01:35:31 · John Coogan, Zach Rash, Jordi Hays
[01:35:31] S1: Awesome. A lot of people in tech are worried about x risk these days. Hypothetically, where's the vulnerable point if I need to disable this with, like, a five five six round? Where would I shoot it? Can't tell [01:35:45] S8: you that, though, but I [01:35:46] S1: Oh, wow. He's on the side of the robots. This guy's traitor to humanity. [01:35:51] S8: Crazy. It... What's funny is that [01:35:54] S0: You need, like, a dead man switch. Okay. Let me test it out, bro. [01:35:57] S1: Send the delivery over. Send it over. You wanna test [01:36:00] S8: it out. Japan. Sick. In Japan, you actually have to have a big red kill switch button on the robot. [01:36:05] S1: Oh, okay. [01:36:06] S8: Which I feel like would uniquely only work in Japan. [01:36:09] S0: Yeah. Because otherwise, in the back emergency. Of [01:36:12] S1: Of course. But separately, jokes aside, I I have no doubt that, that you are feeling the benefits of general model progress and AI progress. Like, just take an image, identify if there's someone there, if you can go straight on this road. All that's probably happening. But what do evals, what do benchmarks actually look like? What are you measuring, and what are you seeing in those? Like, how do you know that the latest iteration of your software, of your AI stack is actually improving the experience? It feels like we sort of feel the failure mode. Like, the hallucination is like the viral video on Instagram of the Coco robot getting stuck on the curb. But how are you measuring that? What's progress been like?
01:36:12 · John Coogan, Zach Rash, Amiram Shachar
[01:36:12] S1: Of course. But separately, jokes aside, I I have no doubt that, that you are feeling the benefits of general model progress and AI progress. Like, just take an image, identify if there's someone there, if you can go straight on this road. All that's probably happening. But what do evals, what do benchmarks actually look like? What are you measuring, and what are you seeing in those? Like, how do you know that the latest iteration of your software, of your AI stack is actually improving the experience? It feels like we sort of feel the failure mode. Like, the hallucination is like the viral video on Instagram of the Coco robot getting stuck on the curb. But how are you measuring that? What's progress been like? [01:36:56] S8: Yeah. Well, it's it's interesting because I think in robotics, [01:36:59] S9: I mean, you see a [01:36:59] S8: lot of demos of folding clothes. And then in the self driving car world, you have a lot of road driving. Right? Yeah. In a very rule based environment. Yeah. The interesting thing about city sidewalks is they're the most complex, Yeah. Dynamic not rule based environment in the world, and every city is different. [01:37:15] S1: Yeah.
01:37:15 · John Coogan, Zach Rash
[01:37:15] S1: Yeah. [01:37:16] S8: Right? Navigating Downtown LA is is is is unique and different than Chicago, which is very different than Finland, which is very different than parts of Asia. And so you have different kind of social norms and behaviors. And so the robot actually needs to learn how to be a great, you know, user of the shared space with the public. It has to be a great sidewalk user. But we also need to get from point a to point b very quickly. So we can't spend a lot of time stopping. So there's this really complex kind of social navigational problem that we have to get very good at. So I think we have, you know, we have millions of hours and millions of miles of driving in these sort of environments, you know, snows, sun, you know, weekend nightlife, all these different sorts of environments. So we probably have one of the largest datasets of this sort of data, and and that's used to train all of our models. But then we've actually open sourced and published some really interesting benchmarks, because I think as these models progress, I think some of these big labs might want to be testing, hey, how did their model actually do at navigating the the public realm? How do we actually do it navigating, you know, a Downtown Chicago sidewalk? Yeah. And we actually have some benchmarks we've created from these different environments that we've published, and I think that will be helpful to the robotics industry in general. [01:38:31] S1: Very cool. Well, congratulations on the progress. Thank you so much for coming on the show. Jordi, sorry. Do you have any
01:38:31 · John Coogan, Jordi Hays, Zach Rash
[01:38:31] S1: Very cool. Well, congratulations on the progress. Thank you so much for coming on the show. Jordi, sorry. Do you have any [01:38:35] S0: Yeah. Excited to see more of these around. It's it's an interesting thing. Like, you got it... You wanna... As a business, you wanna scale quickly. But I think, like, the the lesson from flock is that, like, individual consumers have so much power Sure. Over over, like what's allowed Yeah. To operate in their area. And I think that, making sure that, you know, friendly robots are are never annoying is has to be like number one priority. And and that's such a hard problem versus what you were talking about with like on the roads. It's like we have established laws. When the light's green, go. Yeah. Oftentimes, if I'm behind John, I have to honk after the light goes green. It's not getting any greener, buddy. But anyways, very difficult problem. And, yeah, I wanna see the Coco robot in Japan. I promise not to touch the big red button. [01:39:24] S1: Yeah. I I can I can be trusted? I imagine you hiring a couple of these just to patrol the property. This feels like an intermediary. If you if you don't wanna go robot dog patrol and humanoids aren't working yet, just having a a fleet of Coco's circling the property at all times, that's security. I like it. [01:39:42] S8: Guys, how do we get some lunch orders going to the office, going to the studio? [01:39:45] S1: Well, we'll coordinate. Nick will text you right now. We'll figure it out. [01:39:48] S0: We'll Yeah. We'll do it.
01:39:48 · Jordi Hays, John Coogan, Zach Rash, S2, Ben Liu
[01:39:48] S0: We'll Yeah. We'll do it. [01:39:48] S1: We'll get our lunch delivered with Coco. Thank you so much. Yeah. Great to [01:39:51] S8: hang out. Thanks, guys. [01:39:52] S0: Congrats on the rest day. Of [01:39:53] S1: See you. We do have the phone toss replay live on TBPN ready. Let's see it in action. Jordi [01:40:01] S0: Can we get a little zoom in? Jordi [01:40:04] S1: Bone. Oh, that that... [01:40:06] S0: But I really wanna see Tyler's. [01:40:07] S1: Held up well. [01:40:08] S0: Because Tyler was [01:40:09] S1: Tyler comes in. [01:40:10] S0: Was bolder because he thought I was gonna catch him. Yeah. What's Let's zoom in. Let's zoom in. [01:40:15] S1: Okay. Got a little box. [01:40:16] S0: Let's see what's happening. [01:40:17] S1: Can we see this? [01:40:18] S0: Throws it. Oh, gets out of the way. [01:40:20] S2: Yeah. When Jordi moves back. [01:40:23] S1: You fell for that one, Tyler. You fell for it hook, line, and sinker. Our next guest is the cofounder and CEO of Formation Bio with Ben Liu, launching a huge drug development fund. Ben, welcome. [01:40:36] S8: What's happening? [01:40:37] S1: Let's get this going ready. How are doing? Please kick us off [01:40:40] S6: with Thanks for [01:40:40] S1: having us. Yourself and the organization.
01:40:40 · John Coogan, Ben Liu, Jordi Hays
[01:40:40] S1: having us. Yourself and the organization. [01:40:44] S6: Great. So Benjamin Liu, co founder of CF Formation Bio. We're an AI native company at at building the next generation pharma company with this kind of core thesis that drug discovery is getting more and more efficient, and the fundamental bottleneck of bringing new medicines to patients is the cost and time of drug development. [01:41:01] S1: What's your stance on monkeys? What's your stance on monkeys? You got There's a Joel Lonsdale. [01:41:05] S6: There's gonna be a big shortage. [01:41:06] S0: There is. [01:41:07] S6: They're important. You know? That national security, we gotta increase the number of monkeys we have. It's, you know, the the similarity between our disease systems to kind of humans is probably one of the best predictors. Yep. Right? So you can imagine, like, solving, like, rat Alzheimer's is very different than sucky... Solving monkey Alzheimer's. Yeah. So But, you know, the biggest the biggest bottleneck is drug development in humans. And so, you know, you have this impedance mismatch where drug discovery is only getting more efficient. There's been a doubling of number of discovered drug candidates in the past ten years, even before AI. The number of approved drugs has kept, like, kind of constant 50 a year. Markets are rational, so discovered drugs aren't worth much till post phase two. Sure. So the company, we actually don't do any discovery internally. We look to acquire drugs that are discovered. [01:41:52] S1: Okay.