A

Foreign.

B

Welcome back to the AI Policy Podcast. This week I'm excited to welcome Austin Carson and Josh new from SEED AI, a nonprofit working to establish AI readiness for the United States. Austin is the founder and CEO at seedai. He previously led government relations at Nvidia and and held several public sector and NGO positions, serving as legislative Director for chairman Michael McCall of Texas and as Executive Director for the Technology Freedom Institute. Josh is seedai's Director of Policy. Before joining seedai, Josh was a technology policy executive at IBM and a senior Policy Analyst at the center for Data Innovation, Austin. And Josh, thanks for joining us.

A

Thanks for having me.

B

Yeah, I'm glad to have you here. I think the first question I'd like to ask is for you, Austin. So you've been a pioneer on AI policy issues, recognizing the importance of this area. Well before the ChatGPT moment, it's what led you to join Nvidia. Back before it was the kind of household name it is now. And then to start seed AI, which happened in 2021. So that's where I wanted to start. Um, what drew you to working on AI in the first place and then what drove you to move from government to industry and eventually into the nonprofit space?

A

Yeah, no, I appreciate it and thanks again for having us. It's a great opportunity and I love seeing you taking over this podcast. No shade, Greg, but I love seeing Alec take over the podcast. So, you know, the, the first big moment for me was whenever Google retrofitted Translate with deep learning. I remember one day and I think it was like 2014 or 2015, I noticed that it worked way, way better. And then there was like a New York Times article that was discussing in general modern transformer based deep learning techniques and what Google had done. And then like some of the results and it had kind of some of those fun early shadows of the larger conversation around LLMs and transformer models. A little bit of this, like, we don't know how it works and why, but it works really, really well. Both of which were supremely interesting. So I started exploring some of that when I was in Chairman McCall's office. We had just fin working on a big project on encryption and trying to balance the desire for some in law enforcement to just say nerd harder and figure out how to make encryption that does what we want. And then everybody in the technical field that's like, doesn't work that way. You can't. We all wish we could math harder. It's hard to math harder, you know, so trace that for a while. I think I have the pleasure of throwing the first briefing on Capitol Hill on artificial intelligence, at least specifically as such, in 2016. And then I ran a bunch of coordinating devices when I was in Congress. So like some congressional caucuses, I started a staff association, we did a bunch of educational briefings. And I really enjoyed building that connective tissue that helps people communicate with each other and learn about complex topics. And so at a certain point after 2016, coming in early 2017, another nonprofit, the Technology Freedom Institute, reached out to me about being executive director potentially. And a big part of what they wanted to do was again work more on building this connective tissue and doing some of this technical understanding. Unfortunately, it didn't ultimately make sense and work out for a number of reasons. For our next podcast together, and then I ended up going over to Nvidia, who was trying to start a conversation with dc. So I actually went there before they opened up a proper lobbying shop. And so my job was as the non lobbyist government relations person, just to understand everything happening in Nvidia and then to run around D.C. and trying to explain what parallel computing was and why it mattered and what's happening with high performance computing and why can we now split spectrum in a thousand tiny parts such that your phone can work better and we can have more phones, you

B

know,

A

which is honestly way more interesting than it should seem, which is part of why I think the job made sense for me. And at a certain point I got to see like some more ecosystem development, more like build up AI capability in areas that don't currently have IT projects in the company, which I found very interesting. And there were also major developments happening. One, with kind of like the predecessors to chat, GPT, GPT 2 and GPT 3 and early Claude were getting very good, clearly in a way that was somewhat strange. And then noticed from some of the research internally, some very significant scientific algorithms were starting to get accelerated at like a thousand x. So you're getting to the point where like simulations could now run at a thousand X. The fidelity effectively is what that means. We're entering into an interesting world there. We've already entered into an interesting world where the way that we're processing information is already moving into this fuzzy math place and where the great relative advantage of humanity, which is like our ability to use words and complex thought, is now being put into like, you know, substrate independent effectively. So realize it's not any company's job to go and explain, like to ultimately explain that to everybody just for the public good, writ large, absent Their, like, fiduciary responsibility and their priorities. And Nvidia, I think we spent a lot of time trying to. To square that circle, and especially as an ecosystem company, we could do a lot of that. But just to go and work on the outright policy issues as it was getting really intense, seemed important to me to have an independent voice. And so I went off and started this exactly one year before ChatGPT hit, you know.

B

Yeah. So I think I want to go there next, which is, you know, I first met you when I was at OpenAI and it was right after ChatGPT had come out. And so for everyone working in the AI policy space, that really shifted and accelerated the nature of the work and the intensity of the work. I know, Josh, you were also working on AI issues at IBM before all of this happened, and you joined Seedai after ChatGPT came out. And so I'm wondering how those developments shifted, how you think about CDI's mission and sort of both the urgency of the mission and the type of work you're doing.

A

Yeah, I mean, the first thing was I anticipated some degree of, like, pivotal moment isn't exactly the word I'm looking for, but something that took it from like, this technology is evolving to, holy shit, this technology is here. That was definitely one of the moments. But I think the biggest thing that changed about it was, first of all, everybody that didn't understand what I was talking about for the preceding year was suddenly like, oh, my God, this is that thing you were talking about. What a great idea to start this. You caught the wave. I was like, no, no, no, I caught the wave when I was at Nvidia. That's when I caught the wave. This time I'm just trying to surf the wave with everybody else, you know? And then it turned into something that was more about, okay, communities were conf. Like, kind of confused in the year prior when we were moving around the country talking to folks about building up AI R&D ecosystems, which is a big focus of ours, for national AI readiness, having your local area able to be part of the AI re, etc. And so it went from people saying, that sounds cool. We don't know what it means. We'd love to do something, I guess, to clearly, we have to have a plan. Everybody's saying, we have to have a plan. This chatgpt thing is crazy. What do we do? So, like, the intensity level and the amount that I found myself answering phone calls as opposed to placing them was like a dramatic switch. And I went from spending the first year trying to think about, you know, things like how could we support the National AI Research Resource in this slow and steady way? And what are ways to help communities understand how their specific data and industries could be leveraged for deep learning into like answering phone calls nonstop. And just like when I think we, we met was when I was in the middle of that, like, all right, I'm in a room somewhere and I'm answering phone calls, you know, and then we could move down in a more realistic way towards folks like community college students, because now it's just a natural language interface as your primary tool.

B

And Josh, I'd love to know how sort of chatgpt shifted how you think about AI and maybe how it sort of played in your decision to join SEED AI and work on the policy issues that are central for seed.

C

Sure. So I think before that, but you know, while I was at am, while I was at IBM, but even before that when I was at Italy, I have center for Innovation. I think I had a similar experience to Austin where like so much of the job we were teaching classes on like tech policy to congressional staffers. We're doing a lot of briefings. So much of the job was like, you guys should really pay attention. This is like, important. And then after ChatGPT was like, you guys are paying too much attention. Which is like, no, it's the appropriate amount of attention. I sometimes have quibbles about exactly what they're paying attention to, but people are now, now get that it matters in a, in a very real way. I think ChatGPT was really interesting because, like, if you've been paying attention and like you got a chance to like, see what like GPT2 did, it was like a fun experiment. Right. It wasn't particularly useful at anything. And like the big concern everyone or a lot of people worried about was like, fake news. Like, this is going to generate so many fake articles are going to be. Which like our concerns, our hopes, our dreams were like very much kind of quaint in retrospect, I think about what this technology would do for us. And then when ChatGPT came out, like, it was really, really impressive, but it still wasn't very. It was a sea change. And what I really think, like started to make real for a lot of people was that it was also a very clear, like, oh, progress is going like this, right. Like, we've seen these experimental models. Then we see like an actual consumer product. Like, I don't know, it's not going to be able to do a real Part of my job yet, or I can't really rely on it for much yet. I can't do a lot of coding yet, but I could see how that trajectory becomes real. And I think that has been like a really, I don't know, sort of foundational moment for a lot of people in the space about, like, why they should be paying attention and why this isn't just like a static thing they can learn about and be done with. Makes me very excited. But I still think people sometimes in D.C. over index, I'm like, this is what AI is doing. It's going to stay that way. And we shouldn't think about what the future will be like in five years.

B

Well, let's dive into the nuts and bolts of sort of how you're thinking about responding to this AI moment. So I'd love to dive into, you know, really what SEED AI is hoping to do and sort of what success looks like for you. So what is it that you hope to achieve and like, how do you want to change the world? And how is SEED AI trying to do that?

A

Dude, what a pleasantly small question. So, you know, as you said at the beginning, we are here to work on national AI readiness, right? We're a domestically focused organization. We keep it about the United States for number of reasons. One, it's already crazy enough here. There's 50 different states. They're all very different. I don't need to add Europe to the mix just to punish myself, you know. But second of all, there's a real question what national AI readiness means, right? And some of this is its own discovery process. And of course, we've been doing this for a long time, so we kind of know. But how do we go to communities across the country, understand what it is that they need to have both, like, agency as well as meaningful impact over the future? How can they understand, how can they both contribute to what it is for everybody else as AI comes into the world more fully? And then how can we make sure that we're respecting what's going on there and making them prepared to just interact with the bigger kind of the bigger picture, right? So some of that looks like, you know, we have this program, AI Across America, where we just literally go to different communities directly. And that's shifted a lot over the years. I mean, when I think back when we first started doing some work together, you were seeing the version of it that was more focused on red teaming. And can we make it something where community college students or other folks that are younger can use this as an opportunity to kind of try to break the model, to learn about it and then add value back to the security of the model as like a plan. Right. And that's one thing that helped engage folks in 12 some odd different states. But it's a bigger picture than that. Ultimately, if you're trying to look at how do you make the country ready, Right. And so going to each community, like we just went to Nashville, Atlanta, Baton Rouge, Salt Lake City before that, we've got a project running in Tulsa. But each place has some different need. And then their needs and their circumstances help us illuminate what national AI readiness means that we can then bring back to policymakers and help them understand. And help ourselves understand too, of course, and further refine our work. So, like, at its absolute core is that we want to understand and support what folks around the country are trying to do to become ready for AI, understand how like, the convergence of those things should inform our policy moves and policy decisions both federally and at different levels of government, and then see how we can ensure that folks get to contribute meaningfully. Because I don't like, I don't. The New York and San Francisco and D.C. aren't going to figure this out. And there's not really like a. This to figure out. You know, there's not like a, a one thing we have to understand, a one magical alignment thing that makes AI go well for everybody. Right. It's like, no, we have to do this messy work of supporting people across the country, raw shoe leather stuff, email several hundred people in a place that we've never been to before because somebody at their university told us that, that it would be really helpful if we did. You know, now we try to be invited to places. We try not to be carpet baggers or like arbitrarily drop in to say that we went somewhere. But the biggest and most important thing, I think is keeping this focus on like turtles up, not turtles down. I guess there's a much better way to say this, but something like, I'm not trying to go have like a grand idea of what it means and then go and tell everybody they should do it. I'm trying to go talk to folks with some ideas and understand how we should like adjust and support those and then how we can make the combination of what we know and what they know into something that makes the entire country ready to succeed. I don't know. How does it. Is that ring? Do you need something else?

B

Yeah, I think so. So let me, you know, pull on a thread there, which is that there's this big trust gap around AI in the the US and you see that reflected in polling data where Americans say they're very skeptical that AI is going to make their life better. And you see that reflected in policy. So local opposition to data centers I think is the most concrete example of that. And so I'm curious because you do a lot of like direct community engagement and you do a lot of sort of work to demonstrate how AI can help people directly, right? Like a lot of sort of figuring out how to match the needs of communities to the AI tools that are available. So how your work dovetails with this idea of a trust gap and whether you think that the work you're doing can provide a blueprint, a bigger blueprint for how to sort of address some of the concerns that the public has expressed about AI technology.

A

Yeah, no, and I appreciate that question. I think it is a very central one. It's the same thing we're seeing in, in science too, to some extent, which is like, because these things are alien and separated from regular humans, they can't touch up. Science isn't a thing you can do. AI isn't a thing you can do. They're building a big data center. Whatever happens there, it doesn't have anything to do with you except for in the big picture, right? I mean, maybe there's tax revenue or some electrician jobs or something, but you don't actually have any bearing on the underlying technology. And again, it's not like the thing in your backyard is explicitly doing a thing that's for you at all as far as you know, you know what I mean? So there's one initial piece I think, which is like the more organizations and people that are going around the country and telling people that we care and legitimately meaning it as folks that care about policy and are trying to make policy or even trying to have policy oriented conversations with the companies and like other folks in the ecosystem. I think that really, really matters. Just as an absolute baseline thing. There's a second thing which is like taking it to people and very honestly saying I think you have something to add to this really matters with the trust gap. Now don't, don't send somebody that doesn't mean it to say that we're I think, strange in the sense that we as an organization are full of people that both a think that there is something very legitimate to learn from each community while also be thinking that we have value to add at that federal level and there needs to be this kind of like bigger Picture federal national policy thing that is a little bit easier and more like, you know, alignment and make AI go well kind of stuff. Right. So we, we hold both in at the same time, which I think is very important. And then finally, I really do think people need an avenue to like meaningfully participate. And I think the things that are around them need to have to do with them, have something to do with them. Like we've explored different ideas about how, you know, can there be joint research projects in all of these data centers? Can there be different modalities of data centers, some of which are more designed for like what the community would actually reasonably do or what could be federated between different communities? Because if you really look at it and ask like what is it for? You know, what does it do? And people can't really give you a great answer a lot of the times. And so I think going and talking to people and refining that answer because you're like, well, radiology, it's great. Well that happens like 1/100th of a CPU these days. I mean, if you're using a good thread ripper anyways. Right. So trying to make it less opaque is a big part of it. But really it's just shoe load work that I do wish there were like a hundred other organizations doing that were really that. All of which were affiliated in some way with the institutions that people don't trust as much, which would be like semi governmental or public private. I think, I do think there's a big role for both public and private to play in it. And especially the people that are directly making and deploying the technology. Yeah. And again there's this like kind of fuzzy, touchy feely thing that's like care understand people's lives. One of my first, one of my first political jobs, I worked for a state senator from Georgia. And I remember it was like 2009. And I'm going on to him about something about the Internet. You know, I'm like really talking about his campaign should do a thing about the Internet. And he's like, austin, Austin, this is Georgia. It's like people are just trying to buy peanut butter and tennis shoes for their kids. I was like, damn, that is an extremely good point. That is a really, really good point that for some reason I've become alienated from. Now you think that sounds unsophisticated, but people are also trying to do really sophisticated, interesting shit after they buy peanut butter and tennis shoes.

B

Well, I guess that the midterm elections are going to force at least some policymakers to confront AI impacts on the local levels in a way that maybe they haven't had to think about pre election season. And so we'll just have to see how that plays out. So Josh, I think I wanted to direct the next question to you, which is that's come up a few times in the conversation so far that one of seedai's big priorities and one of the places it has a lot of optimism is around the use of AI to accelerate scientific discovery. And so I think you're leading the initiative that SEED has on this particular topic. And so the thing I wanted to ask is there's this pattern we see in AI where you hear about amazing advances that are just around the corner. We've heard that say in self driving cars even as early as 2004, and I think we've only seen significant progress in that space to the point of real commercial development deployment of those in the last few years. We've heard similar things in the medical diagnostic space where there's a lot of optimism about specific efforts and then they fall short of their promises, like Watson for oncology and even questions about, you know, whether the productivity increases we see from LLMs are real or sort of ghost increases in productivity. So I'd really like to know how and why you think AI for science is different.

C

Sure. So for a bit of framing, I think the idea that AI could be transformed into science for pretty much every stage of the scientific pipeline, right. Like experimentation, design, literature review, simulation, automated synthesis and screening and automated labs and then the messier, harder part of tech transfer and diffusion is something that I think has excited Austin and I for a long time, even before we joined seed.

A

Right.

C

Like if you were paying attention to AI in like 2015, it was a lot about like AI being used to do like really advanced complicated weather simulations and do like interesting geospatial science or do like drug candidate screening or like MRI analysis, things like that. So the potential has been there for a long time. And what I think we are starting to see now in this post ChatGPT era is like the results of scale where we have models that are dramatically more generally capable, can process a lot more data, can do really advanced reasoning to tackle for really the first time, really, really complicated questions in a way that is potentially dramatically faster, adaptive, more accurate, et cetera. And this is true for the person trying to solve a theoretical math problem as well as it is for trying to screen thousands and thousands and thousands of millions of molecules in silico before it decided to pick A candidate molecule for a drug. I could go on and on about all the different kinds of benefits we're excited about and stuff that we're already trying to see. There obviously are limitations in this too. A lot of bottlenecks in science have nothing to do with how good the technology is. And so AI for drug discovery is sort of like this ur example where on one hand we have Alphafold, which is a Nobel prize winning innovation that can predict the three structures of proteins, which like is a dramatic thousand x. I don't even know if you can quantify how much of a productivity enhancer that is for biomedical science. At the same time to like make a drug. This could be like 10 years of safety trials or like studying effects on humans or like messy regulatory processes. Like no matter how good AI is, that is just a messy human political problem that maybe I can help in some way in fixing. But I think the expectation that this is going to change things overnight is perhaps the wrong framing. However, we already have drugs in the drug review pipeline that were designed by AI for the first time. That hasn't really happened before. So I think we're going to start seeing a lot of incremental speed ups at a lot of different parts of the scientific process that over time what the bottlenecks are going to become abundantly clear. And then that could be the focus of policymaking. We spend a lot of time through our coalition that we run called Accelerated Science now, which is a little over a year old. We launch with 20 something numbers an hour at over 80 Frontier Labs, R1s, startups, community college networks, flying computers, et cetera, all focused on different parts of this problem in this opportunity space. A lot of the things that we spend our time on are like what can policymakers or industry in the private sector and philanthropy do together to overcome these bottlenecks or preempt some of them. And so one of the ideas that is like stuck in my head quite a bit right now about how to like actually dramatically unblock progress is like treating critical inputs for AI for science as like or like infrastructure as deliverables of research funding. And so like data sets, right? Like we often find science in the United States as like a test this hypothesis, prove this out or try to validate this model versus like produce infrastructure that would let 10,000 different scientific research teams also solve these problems and then a hundred other ones at scale. It is sort of trying to meet the moment where like as AI is transforming science or the potential for what science can be. The institutions that do Science and how we fund science and how we, like, think about distributing science. Also need to rethought to sort of meet this opportunity. So happy to get into that more.

B

Well, I think, you know, where I want to go is that there is a lot of optimism about the role AI can play in accelerating science. But there's a dark side too there, right, which is that there's this dual use capability from accelerating science. And if you provide these capabilities, bad actors can also get these capabilities. And so they might use them to sort of accelerate their ability to build, say, a bioweapon or engage in some other sort of risky activity. There's an analog to that we're seeing right now. All these discussions around Fable and Mythos have to do with this idea that you can build AI models that are very good at detecting cyber security vulnerabilities. And that's great if you're trying to figure out how to shore up your network, but it's also great if you're a hacker trying to figure out how to hack a network. So, especially from a policy perspective, how do you think about this dual use nature of scientific progress? And you know, what are some of the things you're thinking about in terms of making sure that we are really putting our thumb on the scale and making sure that the scientific capabilities from AR are making society better.

C

So go ahead. One thing I think of is that like, we're talking about like knowledge creation or like applied knowledge. And like that is like human expertise is like inherently dual use in that way. So it's not like these are necessarily unfamiliar problems. Right. I think they perhaps take on a different character or different to mention when we're talking about automating them or doing them at scale or doing them faster than ever before. But I think thinking really hard about what these specific risks are helps us much more effectively understand what mitigations could look like or what risk profiles actually look like. And so a lot of what we talk about or a lot of the really interesting activity in the spaces around AI and biomedical science. And I think as the modern biotech space is evolving, like synthesizing DNA and applying these at scale, and we have all these automated labs that can do this kind of synthesis, sort of fellow travelers in the space are members of our coalition, the Institute for Progress just worked with Sensitive Bio to articulate how you can secure the DNA supply chain. Because there is a specific articulable risk of if you could synthesize DNA either autonomously or at scale, or it can empower bad Actor to synthesize DNA of a dangerous organism. That is understandably a concern, but that's also a very specific choke point that you can legislate around or implement guardrails or work with industry to address that specific risk. And there's good debates to be had about the appropriateness of specific controls and when and where to apply them and how intensely. But I think a lot of these risks are very tractable, solvable problems and it's just an issue of attention and willpower and where we actually decide to focus our efforts to drive progress on.

B

Yeah, I think, Austin, you wanted to say something?

A

Yeah. Well, there is a funny thing to this which is like I was meeting with a biotech company and I was talking about this risk with them and some of our concerns and some of the things we were thinking about from like mitigation perspective and how much weight and time he was like, I don't know, I don't think about it that much because I could already kind of kill everybody right now. And I'm like, well that A, doesn't make me feel better. But B, you're not going to, because the, a big bonus is that you're not going to do this. You understand it. And also like, I don't know, whatever, whatever the good scientist version of the, of the Hippocratic oath exists, you've at least taken it in your heart somewhere, you know. Although I feel like after watching the Fallout TV show and playing a lot of Fallout, I wouldn't say I trust scientists with being fully self governing, but I believe in them way more than random people in their basements that are trying to order stuff off cloud labs. But the main thing is just continuously keeping it in mind, right? And then second of all, keeping in mind that like we don't want to bring a knife to that gunfight either, right. It's like AI is going to be one of the most useful things in understanding where all of those choke points are and how to deal with it. Because there's going to be a desire to just say we will not use this technology or make it. It's like you can say that if you want, but we will use that technology and we will make it. And it will largely be for the, for the good as long as we do things properly. But at the same time you have to approach it with the clear eyes that are like, I would really, really, really like for us to be able to cure cancer. And I know the moment that we're able to do that, we will also be able to Kill everybody. Way easier, you know, in some way that I don't fully comprehend. Right. You have to make sure that you address both at the same time and understand that there's like upside. It's not like you're just doing the, like you will learn something useful. So like, if you think about the way AI itself works, like a lot of the biggest capability enhancements have been because of alignment techniques. Right? Like RLHF, instruction tuning. The things that make them what they are are all techniques for AI safety originally that are purely functional. And I think we're going to find the same thing to be true again and again and again. And especially in the SC scientific field, whatever we need to measure, whatever we need to watch, whenever we need to make, to ensure that like the primary goal is safe and effective will be relevant and important for other goals.

B

Yeah, I mean, I've always thought there, there have been some funny parts of this discussion. So one is like this idea of thinking about AI in isolation and not around supply chain. So especially in bio, there are like a lot of choke points that we, we could use policy mechanisms to strengthen. I think the other issue is, especially when it comes to knowledge generation, again, we often think about AI in isolation. And I think my personal view is the proper comparison should be what can you do with AI versus what can you do with all the other tools that are at your disposal? And that's the real marginal thing that we should be comparing against. And it turns out a lot of that information is accessible through things like search and textbooks and other kinds of material that's out there.

A

Yeah, and I'm glad people have done a lot more work and just studying this. Right? Like studying the nature of uplift and the nature of availability. I mean, I do think about this a lot on the margins too. Like, you know, okay, if there's now 0.01% of people that would find it feasible to take whatever action and they are antisocial, well, that's problematic. That's a huge number of people and we have to care and think about it. At the same time though, that's not like a, we shut down the, or control the Internet in like a crazy totalitarian way, because that also is not like a proportionate response. So the more that people continue to observe and do serious research, which I've seen some, like Abby O'Ver I think, just did a great little survey of this from the, the biorisk. The biorisk perspective. But like, that's just a place where serious work is very Warranted, and I hope people will 10x how much they're working on it.

C

And I will give a shout out to the biotech area, specifically in terms of a bright spot of policymakers recognizing the stakes of this space. The National Security Commission on Emerging Biotech has been doing incredible work for months, years now, and has been tremendously effective at getting ideas introduced. Temporary legislation in the past around how to make the biotech ecosystem more competitive and deliver the benefits, but also address these downsides. I think it is a very tired meme at this point that Congress doesn't get technology. That's often true, but there are really, really promising areas where they absolutely do and they're doing the right things and they're doing things strategically. Doesn't mean it's necessarily enough, but it does demonstrate that, yeah, we have smart people that can think hard about these problems and take steps to solve them in ways that really matter and make a difference. You know, we can scale that. Right? But like, it. It doesn't mean we shouldn't be looking towards progress.

B

All right, well, there's one last question that I feel like I had to ask as we wrap up. This has all been super informative, but the, the thing I really wanted to ask is that, you know, Seedai and you in particular have been big promotants of this idea that we should really think a lot more about sort of non humanoid form factors from ro and so you have this thing called the National Crab Robot Initiative. So tell me a little bit more like, what exactly is this? How seriously should we take this? And when will the first mass market crab robots arrive on the market so I could buy one?

C

What do you mean? How seriously should we take this? This is deadly serious. I mean, this is something that we've been very big proponents of for quite some time, where we do a lot of work on congressional education, on different AI issues. We do monthly events where it's like AI next topic. And everything we do on robotics, there's a ton of staff are interested. And so we were doing this a while ago and I think this has been a thing that Austin has long been frustrated with too. But just the conversation around, oh, are humanoids it? Is that the thing? The robotic perfect. They will matter. There's arguments prone against, but in evolutionary biology, there's a concept called carcinization where the crab form factor has evolved from like five plus different separate lines of evolution. It's like very like evolutionarily efficient. And I think that is a lesson for roboticists. We should be building robots in a crab form factor. Because if you think about it, right, like, if you watch all these demo videos of humanoid robotics and there's all these, like, compilations of them just like, falling down. I've never seen a crab fall down, and I think that is something we should take very seriously.

A

Never have you ever seen a crab fall down, Alec?

B

I can't say that I spent a lot of time observing crabs, but. But it's true. They have a low center of gravity.

A

I think. I think for me, a lot of this is just like we're a little too evolutionarily full of ourselves or something. You know, we like, imagine that we're peak cognition, we're peak physicality, and hopefully that's not true because I feel like both really could use a. A bit more intelligent design. But the main thing is, like, we're certainly not the most efficient thing that exists in the world. And even though we've designed the world for humanoids, there's just this weird fixation on one shotting everything in. In, I think, AI universe. That's part of why our work is a bit weird and different and messy, because we're going and talking to a thousand different little fractured pieces of it that are all different human and institutional structures, right? But people want to make the one thing that fixes everything. We want to make the one robot to rule them all. We want to make the one AGI that is the machine God, you know, and fixes all of it for us. And there's some piece of this that I think we're just very tired, you know, I think we're all quite tired. We could use a nap, you know, and like a new puppy that we don't have to take care of, but somebody else's, and we can just go play with it. But at the same time, the laziness is the truest existential threat, right? Like, our desire to call it a day and just have one thing that fixes it is weirdly dangerous, which I think is maybe the most serious part of this for me, except for everything else, which is also extremely serious.

C

My. My wife texted me saying, I hear you talking about crabs downstairs.

B

I. I think. I think you would say, right, that we should do both, right, like, not just focus on humanoids, but not just focus on other things. I think the logic for humanoids is there's like. Like a lot of our environment is built around people, and so that form factor is. Is useful for a lot of applications you can think about.

C

So I. I do want to push back on that like when we're talking about like value, like I, I think there, there's a, a pretty near future where there's a lot of, there's an abundance of human robots and they, they have their niche. But like, if we're talking about like high economic value applications for robotics, that those things are human shaped is just like a rounding error in terms of like what we could, the value we can extract, right? Like we reorganize factories for like 0.04% marginal efficiency for like workers to get around a warehouse. We should absolutely be doing that for whatever robotic capabilities look like, right? Like, you know, factory plants, operating rooms, all this kind of stuff, like, are designed for humans because like that's what we had at the time. And if we're talking about like whether it makes sense to invest in these kind of things or like we, we decide what the design of the world is and then if, if we have capabilities that make that calculus different, we should absolutely act on them.

A

And I think that like that the completely flat little. What, what are those things called? The robots. Little ones that even.

B

The vacuum robots.

A

Vacuum robots. But there's another one that looks like that. The moot boxes and Amazon warehouses, they're just like the same rough size and shape and they just pallet things around. I think that's probably the most economically productive non manufacturing robot that currently exists. You can't really measure what roombas do, I guess, except for their own sales price or how much free labor they've freed up or something. But the, the, the truest, maybe most important point to take is like, yeah, focusing on crabs is exactly as silly as focusing on humanoids. If you focus on just one thing, it's kind of silly. And I think if you look at now again, focusing on crabs is way less dumb than just focusing on humanoids because it's where we'll converge anyways, if we converge. But just go to national crab robotinitiative.com guys, read up for yourselves. You know, do your own research, make your own decisions. I'll post random drops on 4chan for you to follow over time.

B

All right, I think that's as good a place as any to wrap up. So thank you again for joining me. This was really informative and I'm really looking forward to seeing what CDAI is up to in the coming months and years.

A

Hopefully we'll be doing it together in a crab robot. A huge transportation crap. No, but in all seriousness, thanks so much for having us. Love to see you in the new role and just really really fun to podcast with you.

B

Yeah, it was great.

C

Thanks for having us.

B

Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform. Subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker and Matt Mand. See you next time.