what the cell would encounter in the body, If it's not, you know, if it doesn't have a chemical reaction before that, before it actually encounters the cell, like, how do you, how do you make sure that you are actually, uh, studying exactly the conditions that the cell itself would have in the body?

Yeah, so that's, that's a super good point. So, so, you know, I think for, um, basically everything that's been done, uh, preclinically to date, right, outside of testing in animals, um, you are just dunking, you know, some substrate be an organoid or you have cells in a Petri dish, just basting them with this test compound that you want to test. That's exactly the problem is that's not how transport happens in the human body. I think it's a perfect example. A lot of cell therapies that are designed for solid tumors in a dish will completely kill that tumor. In a day or two, that thing is reduced to little blebs and pieces. Then, those same CAR-T cells, when they go into a patient, just flow right by the tumor through the blood vessels. They don't even know it's there, right? And thus is the challenge, is modeling the transport of that therapy into the tissue. Do they recognize the tumor? Do they concentrate there? And so, in Vivodyne tissues, even the blood vessels, the kind of capillary bed in that tissue, is self-assembled. So, these little blood vessels grow within the tissues, and we can perfuse directly through them. And whenever we dose a drug, It is always through the blood vessels that are native to that tissue. And so, the transport through the interface of the blood vessels and then the meat of the tissue, all of its parenchymal, its functional subparts is just like you would find in a normal tissue. The density is the same. And so, that is just as big a part as does the drug molecularly bind to its target and perform this effect as does it even get there.

How do you measure the output? What kind of sensors or what kind of output readings are you getting? What granularity do you have these readings at? What frequency? What volume of data are you getting out of making the perturbation?

Yeah, so we're able to do full 3D scanning of these tissues at single-cell resolution. So you get this extremely detailed 3D image. 3D scan of each tissue, and then we do single-cell sequencing to get deep transcriptomics of that tissue. And then we also do a 1,000-plex proteomics to see what these cells release into circulation, right, and impact how they signal to the rest of the body and to other cells within that tissue. And so, this is like depth of data that has been unprecedented even in clinical trials where if you take a biopsy from a patient in a clinical trial, at even one time point, it might be considered too invasive to do in a clinical trial. And so, the ability for us to do this at extremely massive scale, to test tens of thousands, hundreds of thousands of different conditions and gather this deep data across all of them is really the big driver of this feedback loop of like, I see exactly how a tissue responds and its structure, how it is shaped, if cells are doing something strange, if the shape they take is different, And then look at their transcriptome to see what could be driving that causally. So, what is the underlying signaling in each of these cell types and then in signaling between them that drives this? And then their proteome to align with anything that happens before or just after a protein is translated.

So— I have a ton of different— go ahead, Prakash.

No, go ahead.

Well, okay, then how about I'll go with one on scaling laws? I'm very intuitively bullish on this. Just go collect an obscene amount of data on this kind of tissue perturbation paradigm. It makes a tonne of sense to me. The sceptical view that I've heard from time to time is you can never have enough data. The world is too big, biology is too complex. You can do that, but you just won't be able to generate enough. What do we know? What have you learned about the scaling laws? And how big of a moonshot is it to actually get to the point where you get the results that we all dream of? And maybe you should qualitatively describe what are the results that we all dream of? Are we solving biology? Are we curing all diseases? Maybe give me the end state first and then project the scaling law to get there.

Exactly, right? So, we think the easiest way to view what we're trying to do is if you just imagine a patient. So, I think something like, I mean, cancer is an easy thing to imagine where the diseased version is you have a tumor, the healthy one is the tumor is gone, but the rest of the healthy tissue is still there. You might also think of something like fibrosis where you have a diseased tissue that's extremely fibrotic, extremely stiff, a whole bunch of collagen, and so on. And then, you have the healthy version of that tissue that is much more compliant but still has all the same cell types. And, the fundamental question is, how do we make this diseased tissue healthy, right? Especially when the disease itself is extremely complicated and it's very unlikely that we just drug one thing and magically the thing is solved. It might take multiple stages of therapy. We might need much more complex therapies than just single-target therapies. And so, that is what we want to do. We want to start with some diseased tissue and say, what I dose this with? What targets do I drug and with what to make it healthy? Then, to address the scaling laws, ultimately, I mean, our target is to make it possible across all of these complex diseases to understand in as general kind of mode as possible, like, what are the therapies that work on as many people as possible? In cases where we have to fragment that, like in cancer, into much more personalized therapies, To understand how those should work, right, on a patient-by-patient basis. And so in scaling, like, how to get there— and so I think the main question there is, like, how do we understand, like, physiology, right? And, like, specifically in a causal manner. Like, we want to know, if I start with this diseased tissue, what happens when I do this thing? So that ultimately we can back-solve. If I know the thing that I want to happen, then what should I perturb it with, right? And the challenge of this argument that it's just like this, just a size of dataset thing, is that currently even the state-of-the-art virtual cell models that exist saturate at a very, very small fraction of the input data that is fed to them. So you can have all the data you want, but their performance saturates after a couple percent, right? And this is like, you know, both published and kind of what we hear on the ground, right, in these Perturb-seq kind of based models. And the reason it gets so difficult is when cells are growing in a dish, which is the kind of mode for this sort of Perturb-seq, they are so far removed from all the feedback loops that are natural in people that they're just trying to colonize that piece of plastic, right? It's like you have a stiff substrate. Normally cells, when they encounter that, are like, I have to encapsulate this. And so they are just growing. Cells that make blood vessels normally, they're not making— they're growing on plastic, right? Cells that, you know, fibroblasts that make all the different collagenous parts of the body, they're just growing on the plastic. Same thing with the epithelium, all this. Their only goal is to proliferate. And so if you knock out genes in those cells that would otherwise affect how they perform their natural function, really there's no big difference, right? They're still just dividing. And so You end up learning this view that perturbations don't really do much of anything, or it's unclear what they do, except if a gene is like required for that cell to live, in which case like it kills the cell, right? Um, and so, and so it's not— there's not even causality to extract from that, right? Whereas in, uh, human tissue, they're, they're like the, the causal basis of of how it responds to something, right? Like, what, what is the sequence of events in, in the, in the response that kind of integrates this thing that we do to it? Um, there, there's far, far more complexity, and the cells are much more functional in what they're doing. They're not just proliferating, right? In fact, only a very, very small fraction are proliferating. And so we think that the, that the, that the fundamental bottleneck to having this causal understanding of biology is that we need to perturb these tissues in the complex environments that they call home naturally, right? And so being able to do this in diseased tissue, being able to do it in healthy tissue, being able to just populate this big latent space of states that these cells can occupy and to understand the paths that they take to get to these different states, almost like the image in my head is we are trying to populate a map And kind of like design a GPS through it so we can say, if I want to get to the state, I need to, you know, like go down this road and then turn left with, with a second perturbation, and that'll bring me to this healthy state. Um, and then, and then the ability to learn, um, to learn those states in biology to a much better depth that we can do so by just, um, you know, giving it like biased features. The number of cells that are positive for this thing and the diameter of blood vessels and so on. We can establish that shared basis by contrastive kind of training, like these multimodal models that we can train to understand what unifies 3D scanning of a tissue and its spatial structure, its transcriptome through single-cell sequencing, and then its proteome and all the proteins it's making through proteomics. And so, that gives you a whole bunch of data. And then now, it's like, To kind of fill out that matrix, you need a whole bunch of observations. So having the throughput to do this testing at scale and the consistency in all this becomes the reason for all the robotics that we create.

Speaking of data, I have heard that in order to get tissue samples, cancer tissue samples, you sometimes have to pay large university labs. it's rather enormous amounts, like $500 a sample, $1,000 a sample, etc. When you do this, and you need to get sampling for some kind of cancer which is relatively rare, do you encounter this problem of, can we find enough diverse samples in order to populate the model? Does that become an issue?

Yeah. So, in the sourcing of cells that we do, I think one of the one of the really nice advantages of having great partners in pharma is they have themselves biopsies from clinical patients of their own. And so in partnership with that pharma, if we're working in discovery, for example, in oncology, we have at our fingertips these results from actual clinical trials, right, where we know the drugs to which a patient has been subjected, if they're on their first line or second line or third line of therapy and so on. And so it gives us a strong advantage there, being able to partner with pharma. That's one of the reasons why, beyond their expertise in molecule making and commercialization of drugs, that we partner so deeply with them across the pipeline.

Indeed.

Can you— go ahead. Well, okay, we're colliding a little bit. Can you tell me more about the interface between the automation and the biology. On the website, I noticed this bit about the most advanced kind of fluidic chip. I'm not sure if I'm even using the right term there, but this sounds like a very— like an interface that is a glimpse into the future of probably many things.

Yeah. So it's a I think like the generalized set of kind of problems that a lot of silicon manufacturing has to deal with is kind of no different here, right? So it's like, how do you interface across scales and how do you interface at the appropriate like bandwidth, right? If you want to call it that, with the things that you're interfacing. And so with chips, it's like, well, I mean, obviously at the level of the transistor, right, you have multiple layers of connections between those transistors as you pattern them larger and larger. And then you have interposers and you have all these things, pins that then go to a motherboard and so on. With our tissues, it's a similar kind of thing where it's like, okay, you have the cells that are in a tissue and they interface with each other, a lot like transistors in a die. But then you have ultimately this chunk, this piece of tissue that has to grow somewhere. And importantly, in the way that we grow them, we have to be able to perfuse kind of continuously through the blood vessels in that tissue, like a little capillary bed. And, we need to do this across tens of thousands of tissues at a time in our automation. And so, the interface there is extremely large but intricately patterned microfluidics where we can control the flow of liquid through all these different tissues and then control the dosing of the tissues, the way that we divert different compounds and drugs between one tissue and another. And, it helps dramatically to have— so, we have this large, maybe plate-sized microfluidic disc that is patterned. Just imagine a wafer, except instead of patterning, doping it with transistors, we are patterning chambers for tissues that are centimeter, inch scale across it. one beside the other. And that gives us the ability to grow many different types of tissue within these chambers, controlled by the different cell types and inputs that we give them. And ultimately, it is like this scalable framework for manufacturing this at scale. Ultimately, what are we making? Well, it's a It's like an integrated factory for growing and maturing these tissues and then testing on them and then gathering the data at the end, right? This, like, yeah, so we need the machine that makes the machine in a way where the machine is this biological tissue and the microfluidics are the layer between those things.

How do you go from a sample received to something placed on this system And growing. I could imagine you might get a chunk of sample and place the chunk, but it sounds like you're starting with much smaller bits, and then they actually have to take and grow in situ, right? In that, that kind of new environment that you've created for them. That process sounds pretty delicate. What's it look like?

Yeah. So, so we are getting a lot of sample. Right? So, a lot of different cell types. And, these can be from biopsies of things like tumors, blood draws for different cell types. And, when we grow these tissues, a really big differentiator for our approach to doing this, and I think the really critical mechanism that you need to get realism in this, is that we use primary cells, which are mature differentiated cells that perform our natural organ function in our bodies, as opposed to starting with some stem cell completely absent from all the conditions that would tell it how to mature into a differentiated cell, and then sledgehammering it with cytokines to turn into some cell type that we think is accurate to what happens in the body, right? We instead just kind of go to the mature source, the ground truth, and we mix together all of these different cell types that comprise the lung, for example, at very high density. And, within that injected tissue, we see that these cells begin to reassemble into the native structure of that human tissue with some little herbs and spices that we introduce. But, the responsibility for forming the native structure of human tissue is on those cells that already know how to do it. Just, I mean, like, they do it in us, right? So they better know how to do it. And so we rely on them to do that. And because this process is a lot like, it's honestly kind of injection molding into a chamber, right? So we, like, shoot in these cells at, like, a high density, and then over the course of several days, They reassemble themselves. They just self-organize to make blood vessels, to partition into, in the case of the lungs, the airway where there's air, and then the parts where there are blood vessels and fibroblasts and immune cells and all of this. So, we are relying extremely heavily on the cells to be able to direct their own self-assembly at these very small scales.

Let's talk a little bit about the clinical process and the clinical failures, et cetera. I think there have been some studies that show something like 8% of approvals for phase 1. There's only an 8% chance of approval for phase 1, right? And so, what does your technology enable? Is it reducing— is it identifying toxicity early in the cycle? Is it rejecting weak candidates earlier, elimination? Like, how do you— like, what clinical output does using the technology actually create?

So I think if we imagine— so imagine without the ability to test in anything else, right? You had some compound, some drug, and I took it. It would answer a lot of questions right away, right? Like, someone like, does it work, right? If, if there's some disease I'm trying to drug, um, you'll, you'll probably learn one way or another if it works or not. And then you'll also know if it's safe, um, right? Like, does it have an impact somewhere in my body that I did not anticipate? And so, and so the, the kind of goal here is like, we— like, these, these, these human tissues should generalize beyond the kind of traditional, uh, specific scopes of— I, you know, it's almost like, um, you know, early, um, early attempt, you know, convolutional neural networks for image analysis, and you had completely different architectures for all these different types of models, would be able to solve these very narrow problems. But the moment you're generalizing or you want to ask a question slightly outside of that scope Like there's no tool for it. You have to invent a new tool and a new tool, and then the results from one tool don't translate to the other tool, right? And are hard to compare. And so, and so the kind of drive here is, well, we know that if we'll have to manifest hopefully the answers that we have for them, right? Is it safe? Is it efficacious? Does it go where it needs to? Does it follow the dosing that we that we thought would be effective. And so, in our platform, by having all these different tissues of the human body, we basically could do the same thing. So, at the very beginning of drug discovery, the big question in a new program is, if we want to tackle this disease, how does that disease even come about, right? Like, what are the differences between healthy and diseased tissue? From that, you get these hypotheses like, it seems like this pathway of signaling, right, or these receptors on these cell types are more active in disease. What happens if we drug them? And we can then test that without changing anything. So we take our diseased airway tissue and we drug that receptor, or even easier, we just delete it from the cells that we want to test it in through CRISPR. At the next step, you're wondering, okay, well, it seems like that experiment worked well. It seems like we have a possible validated target on our hands. What type of molecule should we drug this with? And whether we're testing small molecule candidates or biologics or, you know, siRNAs and gene therapies and cell therapies, we have this perfect human tissue substrate on which to test all of those things, right, and choose which one works best. Then we can get to a point where it's like, okay, well, we need to optimize this modality and this molecule now to work even better on that target. Well, how do you optimize it? Maybe it's going somewhere it shouldn't. There's some nonspecific transport in the liver, right? The liver is like super sticky to all these different molecules. We want it to, you know, we want to put like Teflon, you know, whatever, so to speak, on this molecule. We can also do all of those molecular perturbations, and then we can test each one on our tissues. Then later in toxicity, we can take that same molecule and look across not just the diseased tissue and see if it becomes healthy, but look across our portfolio of healthy tissues and say, are any of these impacted negatively, right? What when they get dosed this drug. And then finally, for clinical trial design, we can look across this breadth of different donors from which we can grow these tissues, right? And we can say, um, you know, this does not work so well on a, uh, you know, the case of cancer. Like, if someone's a third-line patient who's taken these 2 drugs first, it's unlikely our drug is going to do much rescue. Whereas if it's a second-line therapy for someone who's taken this drug first, like, we see— we would see huge improvement, right? Just by virtue of being able to test all those perturbations. Or patients whose tumors have these biomarkers, or patients with, you know, this blood serotype might be more susceptible to autoimmune reactions to this issue. All of these questions we can answer in designing that clinical trial to make sure that the entire data package of, is it efficacious, is it safe, will it work in these patients for which we indicate it, all of those can be answered like before it goes to the clinic. And the biggest problem of today, right, in drug development is Like, all of those questions are unknowns. We have some mouse data for this part of it. We have a little bit of cell-in-a-dish data for this part of it. There's some, you know, custom assay with some, you know, reporter lines that were generated for this part. And you're going into a clinical trial just like, man, I hope it works this time, right? And ultimately, your success rate is 1 out of 20, right? And so, and so By being able to test on human tissue before testing in people, we kind of get answers to all those questions right before we get there.

Can you tell us a little bit about the mix of in silico via the foundation model versus in cell, in actual tissue experimentation today and how that ratio is expected to evolve? Because presumably the big value is It's a lot cheaper, faster, even if you've got all these great wet lab technologies, just ploughing things through the foundation model is supposed to be where you get the insane acceleration, right? But then, of course, you've got to ground it and validate. So what does that mix look like? How is it going to evolve over time?

That's exactly right. So if we look at it like a reinforcement learning problem, which is my actor is this Well, it's kind of the experiment, right? I have this tissue and I can perturb it in very many different ways. And the goal is I have this objective, which is to make it healthy, and I can instrument the state of that tissue very deeply. So I know, like, the value function is very well defined in terms of the state of that tissue. The question of what should I try next? In the large rollout that I can perform, what are the most useful things to try next that add the most information to my dataset? The physical experiments serve to fill in gaps in which the model is most uncertain. The value of the model ultimately, why even make these foundation models is The combination space, especially when we move into 2-target or 3-target therapies, right? If we have just combination therapies, if I take 2 drugs because it's going to work much better than one, is so huge that you have— I mean, you could have like this whole planet could be Vivodyne systems growing human tissues and you'd still not get anywhere close, right, to the just sheer space of things to test. And so we need some way to narrow down to, you know, to, to kind of just, uh, uh, you know, a couple times like the total number of clinical trials in America, the, the experiments that, that we should physically run to, uh, to kind of test our hypotheses, right? And so, and so basically these foundation models allow us to predict what happens when we drug this thing and that thing or all these things together or this thing and then that thing. And, we kind of optimize and we iterate in that space and we find, for example, the 1,000 most likely things that will move this tissue closer to this healthy state. And then, we close the loop by testing it. And, we see where that rollout goes. And then, we just do that on a loop. So, the purpose very fundamentally for us of these models is to tell us what to do next, And then the purpose of this experimentation is to figure out if the thing that we thought we should do next actually does move us closer or how it deviates so we can correct.

Switching gears a little bit, VivoDyne announced something called the TissueDisc 2, which is, I think, your second-generation disc. Can you describe— and as I As we have understood it, it is basically a disc in which you have the cells growing throughout in an engineered matrix so that you don't get this clump, which basically is not that useful. Can you describe to us what was the big design change from Disc 1 to Disc 2? How is it improving from generation to generation?

Yeah, so when we design our microfluidics, one of the— it's like a quality problem to have, I think, now is like we have these systems that can automate the culture of many of these tissue discs within each automated platform. So to give like a mental picture of how this looks like, imagine something the size of like a wardrobe, right? So about the footprint of a large desk, about 8.5 feet tall, you know, surfaces looks like anodized aluminum, like this MacBook. And there's a window on the front where we can see what's happening inside. And these are self-enclosed labs. So fridge and freezer, and there's a 3D scanning microscope in there and big robot arm that can use all these different tools. And The decision is that these discs, these substrates, must house all the tissues within them, right? That is our substrate for growing these tissues. And we can load reagents through other form factors and so on, but tissues must live here. And the kind of optimization that we make is how much of the accessory functions of the experiment should we also put onto the disc? So in other words, You know, in a lot of— I mean, currently, right, the way to maximize like bandwidth in GPU design is if you get an interposer and you just like put it all on there, right? Memory, like, you know, stack the memory, like do all these things. In our case, it's interesting. It's a little bit the opposite where it's like we want to maximize the usage of that surface area of these disks that are optimized for imaging and for all this. We want to optimize the surface area of those disks 4 tissues. And so in our second form factor, we have an increase anywhere from 2 to 4 times the density of tissues that we grow on each disc without any sacrifice to— they're actually much larger than the version 1 of these discs. So we increase the number and the size of these tissues, and we offload a lot of the responsibility for how they're grown to the robotic system around them. Now, the challenge there and the kind of work that allowed us to do this is a lot of work in software on the orchestration layer, right? So, it's like if you think if you had 10,000 experiments to run in parallel, and this is not like I have a different gradient of dosing for some drug or something, right? I can do it on a plate. I am changing the collection point of data. I'm changing the interval between dosing with some drug. I have a first-stage regimen and a second-stage therapy. I'm taking all sorts of different readouts of these tissues. It becomes a super, super complicated experiment. I can define it quickly. I can explain it out loud and say, well, we want to test bone marrow and airway. We want to dose these drugs. this order. We want to branch to these concentrations, and I want 3D imaging, single-cell sequencing, and proteomics. And by saying that, I can define the study decently unambiguously, right? I mean, given the system of tissues that we grow, it's pretty unambiguous what I would want to test in there. But then the robot has to be able to decompile— or sorry, to compile that kind of zipped representation of the study It has to decompress it into this compiled list of like millions of actions that this robot has to perform. It has to remember that if it took the lid off of a plate, that it should put it back, right, as an example. But if there's multiple trays that it has to handle and they're all out on the working deck, maybe it should delid multiple of them in advance, right? if multiple pieces of plasticware are on the same tray and they're coming out of the freezer to thaw, you want to make sure you're not thawing something that should not be thawed too early. And so all like this, this, this compiler behaves a lot like a compiler in code, right, for C, where you're managing like the prefetching of information from memory and branch prediction, like all this stuff, right, that like thankfully our operating system And hardware handle for us. But improvements on our orchestration mean that the system itself, the hardware, the automation, can do much more of this. And more and more of that space on these tissue discs can be used for actual tissue biology instead of having to integrate the little helpers on there.

How close does all of this put you to being able to do N-of-1 studies where some rare cancer, nobody really knows what to do. Can you get a sample in, culture it, grow a bunch of it, run this loop on that one person's thing and come out with a treatment plan? Are we there? What's the roadmap to get there? And is it time to start working on a new kind of postclinical trial paradigm for medicine already today, perhaps?

So we can definitely experiment on I mean, almost like by definition, every tissue we grow is like from a patient, right? And so, actually, like when we're doing these large experiments, we have to be sure to like sample from multiple patients, right? So, we're not overtraining, right, if that makes sense. There's like 2 big challenges, right, I think. I think a lot of medicine It is still a big challenge. If you have— so, to take, let's say, the most promising approach of today for very, very personalized therapy, which is a cancer vaccine. You say, well, this tumor is expressing this biomarker, and we can target it through that biomarker. So, I'll make a cancer vaccine for that patient. I'll take A piece of— 2 ways to do it. One is you just take a piece of the tumour, and you start to kill the cells, and you put it back into the patient, and you let the patient's own immune system try to fight it. The problem is the patient's immune system has been suppressed by a lot of the chemo they're taking. So, it's a little tougher, especially like second-line, third-line of therapy. Another way is you make some kind of— it's called a neoantigen. Like, you take some part of that biomolecule, and you provoke their immune system in a way where they develop autoantibodies kind of against this deformed protein or something on the cancer. And one of the things where you run into challenges in cancer is if you select down for that target that you've now given, the tumour can still escape. So, it just learns not to make that epitope anymore on its surface, right? Oh, I should just stop. It seems like all my buddies, when they flash the sign, are immediately killed. I should probably stop doing that. Or more accurately, it's just like the cells that happen to, by random chance, not make as much of that protein continue to proliferate. And so, it's still not— I mean, if If the means were possible to pretty repeatedly and across many types of cancer do personalized therapies, there would still be at least a little bit of generalization you could do, right? I mean, there's a reasonable amount of consistency in the cancers people get even. There's obviously a little bit that is person-specific or tumor-specific, but also enough to kind of grab a fingerhold on a little bit of generalization. But it is like the lack of understanding, like, what happens when I do this thing really, that still prevents us, right? When you try to attack cells that are quickly dividing in a patient, right, like the earliest stages of the earliest kind of versions of chemo just target cells that are quickly dividing. And you get all these problems that are known in chemo, right? You get immunosuppressed because immune cells are quickly dividing. You start getting problems with your gut and digestive system because all those cells on the surface are quickly dividing, right? Skin gets all dry because, again, your skin is quickly dividing. Hair falls out, also quickly dividing cells. And this kind of challenge pervasively is you don't want to dose too much of the drug that it kills, you know, things it shouldn't, or at least too much of those things while giving enough that it kills the cancer. You know, my kind of take on like personalized medicine and how this whole thing kind of scales and remains cost-effective, right, is like imagine the cost if I wanted to make a Like, if you want to make a product for one person, right, um, that, that takes a very different road than if you say, what do all of these people want that we can provide? Um, with medicine, I think the much harder thing is to make, uh, generalized therapies that can be mass manufactured and like provided cost efficiently, right, to people. Um, now that might take the, the form of like the modality, right? So we could say, well, it is still, you know, patients come in with their own cancers and we take a biopsy and we see what types of antigens they have, right? And based on predictive data we have, we would dose them with this neoantigen or that one or a combination thereof. But I think ultimately, like, what a cure to cancer looks like, right? And this is like, you know, cancer is a collection of all the different forms it can take, is a far, far more sophisticated therapeutic, and probably not one for all cancer, right? But at least like the main domains. I have a solid tumour of this type of cell. I have a broad-acting and intelligent drug that can target the right things and not do bad things where it shouldn't. I think that is like the hard thing to produce, and ultimately might be the way that we sustainably cure these complex diseases in a much more scalable way than one-offs. Though I think personalized medicine will kind of act— it's this weird thing where currently, there's very little personalized medicine, right? There's a little bit of it, but it's not at the individual level. I think we'll go through a phase where there are some cancers that are curable through personalized approaches. And then, as these these kind of capabilities advance and we can add more intelligence to the drug itself, then we get back to more generalized kind of models, right? The same way that I think we see AI generalizing very broadly and kind of feeling this growth through its generalization.

On the Annual Review of Pharmacology and Toxicology's Grand Challenges in Modern Pharmacology, the innovation, a universal in silico or organoid framework that forecasts rare patient-specific adverse events with more than 95% accuracy before first-in-human dosing, is one of their target challenges.

Wow.

Vivodyne is obviously on the path there.

When do you think this challenge So this is a hot take, but I feel like, I feel like, um, this is almost like, uh, picking the color of my Mars suit currently to me, right? Which is like, how about we just like kind of fix the bigger problems? And, and, and the big problem, by the way, we, we, we are working on specifically that challenge, right, of, of, uh, of, of, of like very, very rare outcomes. But a bigger, like, Far more regularly, we have drugs that would have huge potential in their efficacy in patients, but that have routes of toxicity that are pretty common. Like, many patients die from it, right? I mean, many of the cancer drugs that are going to clinic today, the risk is not like, oh man, I've cured 999 people of cancer, but this one patient, like, it didn't work so well. You know, it is more toxic. The problem is, in all the patients we've dosed, they barely had any improvement in their cancer, but it gave them all of these side effects or killed someone from liver toxicity or something. Right. And so, and so toxicology is nowhere near as much of a solved problem as this, like, oh, well, it's just the rare cases, right, that are, that are hard, kind of like Angus suggests. The main cause of those very rare toxicities is usually immune driven. So you will have an immune system takeoff where it's like— to give an example of how it happens in the liver, you take a small molecule drug, right? And then it goes into the liver cells in your body and these cells metabolize that molecule, which means they like break it apart. And in those intermediate states, parts of that molecule might be really reactive. And so by chance, they will stick to another protein in the liver And then that protein leaves the liver at some point, or leaves the liver cell, and makes its way up to, you know, your lymph nodes. And your adaptive immune system, which is now seeing this new shape for the first time, right, this modified molecule, will think, holy crap, this is a foreign— like, this is non-self. And it starts to mount an immune response against that molecule. And if you continue to take the drug, You continue to stimulate the immune system with that kind of side reaction, and you end up developing an immune tolerance to that drug. So anytime you take the drug, you start to mount this big inflammatory reaction. And in some cases, the drug can do some pretty strange things where you get autoimmunity that lasts even after you stop taking the drug, right? Like maybe the antibodies that are developed against this modified protein start to target the unmodified parts of the protein, and then you're stuck with an autoimmune disorder. Now you have to take immunosuppressants, right, for the, for the rest of your life. I mean, um, and so, and so the, these are challenging things to, to identify because usually they happen in a very, very rare percentage of patients, and there will be like one immune cell, right, in their bodies that, that can recognize that modified, um, intermediate and like start to make antibodies against it. And so really, it's like the challenge becomes, can you prove to me that there's no, like, you know, that every grain of sand on this beach is not a poppy seed or something? Except maybe not so crazy, right? Maybe like out of these 1,000 grains of sand. And so the way that we do that is by virtue of just being able to test through a whole repertoire of They're called TCRs, right, on these immune cells that can bind to all these different molecules. We can screen across a whole host of those, including in our lymph node, and we can see if antibodies— so first, this unique receptor on some of these immune cells called the TCR has to grab this thing, and then it takes it to your lymph node, right? Or the molecule goes to your lymph node, the TCR grabs it there, and you get the development of an antibody. And we have lymph node tissues that can produce antibodies. And then the adaptive element of the immune system courses through all of the tissues that we grow. And so, the ability to test across a wide enough breadth of patients is basically how we get there with that challenge. Mostly, though, the big questions of toxicity of today are like, this is a cancer drug, and if we dose a little bit too much, we kill the patient from the drug. But if we dose too little, the patient dies of the cancer. How do we make a drug that has a wider threshold through which we can drug, and how do we find the right dose? Or, this is a fibrosis drug that could treat fibrosis across this huge swath of patients, but it might actually make the fibrosis worse because we don't understand causality well enough. Can we show that it actually does help and not hurt progression of the disease?

I think with your recent TissueDisc 2 and the biological data center, I think you had something like 3 million or so samples that you can do per year. What do you think, how many samples a year do you think you'd need to get up to in order to solve this kind of grand pharmacology challenge? Like, you know, like just, you know, just an estimate, 30 million, 300 million?

Yeah, yeah. Well, I think without increasing the total capacity, I feel like there's not so much something to solve. It is just the breadth of the search for one such— like, given this compound that we want to take to clinic, now is the time where we engage that mode where we're just screening across 1,000 different patients to see if they have the response. So I think the capability is already innate, but we have not— across all the different partnerships we have, we haven't gotten to this point of saying, Maybe there's going to be one patient that responds poorly. Mostly, it's like, will there be therapeutic effect, which is the first big and usually the hardest part of clinical trials.

André, you've been very generous with your time. We've been extremely inspired.

Appreciate it.

I do hope that we see you solve those challenges and help create better drugs for all of us. Um, thank you. Thank you for joining us today.

Thank you guys for having me. It's been terrific. Thanks again.

Keep up the good work.

Will do.

Here we go. There we go.

Uh, let me introduce our first guest for today. It's, uh, Nick Gillian. co-founder and chief technology officer of Archetype AI. The company builds Newton, a family of AI models for interpreting measurements from the physical world— vibration, electrical current, temperature, video, and other sensor data. Its applications include identifying changing machine conditions, finding unusual events, and helping people understand how work is being performed. Nick led the machine learning research in Google's Advanced Technology and Projects group, contributed to Soli, Jacquard, and Pixel. His research there includes RadarNet, a system for recognizing hand gestures using a miniature radar sensor inside a phone. At Archetype, he now leads work on a broader problem: how much of the intelligence needed to interpret a new machine or environment can be learned in advance? rather than engineered separately for every application. Hi, Nick. Welcome to the show.

Hey guys, good to see you.

Let's just dive in. So Archetype builds models that control industrial equipment, essentially. So let's talk about the challenges in dealing with these pieces of equipment, which are often very different from each other at an oil rig versus a construction site, and trying to build models which are able to handle either one without specific optimization or customization for each piece of equipment.

Yeah, yeah. So the big problem we're solving is with physical AI, most people, when you say that term, they might think of maybe Humanoid drones or maybe self-driving cars and things like this here, where maybe the modalities for sensors are typically cameras or maybe a LiDAR or something like this. Maybe they have motion sensors. And typically the output control is some form of actuation of the humanoid. It's you move the joint fingers. It's maybe the drone, you steer the drone. Or if for the car, it's, you know, you're—

Steering wheel.

going to park the car, you're going to drive down the highway. We are taking that idea and expanding it much broader than that, really, to every single type of sensor in the world and every single type of control surface that you might want to plug that up to. So the most obvious use cases for that are today already in large industries, in manufacturing, in energy, in agriculture, in all these areas, there's already billions of sensors deployed in the world. There's already very sophisticated control systems, and they've been there for decades, actually, typically driven by more classic signal processing or heuristic algorithms. If this signal is higher than this, then trigger the alarm and so forth, right? And really what those systems are missing is the intelligence layer to be able to take in all those sensors and actually be able to understand what's happening in the world, adapt as things are changing, because the physical world never stays stationary. And then be able to map that to control surface. So one of the biggest problems we have to solve is with self-driving cars and drones, typically the input modalities are fixed for every robot. You have a number of cameras, you have a LiDAR. When we go and talk to our customers, they have hundreds of thousands of different types of sensors, of different configurations, of different types of fleets. Maybe they have old systems, they have new systems, they all have to be joined together. So one of the big technology problems we're solving is how do you build a foundation model that can generalize to any type of sensor and that when you deploy it, as you mentioned, down an oil rig or you deploy it in a large industrial furnace that never turns off and it's continuously spilling out liquid molten products, or how do you connect this to an energy grid and help understand and optimize this, or how do you connect this to— a data center where you're monitoring all of the HVAC systems, the air cooling, the temperature of the CPUs, and you're trying to balance that very complicated ecosystem. How do you dynamically plug all of those into one foundation model, have it understand the world, and let it control it for you? It's a very big complex problem, and that's really what we're trying to tackle here at Archetyp.

I want to go every layer from the bottom to the top. Let's start with the data. How much data is out there that you were able to just, you know, collect on your own? And then I imagine you must have had to establish a sort of partner network to bring in a bunch of data. Tell me about the data process.

Yeah, so we're getting close to a billion hours now of physical AI data that we've been able to scrape and gather and collate. In many cases, if you look on the web and you look at the different types of physical data that's there, it's very sparse. There's very little multimodal data. It's not synced well. It doesn't cover things well. For most techniques, particularly classic supervised techniques where you're basically building one model and it has to have a fixed number of sensors and they all have to be captured with the same sample rate, et cetera, that's a massive problem. We've been trying to flip that on its head and say, If we have all of this physical data which is captured in different sensors and distances, different systems with different contexts, how do we actually build the most advanced self-supervised techniques that can pull in that sensor data and extract from it in the same way LLMs have done with language? How do we do the same thing for, uh, uh, this very, uh, very different types of sensor data? It's all captured at different rates and contexts. And how do we actually basically take that very low-quality data and turn it into high-quality data? And then, as you said, we are able to then join that and mix that in with a small amount of data we can capture from partners and even in some cases that our customers can give us to be able to really extract and understand and map that to physical language and map that to physical control. So if you look at this from a standard approach and you look at the— The physical AI data that's out there, there's not a lot of it. But we've actually been able to flip this and get almost 1 billion hours at this point of data we can use to train our models and build these fundamental techniques.

So one of the questions I would have is, you often get a lot of data, but the quality is of differing qualities, right? And sometimes even the same sensor by the same manufacturer in the same job site, in 2 different, you know, will, you know, will give you different readings, right?

Yeah, yeah.

How do you actually extract this kind of like useful signal from the data points that you do have? Like, is there a data cleaning method? Is there a curation methodology that you have to use before you feed it into the model? Like, what is actually going on in that, you know, in that pre-data phase?

Yeah, very, very much so. So I guess we're trying to do as much as we can without curating that data. That's kind of in the pre-training phases. But as we start to go to alignment and really try to align this with language, that's where cleaning becomes very important. Understanding the context that goes with that sensor data is critical. We use a lot of labeling companies, for example, to help us bring in domain experts, for example, in HVAC or, you know, in industrial machines and so forth, and really try to bring in some of those nuances of what could be the context of why this machine, you know, why is this large compressor switching between these modes in this very strange way, right? Typically that's not captured in the sensor data and you have to find ways to bridge that with kind of human expertise, right? So a key part of it is like this 2-phase approach is one, how do you extract as much useful data as you can from just the raw sensors? That's the pre-training stage. But then in that alignment and bringing in that context, how do you then actually take a small amount of high-quality data that is curated well, that you can really trust, and make sure that you use that to really steer the model in the right direction? Yeah, so there's quite a bit of resampling and really understanding missing values and sensors. Is that because the sensor had an issue, or is actually the machine itself is connected to— that's actually a feature that helps explain that machine is about to break. The reason the sensor is giving you all these NaNs is not because the sensor is broken, it's actually a feature of the machine in those cases.

Let's move up one layer to the models. I have a ton of questions around this. Are you pre-training totally from scratch on pure signal data? Is there also a language component To the data? Are you using transformers that we all know and love, or maybe something a little bit more? I've seen a lot of results over time that suggest state space models, Mamba type things, can be somewhat advantageous for signal processing type tasks. And what are the predictions that it is making? We take the sensor data in. Is it a sort of JEV-style classification output or a more reasoning, like, hey, you might want to watch out for this. It looks like your machine is about to break kind of story-type output.

Yeah, very much so. Maybe let's start with that last part and work our way backwards. Essentially, you can think of the output of our model as 2 branches. One branch is focused for humans and the other branch is focused for machines. For the human output, that's much more along the lines of natural language. It can be either long detailed reports like, help me understand why the yield of my factory is down 6% this week, for example. You take in all the sensors, you analyze that. It's a System 2 approach where you have a little bit of time to go over and understand this on a large GPU rack in the cloud, and it can spit out a report to helping you understand what can be improved, what the issue could be. Another type of report that would be on the more human side would be, help me understand why this machine has an anomaly, right? You've detected— you've not just detected the anomaly, but you can actually interpret and explain why this may be happening. So you can actually go back and maybe change something in the field to mitigate that in the future, right? So this is more like the human approach. The second output is much more focused on the real-time output of the 4 machines itself. This is where you want to actually build that control loop and you typically want to deploy this part of the model in the physical world in real-time on much smaller, more cost-efficient, even CPUs in some cases. This looks a lot more like a Jev type of output where you are actually directly controlling the machine. It may not even be human language, it could be control signals directly. Or it could be alerts that are then triggering up to a real-time alerting system that says, I've detected this anomaly in this machine, a technician should immediately go and investigate it, for instance. Our model is able to move between these 2 spaces, between the human output and the machine output through a general transformer architecture. We have other components in there as well. But as you back this through, part of that transformer architecture, the sensor-focused components, those are all pre-trained from the ground up. Because there are no techniques out there that can take radar and LiDAR and cameras and time series sensors and fuse them all together. When we start to bridge this in with a language, we typically try to use as much open source language components as we can so we don't have to train those from scratch. But then the technique is how do we align that with, for example, the radar sensor, or how do we align that with not just one camera but 16 cameras across, say, a construction site that are all looking at different parts of the site at the same time, for example. So it's taking that and merging that altogether. So yes, we're pre-training from scratch. Yes, we're using as much off-the-source language components. A large part of this is models that can give you detailed human outputs in the cloud. But then for most customers, the big advantage here and how this is very different from LLMs and coding agents and other things in the digital world, is being able to take that model with all those insights it's gained from your data and build an agent, a physical agent around that, and push that back to the physical world so it can now run in real time as close to the sensor as possible in typically a secure data sovereign environment. So it doesn't have to leave the construction site or the factory or the car or whatever environment it might be running in.

Yeah. The area that I've studied best that has done something kind of similar has been at the intersection of language and just images, photographs. And we've seen kind of, if I had to borrow some of their terminology and map it onto what I understand you're saying, this would be sort of a late fusion approach where you have a language model, you then do a pretrained from the ground up specialist encoder for these different modalities, and then you're doing some sort of cross-attention between those. That's my understanding of, at a high level, obviously, of what the image folks have done. How much have you been able to follow in their path? How much of the techniques you've seen them develop have worked for you, and how much have you had to go back to the drawing board and reinvent?

Yeah, that's a very good analogy, and I can I can highlight where that type of approach has worked and where it hasn't. If you think about that general approach where you have some sensor encoder and you're taking your sensor and you're projecting it into some latent space, and then you're trying to align that latent space essentially with another latent space that is built from language. That's typically what's happening in a lot of the most advanced vision models today, the VLMs. We've definitely tried these approaches and we've had some success with them with other types of sensors, radar and time series and so forth. But the big issue that we have found is twofold. Number one, there's an extreme amount of language-sensor pairs on the internet that can easily be scraped and built into these very large datasets where you have this pairing between an image and a description of that image. That's what's really unlocked a lot of the— if you think back to even the earliest CLIP models and other techniques like this that then have gone through all of these different more complex techniques like you've mentioned with cross-attention and so forth that allows that technique of aligning vision and language in this pairwise approach to work very well, right? This does not exist for radars. It does not exist for time series sensors. It does not exist for all the complex systems that we go and talk to our customers with, right? So we have had to look at different approaches of how do you, for example, take time series and actually align that with human language. That's a large part of our research work is how to actually solve this sensor alignment problem. There's also the case there where with a single image, you may actually be able to get the full context, for example, of the scene and then align that with language. But actually with the time series, you actually need to look at a temporal window. What does a single latent mapping to a single word really mean in those cases? It's not like you have a time series signal that says dog and you can align that with the keyword dog and so So there's quite a bit of work we need to do there to solve this alignment problem of how do you generalize this across many sensors? And that's the key kind of components we're building inside our models, which makes them different from a standard VLM or other techniques, right? Because we simply don't have that corpus of, say, time series and language pairs to kind of apply that previous recipe you mentioned, right? Similarly, we have customers that come to us with completely unique sensors that are not on the internet at all. They've never left their factories, and they want to be able to take those unique sensors and plug them into our model and then use our platform to do some basically fine-tuning alignment so they can plug in their completely unique sensor to our model. So in that case, there's literally zero sensor-language pairs in that case, because the data has never left the factory in that case. It's a brand new type of sensor that doesn't map onto anything else, and we need to be able to learn that from scratch. These are some of the very hard physical AI problems that we are solving that, for example, most robotics companies don't have to solve because they're using a camera and an accelerometer and a LiDAR, for example, and they can build the dataset bit by bit and they can do it in a very manual approach where they build a, you know, large arm farm and they have the robot do things many, many times and they capture those sensors. We're having to approach this from a different direction.

Let's talk a little bit about— I think you guys did a project with Kojima, right?

Yeah.

Which is a construction company.

Yeah.

And I think you had a huge amount of video data from them and a lot of other sensor data. So let's talk a little bit about how this works. Like, you, you take the data, and typically for an LLM, you're trying to predict the next token. So you feed in a lot of text, and you're trying to predict what the next, the next piece of text is. You do some masking here and there, and etc., etc., but essentially that. So what is the input that you're putting in here, and what is the output that you get out of one training step that you then feed in? Like, how does that, how does that cycle work? Like, let's say you put in a bunch of squiggly graphs of different sensors Do you put in the metadata of where each sensor was or what its condition was, its age, et cetera, et cetera? What kind of metadata do you put in there together with that? And what do you expect to get out of that one training step that allows you to iterate?

Sure. So let me give some tiny bit of context then on that Kajima project and map that to what the input and the output of the model looks like, right? So in that case, we're working with Kajima, very established Japanese construction company. They build islands, they move mountains. The project we were working with them on was literally taking a river and moving the river so that it wouldn't flood. So it was a 5-year-long project. They had sensors the whole way down the river as they kind of moved down the river and they're dredging it and they're widening it. And they're basically mitigating the risk of flooding, right? And the big problem they wanted to solve is their company is outsourcing and subcontracting a lot of the work to a lot of other sub-teams. And they want to basically measure to make sure that those teams, if they have to go and dredge the river for 5 hours, did they actually do 5 hours of dredging that day? Or actually they only did 2 because maybe the digger was in maintenance mode that day, for example, and they couldn't do it, or because of weather conditions and others. So for that project, basically the input to Newton, our world model, was the multiple cameras around the site, typically 3 or 4 cameras looking at a specific section of the river. And when you look at those cameras, basically you have a barge, this boat that's floating out on the river. On top of it, you have a Excavator digger. And then on top of that, at the end of it, it has an end effector that is either drilling or it has the bucket on it where it's actually dredging, right? They want to know how many hours essentially that excavator is actually dredging the river, not drilling the river, not a crane, not other things, right? So essentially the input to the system is the sensor cameras the river sensor, hydrometer sensors that basically say how high the river is. And then at different geo locations along that, I think it was multiple kilometer site, where the actual work was supposed to be happening at that day and what the weather was that day. So we basically have a bunch of cameras, we have a bunch of time series sensors with weather and river information. It's fed into the model, and the model is essentially outputting what looks like a Gantt chart. It's actually outputting the state of that team. Where is the barge? Is it parked on the side of the river? Is it actually moving into position, which can take an hour sometimes to do that? Is it now actually in position and they're starting to dredge, or is it just there and it's stationary and nothing's happening? Essentially, our model is transforming the sensor raw sensor data into this business output, which is essentially the yield of the team, right? Like what the Gantt work chart was for that day. And then they can start to map these work charts over multiple years of data and start to look at these bigger patterns that a human can never spot. Like one of the patterns they find, for example, is obviously if there's a storm, The work on the day of that storm goes down. That's kind of, you know, everyone would know that, right? But one of the key things they found is multiple days after the storm, when the weather had cleaned up, the work productivity of the team was still very low. And the reason for that being is the river water coming down from the mountains, it takes several days for it to actually reach the end of the river and go to the ocean, which is where the site was. So there was a lot of debris and a lot of churn coming through the river multiple days after these larger storms, which is actually slowing the teams down quite a bit. They'd never spotted those patterns before because they could never look at these aggregate patterns, but Newton was able to kind of spot those patterns for them, right? So it was essentially able to see—

What did Newton output? Did Newton output like, you know, hey, this is happening, or did Newton output like, this is a projection of some particular sensor in the future? What was the output?

So it was twofold. First of all, it was able to take in all of the sensor data at a specific time window and actually transform that from raw sensor data into basically the state of the team. You know, was the barge parked? Was it dredging? Was it moving? Was it drilling? And so forth, right? So basically, for each— for over 5 years of data, for every timestamp at each location, you basically got the productivity of that kind of geospatial cell. That was then placed back into Newton again, and now you can start to query Newton and start to ask it questions about 5 years of this transformed data to start to look at these other bigger meta patterns. In that case, we can actually feed the data back into Newton again and say things like, help me explain key areas when the productivity was either highest or lowest, for example. You can say, what are some of the key areas of the the top-performing teams, or what is like that 5% of lowest performance, and how do we then start to pull patterns out of that to kind of feed that back in to kind of lift the team back up again? Like, for example, maybe 2 days after a storm, you should actually go and work on a different part of the site because you know the productivity of the river will be low that day, for instance.

So I have a few questions about how this kind of fits into the AI big picture. One, when Prakash said predicting the future state of sensors, that obviously relates to the big question around world models. Do we need predictive world models that, for example, predict the world so that a robot can interact with it smoothly? A humanoid? We have that debate a lot.

Yeah.

But in your area of these industrial processes, Do you feel that you need that? Like, I guess it— and what would determine it? It feels like it might be a question of kind of response time where like if you have a predicted thing and then what's observed is different from predicted, you know, that might set off an alarm. Is that the kind of direction you see this going? Or maybe you don't need that rapid response in some of these places. I don't know.

So yeah, for our customer use cases, we definitely need that type of even, you know, sub-second response. The 2 examples I can give you briefly are, first is in the case of safety, for example, you can use the prediction of what might happen in the next few seconds to actually predict the safety event before it happens and then trigger it before that person, you know, on the construction site, for example, that would be a key area to actually monitor area, monitor workers, and if they start to get into those situations that are higher risk, you can actually use this predictive capability to be like, oh, something happened that surprised the model. It didn't match the predictions. And that in itself can be an indication that something is wrong. Or you can actually predict the safety event itself, and then it's high enough confidence that you can then cause that trigger before the safety violation happens, right? The second key one for real-time systems is we are using this, you know, these kind of world model latent spaces. not necessarily to do prediction on the sensors, but actually to do the control prediction. So this allows us to go from the sensor the whole way back to do the closed-loop control. So actually, how do you use these latent spaces to actually predict what the control of maybe the coolers should be in a large environment based on the input sensors, right? So these world model latent spaces and doing prediction on them have many practical Being able to do it in real time, at least for a subset of our customers, is very important for safety reasons, for real-time control reasons. Being able to do it on the edge as close to the sensor as possible, low latency, is definitely a key use case we see from our customers.

How much does Newton generalize in the sense that when you go into a new site, do you have to take all of the data and kind of fine-tune the model once more?

Very good question. We are building the model so that out of the box you can plug in different sensors and different configurations and the model can automatically adapt to those as much zero-shot as we can. For a subset of users, they can directly do that. The typical analogy I give customers is if you showed this task to an average person from the street, would they be able to detect the safety violation or would they be able to detect that looks like a strange signal and anomaly? If the answer is normally yes, normally that means that Newton can do it out of the box. It doesn't need, you know, a specific, you know, master technician's level of expertise yet to be able to spot that, right? For those cases where maybe the piece of equipment the user is using is very specific to, you know, that factory, for instance, or they have certain parts that, you know, they, they have very, um, internal and, you know, uh, unique names for the— for those processes or parts. That's typically where fine-tuning or being able to bring the customer's knowledge base and actually being able to inject that securely into Newton really helps. So in that case, our platform allows customers to take our model, run it directly on the customer's infrastructure so no data is leaking out of their networks, either on-site, on-prem, or in their secure clouds, and then be able to take their own data and actually do some amount of fine-tuning to be able to customize the And typically with customers, we've seen on the order of a few thousand samples is enough to be able to take the general model and then adapt it very quickly to that customer's use case.

So would you say that means that for a single sensor, Newton's able to detect the anomaly for that sensor, but when you need— when you have a customer job site which has multiple kinds of sensors all over the place, The combined effect of all of those sensors in terms of determining the state of the job site needs to be tuned for the model. Would you say that?

I would focus it more on maybe the use case than the sensors themselves. So we're trying to generalize the sensors as much as possible. It's more what is the use case? What is the output of the model? Like, what is that? Is there some very nuanced domain knowledge the model needs to be able to have to solve that problem? problem, right? We're trying to solve the generalization fusion of multiple sensors as much as we can for all customers, right? And that's really where out-of-the-box performance is— we're really trying to lift. Of course, if a customer comes with a very niche type of sensor that's not on the internet, that's never been seen before, then that's where we would need to help them not just do fine-tuning but really like do the pre-training of an encoder that helps bring that into the common The common model.

Yeah. When you think about the impact of all of this, obviously another big question in AI is, for all the amazing stuff that we see, why aren't we seeing headline economic numbers change? When you talked about safety, I was kind of like, well, sometimes the best enhancements are GDP destroying in the sense that When you have an accident at a site, that potentially causes a lot of money to flow around and the measured economic activity goes up, people go to hospitals and whatnot. So you don't want that to happen. So that aspect of it could be GDP destroying.

How do you—

and that could be great. Sometimes GDP is not good. But how do you think about kind of, does this help your customers make more money, do the same thing with less, do the same thing but with less downside problems? What's that mix look like? And what does that tell you about the future of AI's impact on headline economic numbers?

Yeah, we definitely see customers coming from different angles of that spectrum. I would say for us, probably one of the biggest deltas that we are seeing is while other physical AI companies have to build the robots and then train the robots and then deploy them, We're working with customers who already have a large amount of this sensor control infrastructure already in the world. We're able to come in and typically the main problem we're helping solve with this intelligence layer in the middle of the sandwich is how to help customers solve that problem in the intelligence space that they've been trying to solve for 10 years with other techniques. They just haven't been able to generalize it and make it scale. Or maybe it's taken them 12 or 18 months for a team of PhD data scientists to build some custom solution that maybe only works for one of their sites or one of their wells or one part of the grid. And what we're able to do is come in and really help customers to very quickly, in 10x the time, be able to come in and solve that problem and actually be able to then deploy it in the field and actually start to see that impact either in the safety area or helping them improve yield or One of the big areas that we see is how to bring in a lot of the expert knowledge that's in there, in there, kind of the, you know, the master technicians that might be on the factory floor, for example. How do you capture that and maintain that as a company and then kind of lift up all of the operators in the company with that knowledge? Because you can kind of communicate that knowledge to Newton and then that can actually help other, you know, new technicians that are joining the company. it helps to explain to them why this anomaly maybe have triggered an anomaly, for example, because it actually can lean on the prior input from the master technician, right? So one of the value adds there is actually how to basically take your team and kind of lift up the knowledge base of the whole company, for example, right? So that obviously directly maps back to cost saving and, you know, being able to really take these very large, complex, expensive machines, for example, and keep them in the right maintenance windows because you don't want that million-dollar turbine blade to go off at the wrong time, for example, and be out of action for 3 weeks. Right. Yeah. So we see folks coming from all areas of the spectrum. It could be those that want to kind of help reduce their costs.

It's those that want to help boost safety areas and key things like this here, or it's those that want to capture this kind of knowledge base and being able to bring that back into the rest of the It sounds as though you're capturing a lot of tacit knowledge that is often embedded in subject matter experts, and which I think a lot of, a lot of the frontier labs, for example, are kind of going into right now in terms of trying to absorb legal and financial and etc. expertise.

Yeah, yeah.

It also sounds to me like mechanical engineering specifically, there's a lot of stuff in the real world that only people who operate it understand, right? It's very different from, I think, a lot of electrical engineering or computer science in terms of vibration and friction are often things that, and heat are often things that change the behavior of systems.

Yes.

Quite dramatically. And especially, and also you have the maintenance curve as well over time, which degrades performance, et cetera. So to what extent do you think capturing all of this tacit knowledge will help design better in the future? Can you use your systems to project and simulate so that, you know, you have, you can optimize for, let's say, a longer lifespan or cost savings?

Mm-hmm.

Or strength or efficiency or any one of these axes that you need to design around or design for.

Yeah, very much so. One of the things I'm most excited about in the physical AI space is allowing customers to go from solving a specific niche problem on one machine that they've been trying to solve for the last 10 years, and it's just the physical world is so noisy, it's just so hard to solve it. Not just being able to solve that quickly, but being able to take all of the machines in an entire ecosystem and an entire factory, an entire energy grid, an entire city, whatever it may be, and connect those together into an AI system that can now start to see those macro trends across the machines, right? Where now you can actually start to see the reason that one part of the machine in that one part of the factory is actually operating faster is actually because of all of these knock-on derivative effects from like 3 machines down, right? Or because it's like that thing is near the loading dock and then you have all these strange thermal currents that actually move around the building and so on and so forth. So being able to not just solve these kind of micro problems on individual machines but start to look at these bigger ecosystems, and I think that's where being able to take that expert domain knowledge from humans that's never going to go away and the sensors will never be able to capture all of it. You'll always need to bring in that human knowledge, that human context. But being able to allow the human to spot those larger patterns that maybe a single person can't spot across an entire factory or maybe multiple sites at different parts of the country, for example, being able to actually start to spot those macro patterns that maybe occur over either geographic spans that a human just physically can't capture or maybe temporal spans, maybe it's many, many months of trends before you actually spot the pattern, right? So that for me is one of the really exciting directions that these very powerful AI models are starting to unlock. It's like, how do you actually go and solve— like, discover these unknown, uh, effects in these ecosystems, then map back to these very specific human understandings that get to kind of the mechanical engineering details of thermals and heat transfer and everything else that actually happens in the, in In the limit, right?

As your systems get better, often for mechanical engineering, you have kind of the number of duty hours that a particular component can have, and you need to do a block maintenance every X months, et cetera, et cetera. So you have all of these rules and maintenance schedules, et cetera. And many of these are designed around allowances, tolerances for like 95th percentile, how many failures are we going to get? And we multiply that across the machine, et cetera. As your sensors get better, do you think we can do away with some of these? Like, as the projections get better, can we do away with some of these? And instead of using this kind of like statistical analysis, use kind of projections or simulations of when things need to occur and basically have better tolerances because We just know more.

Yeah. I mean, the thing I'm very excited about is there's always this sim-to-real gap that has plagued various areas of machine learning for decades. The thing that I'm really excited about is being able to, as each part of that loop between the core models, the simulation, the deployment to the world, the understanding and analysis of maybe failures, as we close that loop, And as we get more and more details and each— the models improve at each part, being able to actually maybe not get away with, you know, issues of, you know, yield confidence and things like this here, because there's always going to be, you know, there's always going to be damage in the world. There's always friction. You know, there's always cases where these machines are going to break, but be able to close that loop and have, you know, these models now start to understand each part of that system and start to get this feedback loop where the sim informs the deployment, the deployment informs the analysis, the analysis informs the sim, and you can kind of start to really, by actually getting these models into the physical world and then feeding that back again into the model so it has more time to think. This is exactly where this can go in the future, right? So our initial goal is to get the models deployed and have them bring the best intelligence to the edge as possible. But by closing that loop again, there's some pretty exciting things that can be done here to push this forward.

Last question for me is about your vision of superintelligence and how you might relate to the frontier companies as we go forward in time. I'll just give you my kind of poor man's understanding of superintelligence. You can react to it. Basically what I think is, geez, the reasoning is already getting to, like, super intelligent light with all these Millennium Prize problem style results happening just through super long chain of thought. Like, who would have believed you could get there? So on the reasoning front, we've kind of come awfully far. What we don't seem to have still is a good intuition for the many domains that matter. And what I sort of expect to happen is something similar to what has happened with images, where It's one thing to sort of put an image into a model and get a caption out, but when you can get the altered image out and it's clear that the model understood both your verbal instructions and also sees the image in a very similar way to the way we intuitively see the image, which is something we are really good at, then it feels like, wow, okay, it's really got that domain. So my superintelligence vision is basically do that across a tonne of different domains, many of which we don't have native senses for, with what you're building being a great example of that, right? People don't have a native sense for how to integrate 10 signals from 10 different sensors. So to put that in the form of a question, do you see a role for you in superintelligence being sort of building the deep intuition for this domain? And does that ultimately get joined with elite reasoners via the form of a bidding war between a couple of frontier companies for your company, for example?

I definitely see the part that we can play in a larger superintelligence system. I think the key thing there is these systems may be able to edit images and do all this stuff, but to actually bring that back to the physical world and then actually move that object in the physical world is, you know, it's still a very— it's a very different story, right? So I think, you know, the thing we are really focused on as a company is how do we bridge this gap between perceiving in the world, reasoning, and then controlling, uh, you know, bringing that back to— back to the world again, right? So, um, I think what we're building at Archetype really complements a lot of the— the digital AI that people are trying to solve. Um, by doing that, it won't automatically solve physical intelligence, right? It won't solve problems in the physical world. It will definitely accelerate this and make it much faster. But we're looking at how can we essentially be one of the key nodes that plays in this, that understands the physical world, can reason and perceive in the physical world, can deploy out to many, many small edge boxes where you don't need an entire data center to kind of, you know, do one reasoning loop, and be able to connect that back again to kind of this almost System 2 level of thought that then can actually happen in the digital world and how to bridge that gap. And we're seeing this with other customers too. You mentioned before, like, you know, being able to go and scan your invoices and your finance and all your digital data. We're having other customers where they're using other LLMs to do that, and then they can take Newton and then use Newton to help them understand their physical world and then connect that digital and physical bridge by basically having Newton agents talk to other digital agents, right? So that's kind of how we are seeing, you know, our where we play in this kind of bigger superintelligence loop is like we're very focused on how to understand the world around us and be able to bring that back to natural language and machines so that these things can talk to each other. If we can do that and other folks are solving digital intelligence, that nicely will connect together very well.

So just to make sure I understand your answer there, it sounds like you are envisioning a world where the technology that you're building remains sort of a tool call type relationship to a core reasoning model and doesn't so much get upstreamed and thrown into the GPT-7 pre-training mix to spit out one model that does everything in the same latent space.

We're thinking of it more on the line of we're building physical agents that can be deployed out to the ecosystem of the world. Those physical agents then can talk to other agents and be able to bridge that gap between the digital world and the physical world. How much that's actually done on an agent-to-agent post-training mechanism, how much of that is pulled into the pre-training, we get to see, particularly when it comes to those more nuanced sensors like radar and LiDAR and infrared and all these things that don't look like human visible images. That'll be interesting to see how that's brought back into these bigger pre-training loops versus just— it's more of an agent that can go and solve that problem and tell you what's happening in the physical world in real time and be able to let you interact with it, for instance.

This has been excellent. Nick, thank you for joining us on AI in the AM. The company is Archetype AI, building AI for the physical world. Pretty cool stuff. We'll definitely be watching for updates.

Thank you, guys. Great to talk to you. Thanks very much.

Great to meet you.

See you.

Bye. Fascinating.

And now for something totally different, but also in some ways probably some fundamental similarities. It'll be interesting to find out. I'd say this is a banger episode. 2 very deep tech guests solving really important real-world problems, a bit off of the mainline paradigm of just ever more RL/VR on language models. And the upside is pretty apparent, obviously, especially on the on the biomedicine side, but also in the making sense of all the real-world sensor data that just largely gets logged and kind of, I think, mostly ignored today. I mean, monitored somewhat, but certainly not monitored nearly as much as it could be in a lot of cases.

I liked how you had basically 2 different approaches for data. You had On the one side, you had, let's take all the messy data, let's feed it into the machine and try to make sense of it. On the other, Andrej is like, there's just too much messy data in biology. I'm going to build my own data collection effort, and I'm going to monitor every single thing that goes in and comes out so that I know exactly what's going on. It's the 2 opposing poles of data. One is like, I'm going to exactly control the environment so I know what I'm getting. And the other is like, I'm going to just go out there and just grab everything possible. Exactly the 2 different approaches that people take. And I think Andrei probably has a good chance of solving this thing. You need a little bit more GPU power. You need a little bit more imaging. I think the imaging granularity is probably not there yet, but you can definitely see if they manage to scale a little bit more. The real question is, Do you need to do 3 billion samples, 30 billion samples a year? Do you need to get everyone's genetics, or are you gonna get— is your scaling gonna be able to get you to a point where at 30 or 300 million or wherever, it just starts to generalize a lot more and you stop, you stop needing so many, so many different samples?

Yeah, time will tell on that. But also the willingness to pay for that N-of-1 loop. My guess is that the threshold for N-of-1 to be viable, especially if you're willing to pay and run a lot of the actual tissue samples on that chip in the real world, I feel like it's got to be That threshold's got to be fairly close because a lot of times what people want is just like, just go through all the drugs that might be relevant and see if any of them look like they're going to help me. And the willingness to pay for that in some cases, obviously super high. So yeah, I'm very bullish on that whole approach. I kind of take the meta lesson that we've seen over and over again, that the experts are surprised by the pace of capability advance on the models and on the debate between biology is too big, we'll never get there, and if we just collect enough of this perturbation data, the models can figure it out. I definitely feel like we've got enough evidence from other domains that the latter is where I would put the smart money. Yeah, it's just like, how— what is stopping us from scaling this up faster now? This is the sort of thing that I really think the, you know, if we're going to do socialism of some sort, like, how about putting a real push behind an effort like this? It sounds incredibly viable and just incredibly awesome. I also now wonder, of course, for both of these, like, what's the purchase price at which Anthropic and/or OpenAI come calling on these guys. I think you do have— I mean, we don't know a lot about what efforts those companies have in these exact domains. We know Anthropic is building automated labs. OpenAI has obviously done incredible work in medicine. I think I have less visibility into what they might be doing in kind of more foundational biology research. But given what we see people spending, what we see these leading companies spending on hastily and shoddily vibe-coded RL environments, companies like this that have a unique data asset and some traction on making sense of it, It just feels like it's an obvious upstream. And we've seen the threshold. It seems like the reasoning models are getting good enough to do all this stuff. The ability to paint paintings in digital painting software, the CAD usage that we've seen the last couple of days has been insane. The circuit board design is another one that we've talked about several times. It's hard for me to see why the outcome here wouldn't be one of the frontier labs buys these companies, upstreams the data. The data is the most valuable part, probably. The automated lab stuff is cool too, no doubt, but data alone feels like it's worth billions. to the frontier companies. And it's unclear if it's a buyer's or seller's market. It's clear that they want to buy as many RL environments as they can. It's a little less clear how many BioFoundation model companies they're going to want to buy. But my guess is more than one. Why wouldn't it be more than one? The balance sheet is there to support all kinds of stuff. So is it really so hard to anticipate Anthropic buying like 10 BioFoundation model companies and just Captain Planet-ing them into one superstar?

So one of the confusing things for me is, would it not be better just to use the model rather than having to buy the model? For example, AlphaFold, AlphaFold 2. Okay, so AlphaFold 2 exists. Can't Claude just use AlphaFold 2? Like, why would Claude need to buy AlphaFold 2 and have the LLM reasoning framework be able to directly interpret, like, you know, and directly do protein structure modeling? Would it not just use the model which is specialized for that? Is there any reason to have Claude itself kind of do that? And this is one of the questions about generalization versus specialization. I've always felt that the specialized models do better at their specific thing, even though they don't generalize to do anything else. AlphaFold doesn't chat with you, but it does what it does, the specific function, very, very well. Now the whole, I think, superintelligence fear I've always felt has been that the superintelligence will actually do what AlphaFold 2 does well, better than AlphaFold 2 itself. And that is one way where you don't get any more specialized models and you don't have this dependency structure outside the superintelligence itself, which makes it dependent on other parties. If it internalizes everything, And it did everything at better energy, at better compute efficiency than everything else, then there's no reason for it to externalize anything else. But if, for me, I think the natural thing is always going to be that something which is specialized to do that particular thing at the lowest possible power usage will do that better than the generalist model, which is a generous model. And I think these are the two things, I think that the two kind of axes that I think differentiate probably the people who believe in superintelligence and the people who are a little bit more doomery is that they believe that the model will actually do everything else better and at better compute ratios than specific models. So no dependency structure means you're stuck, you're in trouble.

Yeah. I mean, I think there's also a sort of synthesis of that, that is like, what we seem to be doing now is in many domains, not all, it's definitely still like, as well as Astro might navigate a car around a parking lot, the dedicated FSD models are clearly a lot better. But it does seem like we are seeing this pattern over and over again of data gets upstreamed, the mega model gets better in the image case, right? Claude can call an image generation model with a prompt.

Yeah.

And it's odd that Anthropic hasn't built an image generation model into Claude in the same way that obviously the GPT side has this native image understanding and editing ability. But what Claude can do with calling an image model just isn't as cool, isn't as high fidelity, doesn't give you the same sort of like, oh, wow, you just transformed that image, but you kept the kind of core of it that you get from the proper image language integrated models. And so I kind of feel like the same thing is going to happen over and over again at the absolute frontier of scale. But then you might still see that the way to actually run it efficiently is some sort of distillation. That's also very on trend. So I think you could have a situation where when your N of 1 medical study is to be done in the future, it is being powered efficiently by some much smaller specialist model that sort of only does this, but learned a lot from Claude 7 that is whatever, 50 trillion parameters in the cloud that has taken in all of this industrial sensor data, tissue perturbation data, et cetera, et cetera, and just has this sort of unified model of everything from which we kind of distill out different specialists that actually work fast enough economically on the available chips, et cetera, et cetera. But yeah, put me down for one that has not seen a reason yet to believe that all these modalities don't end up integrated at the frontier. In a way, I wish it wasn't going to go that way. I do think safety through narrowness would be really nice. That's why I'm excited about Jev, in part, because it sort of It has a very niche role to play. You could build all kinds of scaffolding around it. I was a big fan of the Eric Drexler reframing superintelligence piece, which he also known as comprehensive AI services, where you've just got narrow purpose-built AIs for all kinds of different niches. My hope for safe superintelligence is something like that, just based on kind of reading the tea leaves of Ilya's comments. Could we possibly get a model that starts as kind of a stem cell generalist and then matures in a way where it gets really good at its job, but also kind of settles into that niche in the way that mature cells don't revert and turn into other kinds of cells, and that they're kind of limited now? They've lost their pluripotency, so to speak. I love all that stuff. I hope it goes that way. And for efficiency, it might. For efficiency, it probably will. But at the frontier, I just have not seen anything at all yet that makes me think that the single biggest, best teacher model you could make couldn't just learn it all. Given what we've seen, that starts to be a bit of a scary beast.

Oh, I agree. It can learn it all, but I think learning it all at the most efficient compute rate and inferring at the most efficient compute rate, which is what you alluded to with the distillation, I think that doesn't happen, right? So for me, I think the moment you start using more compute and more energy, your efficiency ratio of compute to intelligence is going down. So effectively, I feel like the model is actually getting less intelligent because it's actually using more compute. And I think the more intelligent model uses less compute for the same outcome, which is basically what has happened every generation. You see this kind of like, you get out there and then you kind of pull back and then You get to someplace with reasoning and then the next generation you can get there even without the reasoning. So I think the energy efficiency is kind of key because the energy efficiency means that you have this diminishing returns curve as in economics for getting bigger. And because you have that diminishing returns curve, that means you can't just continuously get better and also get more intelligent. You have to, like, you know, you can get better, but then they sacrifice efficiency first, and then you distill down and you increase efficiency. So I think that always makes me a little bit like, okay, the doomer scenario is less likely to happen because you will get a more intelligent model, but it's going to be less efficient to start out with, and it's going to be more energy consumptive. And that means that it's going to need more energy expansion to get there, but the other models are going to be more efficient than it and more numerous. So I think the singleton fear is that one fear. And I think the singleton thing doesn't happen. Eliezer's next thing is like, oh, they're going to collaborate with you as a swarm against you. That, okay, let's think about that a little bit more. But I think the singleton example, I think, is not likely to happen.

From your lips to God's ears, as always. I think that's a real constraint for sure. I mean, if you made a— let's just get ridiculous, right? You made a quadrillion parameter model, you certainly would find it a bit slow and a bit expensive, and you couldn't run a million agents at that scale of model today. So I do think there are some practical limits there. But then again, remember what we talked about on Wednesday? There's going to be more compute installed over the next 12 months than exists currently in the world now.

Yeah.

So we are definitely going to test for at least a couple, few more years whether just Bulking up on compute solves all, quote unquote, solves or perhaps creates all these problems. And that means another way to think about that is the companies are putting roughly, it may have changed a bit more toward inference now, but historically they've put roughly like half of their compute into research. So that would mean next year, simple model is everybody has as much compute for research as everything they've done this year, right? All the customers they've served, everything else. That definitely can take you pretty far if you have conviction. It's like a crazy thing to do, right? Remember when we talked to David Li from Shanghai and he said the reason that the Chinese companies were a bit behind on LLMs when ChatGPT or like when GPT-3 first came out, maybe when ChatGPT, that kind of changed it. But at GPT-3 scale, he was like, Chinese companies just thought that's a lot of money to spend on something that doesn't really seem to do all that much. And it took the almost ideological conviction around scaling laws for people to say, well, maybe it can't do that much yet, but we'll get there. And I do think that conviction still exists in a few places. where the scientists are much more focused on whether or not they can than whether or not they should. And by default, I think they're going to try to run absolutely mega scaling runs, and we'll see what comes out of that. Or possibly we'll have some sort of agreement to pace the frontier, and maybe we won't. But left to their own devices, I think the mad science will continue. It will scale up insanely. And the balance sheets to go get a trillion tokens worth of tissue perturbation data or a trillion tokens worth of sensor data are absolutely there to just suck these things up and dump them into the top of the training funnel with everything else. So barring some external social constraint, I think that probably is the default path still.

So on bio specifically, Claude— this news came out yesterday, day before— so Claude discovered a CRISPR-like repeating sequence in the genome. So basically what happens, CRISPR, as you know, it is a gene editing technique, There's basically a segment of nucleotides that can identify very accurately where certain parts of the genome are, especially when attached to a guide. So it can go in there. And so this is what engineering has become about, genetic engineering. And so what has happened is that CRISPR is one of those. Uh, there have been other kind of sequences also discovered of different efficacy and for different use cases. And Claude discovered yet another one by looking through bacteria. So it's looking through bacteria, uh, genomes of bacteria, and it found one where— ha! And, and, and it's, it's very funny because it's just reading, it's just reading the, the genetic, uh, code. So it's just seeing like ATGC, ATGC, ATGC, GG, blah, blah, blah. And it says, like, whoa, look at that. There's a repeating form there. And that repeating form seems like there's some system there. There's some systematic use to it. And then it kind of told that to the scientists. And they kind of, okay, let's test out whether this is actually this kind of CRISPR-like guide that can guide, like, you know, therapies into the body, into the genome. And they tested, and it could. So it found a new key, basically, to identify sequences of the genome in order to engineer them. And it reported, and Anthropic reported. It was immediately panned, absolutely panned by biologists. And they panned it because number one, Some parts of this have been reported in literature before, and Anthropic actually identified the parts of literature that had alluded to some of this, but the experiments hadn't been run to actually, like, oh, this is actually a guide sequence that can be used for genetic engineering. Those experiments had not been run, so that was new, but the kind of, like, repeating sequence part, I think, had been identified in literature before. And so Biologists are like, okay, number one, it's kind of been in literature before. Okay, so you took it and you did the experiments and it does the editing technique, so great. We have more of them from other places. Therapeutic use, even for CRISPR, took like 20 years. So if this takes another 20 years, well, good for you. So kind of panned. And I think— Dario was like, oh, this could be an entire PhD's work, et cetera. So the question is, are they marketing it, et cetera? I think for me, one of the clear patterns that we see right now is that the first time the models make a scientific discovery, it's always a search. It's always like, I searched through some literature and look what I found, and it always gets panned. This is what happened in math. And then the next thing is they find something not in literature, but not something which no one had worked on before. It's not in literature, but it's kind of slop AI research. Like, okay, great. It's not in literature. You found something, but it's kind of useless and no one's really worked on it before. This was the Erdős. We solved the Erdős question, but hey, this Erdős question isn't popular. No one does it. And the third step was like the unit circle. And for the unit circle, it's like a problem which every single person has tried. It's like one of the basic problems which is still unsolved. Every single person who is in the field tries it, gives up, and moves on. And so that— and then we'll get to there. And then we'll get to— finally, we get to the Millennium Challenge. And then once you finish the Millennium Challenge, you'll be like, well, it's not that useful anyway. And then we'll get to, okay, maybe it's useful. So I think we got maybe 2 to 3 years more more before biologists finally like, okay, AI can do something. So right now we're still at the literature search phase for bio.

Yeah, I think I'll take the under on 3 years for sure. Probably also the under on 2. If only we could find a way to bottle up and sell some AI Coke to people, we could We can really rake it in.

Ed Zittrain has bottled up and is selling AI coke. And a pretty intelligent man who can basically ignore every prediction that he's made in the last 2 years as they get disproven. Ed Zittrain is out there and he's selling bottles of copium.

Yeah. Well, That probably won't be my business model of the future, but it's proven to be a remarkably sustainable one despite all evidence to the contrary. Yeah, I don't know. I mean, the whole biologists panning this thing, it feels like missing the forest for the trees, really, again. And that's the other pattern that is kind of hard to mistake. The other thing too is on the useful front, I think we're going to see probably a reversal of the order of utility and Millennium Prize solutions in biology relative to what we've seen in math. The issue with math is that we already had a tonne of math that was interesting to mathematicians but not super useful. And so to push the frontiers of math, you kind of had to get into the not useful domain of math. In biology, we have tons of useful stuff that is still not even that well understood, right? I mean, it's just very— it's a far more empirical than theoretical domain. And I would expect a tremendous amount of utility to emerge before we have anything approaching a full mechanistic or fully causal understanding of what's going on in biology. And to some degree, obviously we'd love to have that, but we don't necessarily care at the level of like, does this cancer drug work on this patient's tissue? If it works, you don't have to have a mechanism, right? That's not part of the clinical trial approval process. It's nice to have, not required. So, yeah, the other thing I would just Maybe my last comment on this bio story is this is exactly what Dario set out in Machines of Love and Grace. His core argument there around bio is it advances in kind of step function form when these new general purpose tools that make biology more, let's say, just programmable, for lack of a better term, when they are discovered, they unlock just huge new potential in experiments you can run that you couldn't run before. I think that is not yet, as far as I know, demonstrated for this new DNA editing system that they've found. Does it do more than other things? Does it do different things than those other things have done? I don't think that's been mapped out. But the general strategy that he outlines is we want to find these foundational technologies, and then we'll run experiments like hell on top of them to map out the space. And so this may or may not be one that really moves the needle as to what experiments can be run. I don't know if that's clear yet, but the fact that they're starting to find even good candidates for such systems, I think it's extremely telling as to where we're headed in the not too distant future.

I wonder how not too distant is not too distant, because I think they have to show something within the next 12 months. And I think biologists, bio people are very hard to convince. So it has to be something significant. I don't think an AlphaFold 2 level discovery is enough, is sufficient at this point, because AlphaFold 2— AlphaFold took maybe 5 years plus to kind of filter through and for people to start using it. And it kind of displaced a lot of chemists, actually. It displaced a lot of people who are doing on structure, because typically one protein folding structure is like one PhD. A couple of years of work, and it displaced a bunch of these people. And I'm not also certain that there is— there has been, or there is clearly benefit yet from AlphaFold. Yes, it has made some things easier, but those things were not perhaps the bottlenecks to getting drugs out there. So it has made, you know, some costs cheaper, but the The costs have shifted and the delays have shifted to other parts of the stack. So I think bio is going to be pretty tough. And I wonder what the discovery is going to be that is going to be so interesting. So I have a list of— so I tried to build a tech tree of the next 10 years. So I have a list of like 60 different fields and what experts think are the big developments that they're going to need or they're hoping for in the next 10 years. So kind of millennium challenge for each field, for 60 different fields, including pharmacology, et cetera, et cetera. And that's where I pulled up the Vivodyne stuff from. And I think some of the people that we've spoken to on the show are actually on path to kind of solve some of these. Like, the Vivodyne guy is on path to solve, I think, the organoid simulation challenge. So I wonder to what extent many of these actually do get solved in this kind of 5 to 10 year period, and perhaps even sooner than that. And we'll see the results, which will be very exciting, I think. Very, very exciting.

Yeah. I think the other comment from Dario recently that, if nothing else, suggests that they get it is, I forget exactly how he phrased it recently, but it was like, talking about curing cancer isn't going to help. What will actually help is curing cancer. And I do think they have clarity there on how high the bar is in terms of, I mean, certainly convincing biologists, right? But again, that's not even necessarily the most relevant bar. You could have your holdouts, who claim it's not impressive somehow, even as you go on saving people's lives. But it is clear that they do understand that life-saving breakthroughs are the standard by which they should be measured, given all the expectations that they've set.

And which is also this kind of switch of the chase from artificial intelligence, superintelligence. Stop trying to displace people from jobs and start trying to actually create meaningful progress in the world. It's kind of like, do you grow the pie or do you just drink someone else's milkshake? And until now, I think the growth— and this is also why I think the growth, GDP growth has not really picked up that much because all of what's being done is really like drink each other's milkshakes. And try to scramble for the same group of customers, for the same pie. And I think they need to start showing that they are able to, and also they've kind of maxed out on that. There isn't really that much more that you can grow from just fighting over a zero-sum game. You have to, at this point, expand the pie, expand the set of consumer choices that are available, expand the set of the ability of consumers to actually enjoy those benefits. And I think it's a tough challenge for them, but I don't think the frontier labs can actually get that much bigger just from this kind of adversarial zero-sum game. I don't think it works out. At least politically, it doesn't work out. I think the SaaS companies have pushed back.

And I think I think I'm a little more bullish on their ability to grow even without paradigm changes than that sounds like you are, just because I still think GDP destruction is great in many cases, right? If you can get the same thing for less, or if you can get the same thing from AI at 3% the cost versus what you used to have to go see an actual doctor for, bring in a lawyer for, whatever. I think there's just a lot of incentive for people to use those tools for efficiency because the competitive dynamics across the economy are such that how are you going to not embrace these cost-saving tools? And clearly people haven't done nearly as much as they could. So this is kind of my argument that we could pause for 12 months and the companies would be fine. Just that there's still enough diffusion that hasn't happened at the current capability level, which is just maybe automating tasks that already get done, but doing it faster, cheaper, and roughly as well. I think that can sustain them for another 12 months. Again, I think this is something we're not going to find out because I think they will. In fact, as we see with just these CAD usage things, right? Like, you can design products with a prompt now at a pretty high level. So consumer choice, I do think— I don't see any reason we're not already at the point where consumer choice can't explode in all kinds of little niches where just who was going to design that product before, who's going to make that circuit board? Now it's like, well, maybe the AIs will do it. I think we're kind of there already. I don't think we're going to actually answer this question of If you just automate tasks people already were going to do, is that enough? But for the record, I do think it would be counterfactually.

I think you're right. I think Jamie Sevilla from Epoch, he puts it at 6 to 12 months. And basically the technology is there, but it hasn't diffused. It's going to take 6 months for for most normal people to like, you know, start to do agentic work at— in the workplace and just basically like, I just had my agent do the work and, you know, I just like reviewed it and the agent did all of the work, right? I think we're about 6 months before normies in the workplace, in white-collar workplaces start doing that. And I think we're probably— I think Muse is kind of the start of the consumer wave. And so I think that's gonna be exciting. And on that note, in terms of excitement, I think we are back next week with a new slate of guests and also a new look and feel. I'm very excited. So it's going to be the Opus 5.5 look and feel. Opus is very good at design. So I'm very excited. And so on that note, we will leave you. Thank you.

Yeah, Monday and Wednesday, it's going to be another good lineup of guests, pretty diverse topics. We've got another GPU data provider to try to figure out what's going on there. I thought this week's conversation in that domain was quite good with Wayne from Orin. I'm looking forward to the next one as well. Then we're going to have friend of the show Jeremy Harris on to talk about US-China. They just put out a big report just ahead of the Trump-Xi summit. Oddly, I haven't really seen much analysis or takeaway from that yet. I've been kind of waiting to see somebody put a good analysis forward, so we won't have to wait any longer than Monday, I can promise. And then we will also have author Joel Borgen, who wrote this kind of semi-utopian, semi-dystopian novel that I'm going to be presenting a few chapters of as a podcast episode this weekend. We'll talk to him about his process and how difficult it is to envision a proper utopia or to try to set a story in one. I think that'll be a fascinating conversation. And then finally, another bio, AI for bio startup focused on understanding genomic data, which obviously has exploded and is again dramatically underutilised relative to what we might dream of in the not too distant future. So Monday and Wednesday next week, a little bit of a short week because I'm traveling at the end of the week, but 4 good conversations on tap. Looking forward to them.

Indeed. Good morning, Nathan.

Thanks, Prakash. Talk to you soon.