Fixing American Science Funding

Do you mind, just because I haven't fully tested that mic before, do you mind just telling me what you did yesterday in a semi?

Yesterday I was in Healdsburg, California, and I was having some meetings and drinking some coffee and walking.

You did not drink coffee.

Seeing beautiful California landscape.

So you guys were doing the Edge Esmeralda.

That's right.

Tell me, can you scoot in a little bit? Is that okay? Yep. And then, yeah, sorry, and tell me what you had for breakfast.

I had delicious hash browns at the place next door. I wanted to have a gluten-free crepe, but, uh, Anastasia had some orange juice and, uh, what was your first car? Uh, gray Volvo.

How long you've been driving?

Uh, only a couple years because it broke down on the highway.

We could switch too if that would make that easier.

If that's okay. Sorry guys.

Nope, we're all good.

Let's see. Okay, so got it.

All right, we've done it.

Do you like my gray Volvo too?

I've never seen your gray Volvo.

Okay, cool. Okay, you guys feel good?

That sounds great. All right.

I'm going to do, um, I'll do like a little bit of an intro and make it awkward. So you have to sit there listening to me for a second. Okay, here we go. Adam Marblestone. Anastasia Gamik, thank you for joining me. How are you guys?

I'm great, doing really well.

Doing well. You've been at Edge Esmeralda. For people who have no idea what that is, could you give us like the 1-minute rundown?

It's a pop-up village in which they did not actually build a village, it is just a normal city, um, where a bunch of highly agentic tech weirdos move to be each other's neighbors and hang out for a month and, like, do projects. And everything from— have— we just did normal meetings there. People built solar water filtration systems. Someone built a solar A-frame. Talks. Audrey Tang was there, which is really exciting. Uh, and mostly I think the point is that you're walking down the street and you run into these other people who are building and doing very cool things.

Yeah, I've been hearing about this for a couple of years and now I need to go. I think I'm going to— I will try to make it this year if I can. Okay. To give people some context, you both co-founded and run Convergent Research. We will get into what it does, but, but people are going to be hearing a lot about the concept of a FRO, a focused research organization, and what that means to science. I am going to take you out of FROLand at different times as well. Adam, I learned things about you last night that I didn't know. You were a gymnast, I think, as a young person.

You found that out?

Yes, found that out. Became a scientist, worked at Ed Boyden's lab for a while, worked at Google DeepMind, worked at Kernel with Brian Johnson, a little sort of Neuralink sort of consulting stint or something in there.

Very early.

Yeah. And then now have gone into Well, you're just always a super fascinating person to talk to about all things science.

Looking forward to that.

Yeah. And are pursuing it now. Anastasia studied at UCLA and was not like directly neuroscience or something at the beginning, I don't think.

No, I studied astrophysics.

Right.

And then I learned I don't like math. Okay.

And you were very early at Neuralink and helped them kind of set up. Get going. And, and then you worked at a company that I covered— I'm pretty sure I wrote the first story on ever— which was Momentum Machines, which became Creator. Creator. And they were making a robot for like burger making.

And we made robots that make cheeseburgers. Yeah, highly impactful work.

It was actually awesome. I loved it. Like, who doesn't love cheeseburgers? I mean, it was cool and it worked, but it sort of flamed out, I think, in the end. And then very famously, I think for a lot of people during COVID when COVID was just first kicking off, we're at Curative, which became really well known for actually like figuring out testing in a hurry and doing it at a massive, massive scale. And I think you were— my sense when I was doing my research is I've known you, but like just digging in, I mean, you seem like an operations whiz because you seem to get things off the ground.

Yeah, I was in the DRC and someone called me as I was flying home, and I built their supply chain from like 500 test kits a week to a million test kits a week in about 3 months.

3 months, which like no government organization seemed remotely capable of doing. So well done. And then, and then you both obviously ended up, um, working on, on Convergent together. Um, okay, so that's the intro. I will briefly thank E1 Ventures for making this podcast possible. Um, they've been a big supporter of the show. Thank you. And then off we go. There's a ton of things I want to ask you guys about. I think we should— I feel bad because you have to explain Froze all the time and you've had to do this for years, but I do think we should give the audience like just some context on what this idea is. I know you've done it 100 times, but there are, there are people really who still don't really know this concept. Yeah. Who wants to take a crack?

Oh, I get to go first.

Sure. Okay.

So there's this notion that maybe we prematurely standardized on how we do R&D in the US. There are academic labs, there is industry and startups, maybe there's DARPA, and that's basically how we do science and technology. There are a handful of problems that are not easily solved within those institutions, things that need a multidisciplinary team working together over a certain length of time towards a very focused goal, where there isn't necessarily a for-profit mechanism on the other side, and it doesn't necessarily fit into publishing or academic incentives. And so Adam and Sam Rodriguez came up with this concept called a focus research organization, which is special-built to solve just that problem. If you identify a fundamental capability that we should build, just spin up a new organization only to tackle that goal.

And these are like big things that could crack fields wide open and really advance them in new ways where— and I mean, you just said it, but sometimes a company might not be willing to spend the money because you can't tell when you're going to make money from this thing. A university, it might be too much money for them to like really go all in on this wild idea and see where it goes. And so— and then you put them on— you do try to contain it. I mean, they get a fair bit of money, but you put kind of this 5-year time limit around it, I'm sure, so that the— these academics aren't just sitting there forever, or engineers trying to make everything absolutely perfect, but, but to get these ideas off the ground and, and, um, and make more ideas.

Yeah, something like, we know the thing we want to build going in, we're pretty convinced we can do it in 5 to 6 years with $30 to $50 million, and being at early-stage startups, there is like an existential forcing function to get there that there isn't necessarily at other larger institutions. So mimicking that kind of existential, I have to build this thing, I have to ship a product, I think is what we're trying to do with this time-limited nature.

And I don't know how sensitive this is. I hope it's not too sensitive, but I know, I mean, Eric Schmidt, I think very publicly been a big backer. James Fickle, I see on a lot of the companies that you guys are backing. Thank you to James. Helper to many people. And so people who are interested in science, seeing science flourish, seeing big ideas flourish, are willing to put their money behind this. I've written about a few of these companies and the ones that I've gravitated towards are in the brain neuroscience area. So E11 is trying to build a more detailed, richer map of the brain by, by using viruses to, to put colors on neurons that, so we can trace their paths. Um, Forest Neuro has a brain implant using, I think, ultrasound to see a bit deeper into the brain than, than some of the current implants from Neuralink and others. And then I know I visited you guys up near Boston and, um, there's other stuff around what, like enzymes and, and, uh, I don't know, I get focused on the brain stuff. What is the other like scope of what you guys are interested in?

Yeah, often, I mean, there's a broader frame which we can come back to, which is sort of what are the most important missing fundamental capabilities. We think a lot about the tech tree, but basically, often we're making a tool system or dataset which catalytically makes downstream work in some area dramatically easier or faster. So one of the projects that you saw in Boston was called PTI. It's on proteomics. So proteins are kind of like the little machines in cells that do most of the work. Most of our technology in biology focuses on the genome, nucleic acids, so DNA or RNA. And so this is a project that's trying to make proteomics about 100 to 1,000 times cheaper per sample. So that way you could do many, many samples, which you might want to do if, for example, if each sample is a single cell, you might want to look at many different single cells to understand how they're different. And so, so far PTI has increased the cost performance of proteomics by a factor of 27. And they're hoping over the 5 years to do 100 or to 1,000.

And it's super cool. And I'm forgetting the exact name of it, but you just put this map of unsolved problems. Is that the—

Of gaps.

Of gaps.

We have a gap map.

The gap map, which is incredible. It's on your website, Convergent Research. It's really cool. I mean, you very specifically outline in all these different fields what needs to be done. I would like fully recommend that people go check it out. It just got my mind. It gave me like just free story ideas. Bits really of people to go explore. And so, and when you list that out, I mean, you are doing research either adjacent or in some of those areas, but this was just stuff that you guys wished people would also tackle.

Yeah, it's interesting. It exists across all fields. I think the gap map is sort of a digest of what we've downloaded by talking to hundreds and hundreds of scientists across different fields, scientists and engineers. I mean, at a high, at the highest level, I think what makes our approach unique is actually something that it seems fairly obvious, but it is just the recognition that we have capabilities, like let's say making computers, right? As humans making computers, those capabilities depend on more fundamental capabilities. So this is the idea of a technology tree, but I really think of it as a technology or technology and science tree. So the ability to make computer chips depends on the ability to process silicon and understand solid-state physics and doing Those things depend on being able to see the atomic structure of materials and things like that. And so as you kind of go further and further back towards, if you want, the root or the trunk of this technology tree or technology and science tree of more and more fundamental capabilities, our everyday products and services are sort of the leaves of this tree, right? As you go further back in that tree, our main like hypothesis is that you can actually just identify what the missing fundamental capabilities are that would be kind of thick branches on this tree. Whereas most research is kind of, it's more just thinking forward, right? It's like, what would be the next interesting thing to do? What would be the next curiosity-driven thing to do? Or it's any very economically driven, what would be the next product or service that I can make that I can get money for? We just say, okay, well, let's look at these different missing spots on the tech tree where we could sprout a new branch, those more fundamental capabilities. We think these are identifiable by like roadmapping, but roadmapping also just means going and talking to people and saying, what would be the capabilities that would change your fields if you somehow could get the money and the team and the organization and the operations to build that capability? And so the gap map is kind of just— we've gone and asked lots of people that question, and that's a way of organizing it by thinking about research gaps. But really the capabilities is kind of the essence of it.

Sometimes people ask us, how— what tools or processes we use to do road mapping. Um, and I will say that is, uh, kindness and networking, um, and asking different questions.

Did you say kindness and networking? Okay, I thought—

yeah, yes. Okay, I really think that like there's an extent to which Adam has traversed a very large network of scientists very generously and asked them a different set of questions than they normally get asked. Instead of saying like, oh, what are you doing next, what is cool for you, what are you interested in, how does work, he's like, well, what is the fundamental thing you wish could be built? What would happen in a different system? What is missing from this field? And oftentimes people aren't asking themselves this question, and so they don't really have an answer. But then they go away and they think about it, and they're like, oh, actually, this is like, if my lab works differently, or if the system worked differently, this is what I wish existed. This would allow me to do all of these downline things. And by going and asking that over and over and over again we've collected information that then we put out via the gap map. I don't think we knew that we were building that in the first place. Adam was talking to cool people about cool things, and it came out as this map of fundamental capabilities, um, that then would solve these gaps in science and technology.

And yeah, it's just cool to have somebody focused on doing something like that. It's so— it's funny, just as you were talking about the gap map and I hadn't thought of this when I first saw it. I remember when I was working on the— this must have been like 2014 or 20— somewhere vaguely around there. There was this guy, Jonathan Hubner. He published this paper that was like kind of controversial, I guess, at the time. But also I felt like people were mostly agreeing with it. He was arguing that most of the big ideas had already been found. And Peter Thiel really jumped on this at the time and was like, yeah, he's right. And people were depressed. And then now I'm just bubbling through my head. I mean, it feels like we're in— maybe it's real, maybe it's not. This is what I want to ask you. I feel like we're in such a different place because even if you just look at AI, everyone's like, oh, this is going to unlock all these things. With the gap map, there's very clear things, specific stuff that you're pointing to that could open up new fields. The FRO idea in and of itself is sort of—

I've never really grokked fully the— I think you can talk about productivity of different systems of science. I do think that this institutional question is key. And the question of what sorts of activities can you actually do in order to unlock new ideas. So, and sometimes it's not really about an actually new idea, if that makes sense, or that's not the limiting factor, right? So there's much research that is limited by new ideas. I think a lot of academic research is limited by new ideas and creativity, you know, maybe fundamental math research. We can come back to that. We have a fro related to math, but a lot of fundamental math research, yeah, maybe related to ideas. But if you think about the Hubble Space Telescope, for example, It's kind of obvious that you, if you want to boost the idea-generating capacity of astronomy, you need a bigger mirror in space that collects more light. And, and that, you know, by that time you could have a CCD camera and so on, and you could beam it back to Earth, right? And so a big mirror in space is sort of, it's, it's, it's almost obvious that that would be a catalytic boost to the field of astronomy. But How are you going to go build that? Who's going to provide the money? Who's going to run the project? How are you going to actually get it into space? And so what I think our observation in part is that there's a lot of things like that, not necessarily as big as the Hubble Space Telescope. That's a, you know, the way astronomers do it is they actually have this decadal survey process where the astronomers all get together and they say, what are the few things that we want to ask the NSF and NASA and so on? The National Academies of Sciences coordinate some of this process. But In other fields, there isn't a decadal survey. And also the projects you might need, they're not necessarily billion-dollar megaprojects. Even getting 20 or 30 people to coordinate and say, if we had this missing capability, we could boost that. So, so yeah, I just kind of naturally just see the world in terms of these missing capabilities, some of which are actually fairly straightforward, but somehow nobody's doing them. Yeah. And so why is that? Is that institutional or financial or— yes, it's not really intellectual. There is a chicken and egg problem, though, which is interesting. I mean, so Tom Kalil, who's our board chair, often talks about, you know, you don't spend a lot of time planning for what would happen if you won the lottery, you know, exactly how would you spend the money. And similarly, if you're a scientist and you're thinking, well, to get my next grant, I need to have a project that's roughly the size and shape of a couple of graduate students and postdocs in my lab, that's the type of thing you're mostly thinking about. The missing capabilities are that size and shape, right? But if someone came to you and said, okay, well, here, think for a while about this. You can be the CEO of a 30 or 50 person, you know, engineering-focused, almost like a nonprofit deep tech startup to develop any capability that your field really needs. Then at least some people would actually come up with what would be those ideas. But they're not being asked that. But similarly, if you're a funder, And people aren't sending you those ideas because no one ever asked them. And what would be the point? You wouldn't plan how you won the lottery if no one was asking you, how would you spend the money?

So I think one version of this is even if you think about PTI, if you were a grad student or a postdoc in Nikolai's lab, you'd be like, great, what is the next piece? Like, okay, now I do more mass tags, or now I do a different type of protein or a different type of imaging. And then that's how you think about the problem. Instead of, I need to make single-cell proteomics 100 times cheaper, how do I do that? The answer involves $44 million. That's like not actually how they think about solving the problem. And so by giving them a different problem to solve, you get a different answer.

So you guys see this as just because of that, this angle of attack is very different to— if I think historically of Xerox PARC or Bell Labs or the university system or, um, company R&D labs, or even things like, like Janelia Farms, Howard Hughes. So there are places that have historically allowed for this open-ended, well-funded research. The United States obviously has funded science, you know, quite successfully in many ways for a long time. But, but it is— it's this, this intention, this angle of attack that you're doing that you feel like is different.

I think there are some similarities to what some of those organizations have done. In at least parts of their research. So Bell Labs had a sort of fundamental development track. We also like this term fundamental development. It's not just research on new ideas, but you need to really do the engineering to make it possible to use that as a capability, whether in further research or in something more commercial. Janelia, HHMI, this research campus in Virginia, some of the work that they do looks more like open-ended academic research, but some of it, like they did really great work on the fly connectome, map of the fly brain. That looks more like this directed missing capabilities development. But there's only a few such institutions. There's not really a general mechanism either for finding what would be the most important ones or for saying, okay, what if you don't work at Janelia? What if you don't work at the Allen Institute? How do you actually create a mechanism to identify and then build against one of those fundamental capabilities? Yeah.

And I don't mean this to sound as loaded as it probably will come off. But if I, as an outside observer of China, you know, it seems like they, they think a lot in some of these terms. And it's like, these are areas we care about. We care about quantum computing. We're not just going to fund one quantum computing effort. We'll fund 10, you know, on every approach. And then whatever shakes out shakes out. Am I reading Yeah, right.

Well, the cultural— what Anastasia said early on of sort of the US has prematurely standardized, and I think that, you know, Western Europe and so on maybe somewhat is in a similar mold, but the US has maybe prematurely standardized on sort of scaling up this what was originally Vannevar Bush kind of heavily university academic PI-based plus national labs and a few other commercial companies and so on mechanisms. And China has a different— I don't, I don't know in that much detail, but they do have a different set of maybe strengths and weaknesses, right? So if they want to do focused capability development, let's say for the next generation of computer hardware, new kinds of semiconductors, quantum computers, it may be easier from an incentive structure and funding perspective in their system to do that compared to ours. Yeah.

To do something like roadmap, what are the top 5 potential fundamental capabilities here? And then just decide to allocate resources towards those top 5 and see what happens. I think it's actually quite tough in the US right now.

Yes. Okay. Which is leading me to another question I want to ask. If you know, no idea actually if this gets you guys in hot water or not, because I don't fully know where all the complexity lies. But, but as we look at what's kind of been happening to science funding over the last few months, I talked to people who I think are fairly rational and actually all pro-science, and they they actually have mixed opinions. I mean, there's a whole camp of people I talk to who are big science funders who are like, you know what, the NIH is getting what it deserves. The universities are getting what they deserve. They were bloated. I know this sucks, but something's got to be done. And this is the change required. And then obviously there's other people who are just like, this is horrible. We're— this is going to have, you know, that may even be the case, but we're going to have lasting effects from what we're doing. And we're also scaring away the best and brightest people in the process. So yeah, you guys seem like very well positioned.

Maybe I'll start with a couple of things. So one thing to note is that we saw this gap, this gap in the sort of mid-scale fundamental capabilities development and identification. These, if you want, FRO-shaped problems, right? Even we started doing this back in like 2020, right? Previous administration. Yeah. This was a pre-existing problem. So thinking of this as a gap that needs to be filled or as an additional diversification of the research enterprise or ecosystem that pre-exists any of this stuff. So our original intention certainly has been to add something new, that the more mechanisms the better, because I actually do think that university research is actually really, really pivotal. There isn't, to my mind, there's no other way to do certain types of things that universities do. So imagine that your goal, imagine I told you there was a scientific field and in order to make progress in it, you need to spend a long time kind of almost cloning the thought patterns of an expert in that field while at the same time being like really strongly incentivized to create your own diverse angle on it, right? Um, think about, you know, Geoff Hinton's students, right, that went off to start these AI companies, right? Um, they kind of cloned Geoff Hinton's brain and made their own version. Yeah, that's academic research. That's what it's like to get a PhD in a research field. I think that's, that's a really important part of the system. So the FROs don't solve that. They solve this directed development of these missing, almost more engineerable capabilities. Um, but we still think that university system is really important. On the other hand, at large, I mean, except for this kind of philanthropic intervention and kind of manually going to funders and saying, hey, hey, will you, will you help us out here? Um, the system really hasn't made an FRO-type mechanism. So the change that's going on may be an instability that allows more change at the federal level. Yeah.

Okay. So, okay, that is a possibly optimistic take on things then. But writ large, I mean, you get to talk to so many scientists and you probably— I know we know a lot of the same people. So yeah, I mean, is there any sense to what is been going on, or are you— where do you fall on it?

I, I've spent a lot of time with meta-science. I've learned that there's two versions of meta-science. There's meta-science that happens by analyzing citations and describing the way that science happens, and then there's like this— Ben Reinhart called it once meta-science pirates, and Michael Nielsen calls it meta-science entrepreneurs— but people running experiments. Um, I think that it was really hard to run experiments like golden tickets at the NSF. Or lottery-based grant allocation. And I think that those things should be easier to run experiments about. How do we make decisions about allocating funding, about doubling down on funding, and the types of projects that people can work on? There are certainly a lot of people who say things like, I couldn't get my PhD work funded today that I did in the '80s or the '70s. And so I think that there is some amount of needing to encourage innovation in how science is chosen to be done. Um, I also would like a— it is— it feels amazing to spend time in the US with the best scientists in the world who all want to work together and want to come together and feel comfortable traveling in and out of the US and bringing, uh, as much talent into one dense ecosystem. And I think that it would be a real shame if we ended up losing that really dense talent ecosystem in science and technology.

Right. And then do you think there's any— just because you two do study this and these, these types of organizations, is there, is there validity to the critiques on the overhead at these universities? And, and, um, that I feel like we're Elon Musk-ing science right now, and it's kind of like, well, you guys need to clean up your act, deal with this overhead thing.

There are certainly problems. There are certainly problems in the existing system. Major inefficiencies. But I, I think you have to be really thoughtful in, in terms of surgical interventions because there's— you don't want to throw the baby out with the bathwater. In the process of doing that, a lot of scientists are really struggling right now. That doesn't mean the system is perfect, and it doesn't mean it doesn't need major catalysts for, for structural change. But I don't think just decreasing the amount is the way to get that structural change.

Yeah, I have, I have one One thing that annoys me about the word overhead is that actually often captured in an overhead number is something like the machine that everybody uses, or like the air handlers that allow you to do BSL-2 work in your lab. And that's captured in overhead, but it is essential for the project. And so I think that actually the way that we describe things, the way that billing happens, the way that even like grantsmaking happens, because you can't actually get a grant to support this like shared piece of equipment, a grant has to go towards equipment that is dedicated to the project, and so therefore shared equipment has to go into overhead. Um, like, the system makes it very opaque to answer this question, to the point where I don't know the answer to your question. Okay, because I don't understand the system.

Yeah, I just, you know, I— last week I went to talk to a group of Stanford aerospace students, and they were really sweet. They, and they just took me on a tour of the lab afterwards. It was like, oh, This professor, all his programs have been shut down. This professor, all his programs have been shut down. And then my friend— and then there was another kid there who was, uh, well, I saw the next day, who had studied at Caltech. And, um, I get— you would know better than I, and I'll try to get this right, but I guess Caltech is mostly NSF-funded, is what I was told, and, and is maybe like a little more immune to these things. But even you saying, oh, Frances Arnold— oh God, I don't know if I'm gonna get into someone in trouble. I think her— she's a Nobel Prize winner. Her lab was, was under duress. And so I came away—

multiple Nobel Prize the founders we've talked to struggle for funding. So, so I came away super depressed in a pre-existing system, and especially as a result of what's happening.

Yeah, yeah. People that I think should easily be getting funding who are doing work that will change society are currently struggling to maintain their grad students and postdocs.

So I tweeted about this because I had this visceral reaction for these two conversations right next to each other, and then, and then again, a bunch of friends of mine who are extremely pro-science were just in the comments saying, this is what they get, you know, this is what needs to happen. And, and yeah, I didn't know. I mean, that seemed to actually— I didn't tally everything up, but that seemed to be the majority opinion.

And then, yeah, I guess, um, so there's a proposal, um, from Caleb Watney, um, at the Institute for Progress that's called X-Labs, which has to do with diversifying, setting up non-traditional institutional structures. So what one type of X-Lab would look sort of like, things like the ARC Institute or, or other large institutes, one would look maybe like an individual-focused research organization, let's say, or something roughly like that. And one category would look like things like Convergent Research, which is a sort of engine for creating and operating these focused research organizations. But it might be for something other than an FRO. It might be sort of the kind of meta organizations. I think that type of proposal is a really great idea. Let's try some more experiments. Let's try funding in different structures. Um, but there's a lot of work going on in the existing structures where you have to be very careful not to throw the baby out with the bathwater. Yeah, I, I think this— the people are struggling right now. It's, it's a problem.

Yeah, there's a, a line from Hamilton which is, "Revolution is easy, but leading is harder." You gotta sing it. No, it's because I would have to rap it, actually. I think that would be bad for everybody. Um, but the notion that, like, it is one thing to, like, be like, there is real bloat in science, there are real institutional or, um, bureaucratic problems. Like, the university systems have been around for hundreds of years. Michael Nielsen and Kenjun have an amazing article on meta-science and sort of how institutions change over time. And there are issues that come from that. It just means that now you have to build something better. And I think that building is a big— consciously making decisions about what to fund, how to fund them, how to keep the momentum going is like the moment that we're in and making, making sure that someone is doing that is important.

Do you guys think when I look at how you're funded somewhere like Howard Hughes, you know, often it takes this incredibly wealthy person who happens to care about this particular thing and then amazing things happen. They, you know, Howard Hughes has been going on for a long time doing pretty awesome stuff. I think you were a member. You did the Hertz Fellowship, right? I mean, when I look at the list of people who have done that, that's what— that's like a rental car magnate from ages ago.

Relatively inexpensive, right?

And like Dario, who's the CEO of Anthropic, you know, Andy Bechtolsheim, I think, was a Hertz Fellow. I mean, you know, he invented like half the stuff the internet runs on.

And, and, um, yeah, I mean, it's a very interesting intervention, actually, the Hertz Fellowship, because it, it funds the student directly. It doesn't fund the university PI that the student works for. It just funds the student. So that means that they can then move between labs, and it also really encourages them to sort of uplevel their ambition through, through the community of Hertz Fellows. And so the Hertz Foundation is an example of just a purely additive intervention. You're not taking away anyone's anything. You're just giving these high-potential people this extra kind of wiggle room within the system. That's really powerful, right? And it's hard to know what's the selection effect versus a treatment effect, but it's really cool. The results seem to be very interesting.

Nathan Myhrvold, I think maybe it was what I urge people to go on the Wikipedia list and look at the list of the Hertz Fellows. But my, what I was kind of thinking is one of my life theses, I guess, is that, you know, we live in the age of the engineer. Most of the world's wealthiest people are tech people. Um, this has been the story of the last, maybe call it 30, 30 years. And, and clearly these tend to be people that care about science. They've been in the tech industry, and clearly they are funding some of these things. However, you know, it's, it's not— I— what I was curious about is like, yeah, do you guys feel like enough? I see, I see everybody on the Giving Pledge, they're going to give their money away when they die. Um, we— I see people griping about this one.

This one's easy.

Okay.

No, it's still, it's still not enough. We're very grateful. We have 30-something donors now. And that has been the decisive factor in being able to do this. It's interesting. It's actually, it's not, it's expensive, but not that expensive. I mean, you mentioned the Hubble Space Telescope. I don't know the exact cost of it. My guess is on the order of $10 billion plus, right? To do something like the Hubble or CERN, right?

You could probably do it cheaper.

But here, right, exactly. You could probably do it cheaper. But so if you imagine each FRO is 20 or 30 people times 5 years, something like that, that's on the order of tens of millions of dollars if you add lab space and equipment and so on. So imagine there are, let's say, 20 or 30 fields of science. There's— treat neuroscience as different from molecular biology, but don't treat Alzheimer's as different from schizophrenia within neuroscience. There's like 20 buckets. I don't know how many departments Caltech has, something on that order. And imagine for each one of those, there's like 5 5 to 10 pivotal fundamental missing capabilities, as important as just mapping the wiring of the brain is for neuroscience, like with connectomics, right? Or as big telescopes are for astronomy. And so imagine there was sort of 5 FROs times 20 fields, that's 100 FROs, each one of those a few tens of millions of dollars. Well, that's like single-digit billion dollars to like, in some sense, do all of the focused research organizations in a period of time. Time. So, you know, you need, you know, $5 to $10 billion, something like that, to kind of solve this. Yeah, um, categorically, um, that's something that's within the reach of philanthropy, but we're nowhere near that in terms of people actually committing. Yeah.

And I, I mean, if you look at the way the world is going, to me at least, you know, people have gotten extraordinarily wealthy. You see this swing, at least again, this is in my I mean, please correct it if you think differently. But when I look at like today's data centers being built by the AI companies, when I first started as a tech reporter, almost all the largest supercomputers were built by the Department of Energy or a national lab and/or a university. And now I don't think whether it's the NSA or the DOE or whoever could even remotely kind of compete with a Google or a Facebook or an OpenAI data center. I think about stuff like Tesla building like a nationwide charging network, SpaceX outflanking anything the US military or government could do. And like in my head then, you know, so there's a bit of responsibility for all this wealth that you've gained because— and then if you're going to sit there and complain, well, like China's racing ahead of us, they're doing all these things, I feel like some of this burden has shifted from the government and the university because of partly just how powerful tech has gone, where the wealth health has shifted. And yeah, I don't know, somebody has to backfill this in my mind if it's not going to be centered at the universities anymore.

Some of the big AI labs, of course, are not ultimately nonprofits or public benefit corporations. They could do a huge impact on science, right? Even the nonprofits, as could their—

But that's an interesting— I mean, you guys must think about this all the time, though. I mean, if AI really does advance in the way these people are saying, I mean, then there would be 5 companies that theoretically should be the the epicenter of scientific breakthrough and how that gets doled out, I guess.

I mean, I guess people can tap into their systems, but I think that stopping to think through what does AI— how does AI change our thesis? So we, we started in 2020, we're like, we're gonna do 2 of these, let's go. You know, it's like, how, how do I hire anyone? What is it legal? What does a contract look like? Like, it was very much so like, can we make this happen? And then it was like, well, can we repeat this? We have E11 and we have Cultivarium. Can we grow more? Now we're at 10 projects across literally math, biology, marine-based carbon removal. We have this big, big, big portfolio and we're like, great, what is the right thing to do given where we think we are on the timeline? And I think that it's really had us take a step back and think about the future. What projects aren't going to be easily obsoleted by AI? If I do something now, in 5 years, would it have been done way faster and way cheaper? What will work? What will be accelerated or will accelerate AI? I think that there's this big ground truth problem. I used to call it a data problem. Maybe it's closer to a ground truth problem. If you can get AI AI models to predict things. How do you know if they're— how do you verify that they're predicting the right things? How do you fine-tune them to the real world? You actually need to go and collect data from those systems. How do we mitigate some of the downside risks of AI? How do we build better biosecurity? What are the technical problems on that side? How do we solve the fact that we're going to have potentially amazing hackers in everyone's pockets? What are the cybersecurity systems? What are the technical problems there. And how do we build a better society given that we all have to now interact with a crazy AI-mediated world? How do we do epistemics or decision-making or maybe even some cultural things? What will it do to the economy? And so I think that it has really updated our thinking of, well, now that we have this framework of fundamental capabilities, now that we know that we can launch projects and we can do this roadmapping, how do we apply that to 2025.

And so you guys have placed these 10 bets. Are you waiting to see how those play out or are you making continued bets or especially with that context in mind?

Yeah, we're making, we're making, we're trying to make continued bets, but with a certain increased clarity of focus given just in some sense the urgency of dealing with the changes due to AI. But it even affects some of our existing projects. Again, can I give you one example? Yeah, I would love to.

I was going to ask you for it.

So one of our projects is called the Lean Theorem Prover, so the lean-focused research organization. And this is weirdly, it's kind of like the Hubble Space Telescope analogy for astronomy, but this is for math. We didn't even know there was something like a scientific infrastructure, engineering infrastructure for math, right? Because what do mathematicians do? They go on the whiteboard and they have good ideas and they talk to each other and it's not very expensive or anything. They just kind of do interesting work. So it turns out that when mathematicians mostly write proofs, so a proof basically trying to deduce some mathematical conclusion from some underlying assumptions and you write down all the steps, they're typically doing this basically still on pencil and paper. And then that gets published in an article where you can see the full typeset, math symbols, and then some other mathematician goes and they read the paper and they say, okay, do I think this is right or wrong? Have you missed a step here? Well, it turns out that with proofs, with mathematical proofs, if you write it in a certain programming language, this is the Lean programming language that this is based on, you can actually just click a button on the computer and the computer says, is the proof correct? You can just check, do all the steps add up, you know, from the assumptions to the conclusions with no missing pieces? Is it a complete proof?

This is what the FRO you're funding has, they've built that application.

Yeah, so we're building the software infrastructure of Lean. They had already started doing this, but they only had a couple people working on it as an open source project, and they now have 20-some people in an organized engineering project building out this Lean theorem prover.

And did Lean exist before?

They, they did exist before, and we made a Froh as a way to accelerate their progress.

Okay.

Um, and so now they're organized as a Froh versus a kind of ad hoc open source, um, uh, set of people. And anyway, so, so you can now verify these math proofs if you write them in Lean. This has a couple of different types of implications. So one, imagine I'm one mathematician and I want to know if I trust the other mathematician's proof. Just say, did the Lean compile, right? Check. Okay, then I know the proof is correct. So that's interesting for mathematical collaboration.

Is it like 100%?

If the proof is correct, it will tell you that it is correct. And if the proof is missing something or is wrong, it will tell you it is wrong. It doesn't necessarily tell you what you have to do to make it correct. That's where some AI would come in, is can AI suggest the next step. But okay, now you have the verifier, right? So this is a little bit like in, uh, playing the game of Go, right? Where, where DeepMind previously had big breakthroughs. Um, because you can tell in the computer quickly who won the game of Go, you can now crank out reinforcement learning on this, train AI based on this verified ground truth reward of did you— who won. Now here you can train AI on the verified ground truth reward of did you correctly prove the math theorem. So that's interesting in a few ways. So I mean, one is, can just the AI start to do math? Can you— now, if so far everything we've been able to do, uh, reinforcement learning from a ground truth verified reward, eventually you get basically superhuman if you have a good in silico, you know, in the computer you can compute the verified reward quickly. So now does that mean math, proving math theorems, including like ones no human has ever been able to prove, conjectures no one has ever been able to prove would fall to this. But so that's all interesting for math, but then now with AI, we're thinking about this in a different way, a couple of different ways. So one is like, that's a verifiable source of true reasoning, complex reasoning. Lean will tell you if the thing hallucinated a math proof, right? So it can reward it for not hallucinating. That's part of it. But also, if you think about software, software is just basically some complicated mathematical object, a software program. Program is just basically math. And so, uh, what if you wanted to prove that the software couldn't be hacked? Even if you had a super intelligent AI that was trying to hack your software, I have a mathematical proof now that the software cannot be hacked, no matter how smart you are. It just can't be hacked in certain ways. It's mathematically impossible. Super intelligence can't get around a mathematical impossibility, no matter how intelligent it is. So that's this idea of formally verifying software. So now we're thinking, okay, Lean, which we thought was just a pure math project in some sense, actually might be really important. What if we could formally verify and have defense-dominant cybersecurity? Things could not be hacked in the limit because if you can have AI to prove that the software can't be hacked or synthesize the software and prove that it can't be hacked, software becomes defense-dominant now. So now we're thinking a lot about should we go into formal methods, formal verification of software rather than just thinking about this as math and that's because That's because we're worried that obviously that when AI gets advanced, people are gonna be hacking right and left and our societal infrastructure is gonna be very vulnerable to attacks.

Yeah, sure. And were they one of your first projects you funded?

They're kind of in the middle of our, somewhere in the middle of the 9 or 10 that we have now.

Okay.

It's been amazing to watch their user base grow. DeepMind has used them in their research. Terry Tao is a Fields Medalist, probably one of the leading mathematicians right now, and has disproved one of his own proofs using Lean and is like actively using Lean. Harsh. And he's like live streaming using Lean. I didn't know it was the thing you could do.

But actually getting data, getting data on what Terry Tao actually does with it is actually very interesting.

Influencer or something. Let's do— I'm going to— we're going to move off Froze in a second, but just— and don't do E11 or Forest because I'm going to take you guys into neuroscience land. But let's just make this concrete for people one more time. Give me, give me one other Froh and, and, and why that the way you're solving this problem, um, you know, fits with the, the hardcore philosophy of the organization.

Yeah. Um, let's talk about the climate. So, uh, assuming climate change is real, lots of efforts to reduce emissions, um, we probably also need to draw carbon out of the atmosphere. If you do math, a lot of the ways that we can do carbon out of the atmosphere are actually really energy intensive. One of the least intensive ways is to use an open system, so something where carbon already fluxes in and out of the atmosphere. The biggest carbon sink is the ocean, and, you know, tons of carbon regularly move back and forth. If you can increase the alkalinity of the ocean, it then actually can draw down more carbon. It's a big buffer, basically. We're already already pulling a ton of carbon into the ocean as more CO2 goes into the atmosphere, which is why ocean acidification is coming. So basically, you can take a big alkaline, a boat full of alkaline material, a boat full of Tums, toss it into the ocean, and we should have carbon removed. And it should be removed in this way that is abiotic. It shouldn't really affect biological systems. But there's a but. I have no idea how much carbon is removed because it will happen over the next decade. That carbon may, like, or that, that ton of Tums may float along and then, like, sink to the bottom. It has to go out through the estuary. There's, like, all sorts of physical processes going on and then chemical processes going on. No idea how much Tums is actually— how much carbon is actually removed. Um, you just can't measure it. And so it has held back the industry, because if you're a buyer, you want to know how much carbon removal you have purchased. And these companies who are spinning up in the space can't tell you. And so we have a fro called Seaworthy, which is building a set of physics-based modeling tools to tell you that if you did a perturbation intervention and you put in an alkaline material in this estuary during this season, where that carbon flux will happen over the next decade. And it's an incredibly complex physics problem, right? There is the, like, very, like, surface-level chemistry problem. There is, like, the local estuary flows, the regional flows, and then actually into, like, global flows. And building models at all of these levels, they exist in each level, but acting between the different levels of granularity is incredibly mathematically complicated. They— there's a bunch of orgs, you know, NCAR, NOAA, who can do pieces of this, but the problem of solving how do I solve, uh, measurement and verification for marine-based carbon removal is just like not their agenda. So no one is going to build the giant modeling software. And so there is a team in Boulder, Colorado led by Matt and Alicia who are just killing it building building only this infrastructure. So over the last couple years, they have, um, released a map of all of the different kind of carbon fluxes. You can check out their website, and they're working with partners to, to develop this tooling.

Okay, that's— so that's great, because I think some people might— there can be startups that come out of this research, but there don't always have to be. And I mean, this is an example where, like, there may be a business case, there may not, but, like, we want this to happen, and you guys are willing to help make it happen. And sorry, I will push on past this, but just it occurs to me we've talked about this and people who might, they might want to know what the business case is around some of this. I mean, the, the people who work on these technologies, if it's, well, any of these throws, I mean, they are able to after the fact, like if they see a company in the, from the tech and they would have, I guess, like a head start over anybody else in this field.

If it makes sense, if it makes sense for the impact, right, for them to go off and start a company, if that's the best way to take it forward, generally speaking, you know, we can support that. So I mean, in general, we are looking for what we think of it as sort of at-scale or scaled revolutionary outcomes. So scaled revolutions that could take the form of commercial breakout, right? And sometimes it needs to in order to actually have that impact. I mean, the DNA sequencers that everyone uses, that made by Illumina, which is a multi-billion dollar company, or you can think about OpenAI or, you know, other things in the AI space needing this compute scale. And so it, it sometimes it makes a ton of sense. And the, the FRO would basically get it to the point where, whereas initially you would have to just raise philanthropically, it's essentially, but if it works, You now have such a powerful technology and you have such an ecosystem that you've built around it, partly by being very generous and very open, that now the world is going to interact with you differently. You're in a different world, essentially. You've shifted the difficulty of now doing companies in that space. But maybe you've— maybe it's not through a company. Maybe you've shifted the difficulty of partnering with the government or partnering with another large institute or creating a long-lived nonprofit in that space. Sometimes we think about it. Let me give you just— I know we'll come back to neuroscience, but if we could get brain mapping 100 times cheaper, that's going to make it a lot easier for philanthropic actors, government actors, and commercial actors to take connectomics brain mapping as a forward, right? Isn't that not a multibillion-dollar endeavor to map a mouse brain? Let's say it's now a millions of dollars question. Now you've changed that landscape. The donors didn't, you know, profit from that, but they enabled that to happen and somebody else can take it.

Yeah. So what are the trade-offs? I mean, so if, if, um, your donors do— does Convergent own ultimately the intellectual property that comes out of this? And then also, you know, for the people working on this technology, like you said, I mean, I see all the pro cases. I could see how in their head A lot of what you guys do is open sourced. You're sharing your data. That's sort of the point of all this. So you could make an argument that yes, you kind of have a head start and maybe that's, and that is great, but you're also kind of handing opportunities to other people to compete with you. So yeah, I mean, who owns what?

Just like a university, essentially the, the FROs, which are part of our, in the US projects are part of our US nonprofit umbrella with, with the— we're doing some projects in the UK and that have a slightly different type of structure where they're a bit more independent of the parent, but they roll up into our nonprofit structure. And so just like a university can license IP out to a spin-out company, I mean, Stanford owns some IP on what became Google, for example. We can license out to a new company, for example. And yet what you're saying, I mean, I think is quite right. The team and the focus and the organization and the know-how of that is tremendous. To actually go and do something versus the abstract IP as well. But in any case, we can license the IP at fair market value as a nonprofit to spinouts.

Okay. I feel like it'd be doing a disservice if we didn't talk brains because you're both such big thinkers on this area and have lived through it. Again, not sure if this is an uncomfortable question or not, but as you both know, I know Elad and Brian Johnson pretty well. And so, you know, they were kind of doing this kernel, people People probably don't know Kernel as well, but it was trying to make these helmets that, well, it started out wanting to go kind of full Neuralink with a more invasive kind of thing and then ended up in this place trying to read neuronal and blood activity from outside the skull without doing an implant. But, you know, I've been with both those guys. I mean, I felt like you might have, which is what I was reading last night. I felt like you must have been in the middle of this moment that I've talked with both of them about, which is I kind of think Bryan was there first and then was trying to spin Elon up on this idea of doing some brain-computer interface of 2016, right?

Right.

And then all of a sudden, yeah, yeah. And then as I've seen happen with Elon on some other things, he kind of decided to do it himself. But so yeah, were you— um, I forget what I read, but it sounded like you were—

um, I was involved in some early discussions with the team that became the Neuralink founding team.

Okay.

I didn't end up joining that team, but there was a very interesting early discussions prior Statute existing as a company. And then I, um, very shortly after that joined, uh, Brian as Chief Strategy Officer of Kernel.

Okay.

And helped them start the work, particularly the work that went on the non-invasive direction.

Do you want to settle the debate, as far as there is one, as to who got there first?

I actually, I mean, I think there, I think there were some, there was, there was, I honestly think there was just, there was two things going on and they both were just discussing with each other and ended up not working together. And they ended up taking extremely different directions. Yeah. And I think it's great actually that they took two different, different directions. Yeah.

And so why, why then? And, and we've got the, this Utah Array that's around for decades. It's a device that is invasive, that is accomplishing amazing science, but it's trapped in universities. And then, as you said, like 2016, all of a sudden, um, people decide, okay, this is, we're going to move this into startup land. And then that has only, only dramatically—

it was very inspirational, certainly, for this, the idea of the FRO is to see what can you do in an academic lab, the brain-computer interface problem, right, it requires things like chip design and microfabrication and medical device, you know, hermetic packaging and FDA regulation. Uh, it's not really a neuroscience problem, but it was mostly left upon the neuroscientists to mostly make their own tools, right? And so being able to shift it into this cross-disciplinary deep tech startup type structure, it was like, it's— it felt very enabling. That doesn't make it not a hard, incredibly hard problem to do that. But it was, it was a major shift. And I think I actually think it was a positive shift.

Is that the shift that took place though? It just happened to be the two very wealthy individuals happened to get interested in the brain at that particular moment.

I can't speak for them, obviously, but I think part of what was happening— what year was OpenAI founded?

OpenAI was around 2015, 2016.

Yeah. What year was AlphaGo?

Or yeah.

People were seeing AI, therefore we have a renewed interest in the brain. And I think we actually share that. We think our neuroscience projects are even more important in proportion to how fast AI goes, actually.

Okay, this is where I want to go to. So, which is interesting. And we've talked about this before, I think. And I don't, I'm usually not very eloquent in how I explain this. But there's this fascinating thing where we are, we, okay, yes, we have brain-computer interface startups now. So there is is money going there. It's like, and it's large, but it dwarfs what people are spending on artificial intelligence and creating brains in the, in the, in buildings. Um, we've talked about like the Connectome in the past, you know, so that's an effort to map the whole inner workings of the brain. I would argue, you know, relatively little money overall is kind of spent on what seems like the thing that defines our entire existence that we should probably know about. I, I can see there's an obvious answer to my question, which is, well, AI has the potential to make money, and knowing how our brain works, like, may or may not. And, and, um, so I get it. I get why there's this divergence of resources. I also find this like an interesting philosophical question, especially in this moment. You would think that, um, rich people and rich countries would devote just— I would think so many resources to understanding the most important organ, imagine the thing that defines everything. But we're clearly spending way more inventing the artificial version of that.

It's also, it's also hard to find the right focality of those efforts. So there's obviously a lot of money spent, for example, on Alzheimer's, whether that be the research or failed, often therapeutics, right? That's part, that's one piece of of, of what's going on with the brain. Um, what we're trying to do is say, okay, let's step up on the tech tree. The problem is connectomics is too brittle and too expensive. If we can first do a FRO that makes it less brittle and less expensive, can that unlock what really should be not even that expensive, but it should be—

I—

we need to map entire mammal brains, right? It's crazy that we haven't mapped entire mammal brains.

Yeah, what we've done, a worm and a fly, right? Yeah.

I think that there's a that physicists have it easy in that you can, from first principles, basically all agree on the right experiment to run, um, right? You like, we have to build a collider. Like, everyone knows we had to like figure out how these particles interact, at least in a few key cases.

Yeah, the accelerator has to get bigger, or else you can't see smaller particles. It's just like, gotta make a bigger one.

Yeah, I think that in things like biology and neuroscience, that's much much tougher. How do you get everyone to agree on what is the right rig to buy or to build that enables the field? Because it's much harder from first principles just to do math that leads you there.

And should we be studying how the mouse brain does spatial cognition, or should we be studying how the monkey brain neurodegenerates? There's all these different possibilities. Now, I think if you have a cheap enough technology, you don't have to choose. You can do 100 mouse brains that have different diseases and normal. And, you know, so this is why we think neuroscience is so technology-driven, um, because it is hard actually to agree on a scientific megaproject, right?

Okay, okay. The— I mean, I'm like, I'm like excited in this moment because it feels like we are— it seems impossible that over the next 10 years we do not gather more data about the brain on on like some just insanely dramatic level than we've ever had before. I mean, we've got from all these startups, we're just going to have probes sitting on brains in ways that we never could before. So, so that seems to happen. I am like a slightly— I've spent a lot of time with these BCI startups. It's like my favorite thing to cover and write about. I'm concerned though, you know, because they— Neuralink, the Elon of it all, Neuralink has done great things. They've brought all this money and attention into the sphere. We just talked about that. These companies probably wouldn't exist, at least like 20 of them, if it wasn't for that. They just got this massive new valuation. When I spend time with these companies, it's so clear to me that in the near term, this is all about helping people in quite dire circumstances. To me is a limited market. You're talking about people who are paralyzed, have ALS. Maybe if we get to like mental disorders, Maybe it gets big enough. I, I just, I worry that there's this gap between what I see these companies are being valued at, what are almost like consumer electronics companies, which is the vision for Neuralink, and it's vaguely the vision for some of the others. But man, we're on a timeline because if, if you— I just worry that if we don't, I think it will take longer than some people think. And then I think the You could just have the rug pulled out of this whole thing if you're not able to hit these numbers. And yeah.

What about— no, I want you to tell.

Well, one of the reasons that we— tell me if this is what you want to continue. But one of the reasons that we got excited about Forrest Neurotech is that I think it— if you look at Neuralink has gone in a direction that has invasive probes that can be very fine details. But in return for getting those fine details around any one link, you can only get a certain number of channels. They're not going to— one link is not going to see things that are extremely far from that link.

You're like in the motor cortex.

Yeah, it's a motor— local stuff in one part of the motor cortex, and then you need to go— if you want to look at some other part, or especially deep brain areas, you need to add on something else, another link or something. On the other hand, you have these non-invasive methods, and that's sort of where Colonel has gone using light, for example, to try to go in and the light enters and interacts with the neurons and then comes back. But that's kind of fuzzy and coarse because you have to go through the skull. And as things go through the skull, they kind of spread out and it's hard to localize something. So ultrasound is a really, really interesting trade-off point where it's a kind of nice property of sound basically that the, the wavelengths of ultrasound, okay, it's a wave, sound is a wave, and the wavelengths of ultrasound that can penetrate relatively efficiently through brain tissue, and therefore maybe you put your probe on one side of the brain that emits and detects the ultrasound, and then you send the ultrasound, it could see all the way to the other end of the brain. So the ultrasound propagates through the entire brain. It can do that even at small enough wavelengths Okay, that it can also see fine details, right? So it's a balance between how fine of details you can see and field of view. And so ultrasound has a unique trade-off between field of view and resolution. And so if you can, especially if you can put an ultrasound chip, miniaturize it so that the emitter-detector chip of ultrasound is on the inside of the skull, but it can do that without having to touch the brain, right? So that's very powerful, right? That could be safer but also allow more general because it allows you to see the whole brain, right?

Absolutely.

So there's this, this resolution, um, scale trade-off. There's this author, uh, how invasive, how dangerous, what are the downsides? I think that we are a, a far place from a healthy individual opting into a, uh, highly invasive, potentially very dangerous surgery. If you talk to neurosurgeons, and this is both like just a craniectomy, we're just taking off weirdly importantly a non-circular part of the skull because of how billing works. They only draw a circle. It's really bad for the neurosurgeons. It's not a circle because of the way the tool works.

Okay, the procedure. The procedure in terms of reimbursement of the procedure, it's better if it's not a circle. Oh, that's a That's a detail.

Yeah.

Okay. It's a small— the smallest, smallest removal of part of the skull.

Oh, you said because of the way the billing works?

Yeah.

Okay. Okay. You— should we explain that?

How the— The way that surgeons get billed is per procedure. And the procedure that drills a round hole is less lucrative for the hospital and for the surgeon than a procedure that doesn't drill a round hole.

Okay, great.

And the surgeon just has to drill with a circular drill.

That's all.

They don't have to do any work, you know? Oh, they just have to drill some things.

That's okay.

A neurosurgeon saying, "Oh, just the skull," versus past the dura into the brain are just, like, radically different. The chance of infection is radically different.

Yeah.

The chance of, like, negative side effects is radically different. They're just like, you know, to me it's all brain surgery, but to a neurosurgeon, they're vastly different things. And so ultrasound has this ability to both be fairly fine detail, fairly large scope, can see most of the brain, and is not nearly as dangerous as things that actually have to enter the brain.

And I don't think— I don't think I'm revealing anything I'm not supposed to, but with Forest, you do this implant, like you're saying, I think it sits on top of the dura. That's what we're talking about. And then the idea is that the patients can have this in when they're, you know, Typically, if we're doing stuff like this, it's like you're trying to measure mental illness on someone while they're in a hospital in unusual circumstances for just a few hours. And it's like, what are you— you're not like, what is the baseline you're even seeing from this data? Whereas with this, you could have it for a month, for 6 weeks, you could be at home forever.

That's another thing overall that these companies— I think Neuralink deserves a lot of credit for this— is sort of making a really miniaturized, really good piece of electronics so that it's not connected to this giant mainframe kind of machine.

People do not get this.

That means you can really, this context of use really changes. And so that's, I mean, that's powerful. And of course Forest would have a similar property of being able to be fully wireless.

So then you could actually see when you're depressed in your normal environment and get something that's like possibly real data. Yeah, people do not appreciate this. Even the, it's like, I get that lots of people hate Elon and stuff and Neuralink seems weird, but when you're with these patients, it's like they're at their house.

The low power electronics involved in this is incredible.

And some of the other startups, I kind of, I'm interested in all these companies. I'm really not trying to besmirch anyone, but they always just tell me like, oh, it's like, you know, well, Neuralink has these problems. And then, um, we're also going to make this compact device like in 9 months and stuff. And I'm always like, will it take, you know, you guys might be underestimating the, the, the, that part, which, 'cause I, I, I think they don't get enough credit for some of that because, um, it's very hard.

It's just engineering, right? It's incredibly hard engineering.

Like just the manufacturing line just has to like not leak anything and last for a really long time. Time and be tiny and be wireless. Yeah, um, many steps. Um, okay, I'm not going to keep you guys too, too long. Um, oh wait, so I knew 2017 Brian Johnson. Did you imagine him becoming worldwide longevity guru?

He's— I, I love the guy. Yeah, no, I mean, I think, I think that there's, there's some shared elements of his, his zeitgeist over that time. I didn't imagine the specific—

there were hints for sure. Um, but he definitely has—

he had— I I think his overall personality has stayed pretty similar. Yeah, he was great to work with. Yeah.

And, and when you guys— what else? We talked about AI and how it's changing your thing. I've ordered— like, what are you most excited about? What are the— since you've done these 10 areas, what's like the coolest thing that you're hunting for right now?

I mean, a lot of it is actually just AI-related safety, security, and, and biosecurity. I mean, those are two areas where we really want to push more.

This is Are you doomers?

Reducing the downside risk. Not really, but I think that there is such a thing as, I mean, I think we really like this defensive acceleration framework. Can you accelerate a technology which makes the civilization more robust to these problems, right? So the formal methods thing I mentioned with provably secure software is one example of that, yeah.

I actually like pretty concretely we're not doomers. I like how we have like a shared—

Admit it. No, I'm just kidding.

We have some concerns.

I mean, I'm not not a doomer, I'm not a doomer.

I think that we see a version of a world, I think a lot of people are like either AI bad or AI will solve everything black box. And I think that we spend a lot of time saying, what is the version of the world or versions of the world that we can get to where this has gone extremely well? And so what are the downsides to this defensive acceleration point where we're like, okay, we should like mitigate biosecurity risks, we should mitigate some cybersecurity risks, whether that be on the hardware side or formal verification, but also enable— I don't know, why don't we have replicators? What are the fundamental material science questions to this tech tree problem? I'm not building a replicator anytime soon. What are the fundamental material science questions that underlie where we could go if AI could help us advance it quickly? And so how do we work on those technologies? On the neuroscience side, if we think that neuroscience, really deeply understanding the brain, will help both with what happens if we create other intelligence and/or how do we make better versions of ourselves with new capabilities. And so picking these things with these broadly optimistic visions of mitigating our downside and kind of creating as much possibility for the upside as possible.

AI is also even just thinking a little bit out on the horizon of what we might do. I mean, it's caused us to think a little bit more about social sciences actually than we had previously, right? Most of these things we've been talking about are, you know, microscopes or something like that. But are there techno-social questions, right? So can you use AI to improve democratic deliberation processes? Can you use AI to improve human thinking and epistemics even without a brain interface, right? Those types of questions are areas where we want to understand what the, if you want, fro-shaped problems or the directed development of fundamental capabilities. There's lots of questions about the ideas and we're not trying to solve necessarily deep theoretical issues like how to do alignment, for example, although I think some neuroscience may contribute to that. But we're trying to find the actionable engineering or directed research problems, including in those areas of social science.

How do you find those people?

I hope you go on podcasts and you talk about how you want to find these people, then they reach out to you.

Are those sociologists?

Are they philosophers? There's a There's a mix. There's a set of people thinking about democracy that maybe are using some software. There's software called Polis, which is about scaling more deliberative processes. Okay, how does that change with AI? There's been some research from DeepMind, for example, on using AI for democratic deliberation purposes. There's a set of other people that think about recommender systems or think about prediction markets or think about other types of epistemics, human thinking type questions. We go talk to those people and we say, okay, well, can you do what you're doing now in your startup, in your lab? Do you need some sort of fundamental change to society for this to work? In which case we may not be able to help. But do you need like an engineering project? Yeah, let's, let's talk, you know.

Yeah. You nodded when we were talking about this before pretty emphatically, but, but I still have my mind just talking to you too. Just keep, you know, why, why, why are these nerd billionaires not just dump dumping. I don't understand.

I remember I interviewed— giving this a finite size. They don't have to dump everything. Yeah, they have to dump a meaningful amount.

Well, I interviewed Larry Page many years ago, and I think at the time— I'm doing this from memory— and I think he was, he was in the top 10 richest people in the world. I think he probably had roughly like $25 billion. And at the time, because Elon was not Elon, these companies were like doing okay,, but kind of not anywhere near where they were. All his wealth was tied up in Tesla. See, I think he was worth like $400 million or something like that, which is great. But you know, Larry, Larry was like, oh, I'm so jealous. He told me, he's like, you know, I wish I could do what Elon does. And, and I, it's like, I kind of got some of it was that part of it I think was like he was CEO of a public company and, and there's baggage that comes with that. And Elon could be this free radical, uh,, to some degree. And that, that if Larry does a flying car company and it doesn't work well, it's sort of like reflects kind of badly on him and on Google. And it was like, why, you know, who was funding what? And there were extra questions. However, you know, when I look at the list of like the world's 50 richest people, and now we're talking about Larry's net worth, and I'm not just— he's a very nice guy, I ran into him the other day. I'm not I'm not like totally besmirched him, but you know, these people are worth like $80 billion now, $120 billion. Yeah, I just, it's crazy.

They could fund the catalytic activity to unlock the entire gap map, for example, I think with private, you could do it.

And they're all, they're all science-minded nerds.

I think partly, partly they're busy. Yeah. They're running companies and busy and stuff. Partly they want to make very high quality decisions and the marketplace of philanthropy is very opaque. How do you know who is the right scientist? You can go to Stanford, go to another university and there's lots and lots of options, right?

If your money's being spent well.

How do you know it's doing it well? Yeah.

I think that in some ways I have a lot of empathy or sympathy for philanthropists. If I'm running a venture firm, I'm making a pretty clear deal. I'm optimizing for dollars in the door later. If I put out dollars now, will this or will this not return multiple X in 10 years? And it's like a very clear trade-off. If you're a philanthropist, you're like, well, do I donate to the Met? Do I Do I fund new artists? Do I help save lives in Sub-Saharan Africa?

Science storytellers.

Science— do you fund science storytellers? Do you fund scientists? Do you fund, like, the existing scientists who already have a network? Do you find new scientists? Is it— there's no money back later on the other side of that experiment. You have to be like, do I think that worked? And I think it's a complicated thing. And I think that Many philanthropists then build organizations to answer the complicated thing, and then those organizations become beholden to the things that all organizations are beholden to of internal decision-making and internal processes. Um, and I think that it's complicated. I think that our hope is to build convergent research into an organization that can start to build things like the gap map and a list of what we think are the highest leverage bets you can make in science and technology, and say, actually, here is a bet that is, you know, $2 million to unlock a seedling that would then, if that proved out, be a $40 million bet that then would transform this field, and make a series of fairly robust roadmaps and conclusions to help people navigate this decision. Um, conversion research, much like Science Storytellers, also has to go. So figuring out how to build an organization where that is a valued thing from the philanthropic community, of really doing the deep dive into all of the different areas of science and across the board saying what would be highly impactful.

Okay. Um, all right, let's end on a completely unfair, um, open-ended, um, impossible question, just because I feel like you're both so into deep science, you're kind of sci-fi type thinkers and everything. I mean, give me— give— and, and look, again, I just said it's unfair. There's many, many wild cards. Like, give me— what's your— what's your vision of the next— where, where are we in like 100 years? I mean, are we about to unlock, you know, some great flourishing? Um, given the caveats I've already put in place, when you, when you go to sleep at night, What are you guys picturing?

I mean, it seems very volatile right now. It does. I mean, the question of, doesn't mean I'm a quote unquote doomer, but I think it does feel very volatile. So I think we have to navigate it carefully. But I mean, one of the things I do, I do think the neuroscience thing is a big question. I think, and I think maybe in the years shortly after we have super advanced AI, we need to spend a lot of time on neuroscience, right? We may have super advanced AI and not know if it's conscious. Conscious. We may not have, you know, we may not know what determines whether we're conscious. We may have super advanced AI and not know how to retune our own brains, right? So basically the tunability and understanding of the human brain, I think that's kind of the— it's actually really hard for me to know what 2075 looks like without that. So I think you can say, oh, we'll have rockets in space and we'll have a lot of low-energy compute capacity and we'll have secured the world against bio threats and stuff like that. And we'll be be able to then do incredible, um, you know, accelerated, you know, research on aging and other diseases. I mean, this is, this is all great stuff, um, but like, what does 2755 actually look like? I mean, I'm not sure. Greg Egan, maybe as a sci-fi author, he tried to go out there, and very few have tried to go out.

I mean, you make that— you make the point that I'm, that I'm betting on for, I think, I don't know, who knows what it is. Let's call it like the next 20 years of my life. I mean, I think it's the— yeah, it's this brain, brain on brain interplay, questions, knowledge. I just think this is— so that part, I think, I think that's—

there's still science in between. Yeah, here and there, actually, which would— I think our concepts in 2075 of what are we trying to do, yeah, may be different because of neuroscience.

Yeah.

All right, Adam took us back to 2075. I did 100 years. It's probably kind of gross and not really unfair. You pick any timeline you want.

So I actually think I have high hopes for 100 100 years. I mean, I really think that we're at a steep part of the curve, and 100 years from now, there's so much that we can even accomplish. I think the next 20 years will be weird, man.

Weird.

Yeah.

Yeah.

And so I think that in 100 years, I don't think that the replicator is impossible, right? I don't think that, you know, true high-bandwidth BCIs are impossible in 100 years. Like, I'm not sure about escape velocity for longevity. I'm like not convinced that our meat sacks are meant to do that, but like I think there's a lot possible.

You are? Yeah, hang out with Brian, man.

Anyway, I'm sorry, but like I think like 100 years, a lot of these things are possible. I think that building— what are the bets that we can make in these next 3 years that make the following 20 the best version is like a thing I'm thinking a lot about.

Okay, okay. All right, next time you guys are on, we're talking partial cell reprogramming, iPSCs, and NeuroPixels, which are 3 things I didn't have time to get to. Um, thank you so much for coming on. Um, are you going back to EDGE, Esmeralda, for a little bit?

Yeah, actually, at EDGE, if you go, Avery— whose last name I don't remember—

Avery Krieger.

Avery Krieger is there, uh, collecting a bunch of neurodata and phenotypic behavioral data to build a giant dataset all month. So he's just getting like hundreds of hours of data as a pilot project for this crazy thing. And this is the kind of stuff people are building at Next.

Like an EEG or?

A lot of it is about the behavioral capture.

Okay.

Yeah, it's interesting.

So you don't get a neurochemical activity.

I mean, they ultimately want to say, you know, neuropsychiatry or neurosurgery or so on, neurology. They want to be able to just capture richer human behavioral data.

He has an EEG there too. He got a bunch of equipment donated. Okay. Doing like an hour of neurodata for 10 hours of behavioral data and just building a big dataset.

Okay. Okay. Do you guys— you must get bombarded with like, should people send you their wild ideas?

Abstracts@convergentresearch.org. We have a particular one-pager format.

Yeah.

Okay. And we can canvas those. And I think hopefully from this discussion, you understand a little bit what the flowchart is don't send us a 1-person project or a $1 trillion project or, you know.

What's the sweet spot? It's 20 to 50?

Yeah, think about what you do with, let's say, 15 to 30 people over 5 years when those people can be professionals in the same way that, you know, deep tech company hires professional operators, professional chip designers, professional biologists, whatever.

Okay. Well, there you go. If somebody pitches you guys and it gets turned into a thing, we get the exclusive. Exclusive and claiming, claiming rights on anything derived from this, this podcast to, to, uh, not on the technology, but on the story, on the story. Um, uh, thank you guys so much for visiting. I really appreciate it. Two of my favorite, most interesting people.

Um, thank you so much.

All right, cheers. Core Memory is hosted by me, Ashlee Vance. The show is produced by David Nicholson and me. Our theme music is by James Mercer and John Sortland, and the show is edited by the John Sortland. Thank you all for listening. As mentioned, the show is sponsored by Yuan Ventures.

Thanks.