WEBVTT - Josh Wolfe: The ChatGPT of Robotics is Coming 0:00:03.120 --> 0:00:08.360 Bloomberg Audio Studios, Podcasts, Radio News. 0:00:20.200 --> 0:00:23.640 Hello and welcome to another episode of the Odd Lots Podcast. 0:00:23.720 --> 0:00:26.040 I'm Joe Wisenthal and I'm Tracy Alloway. 0:00:26.200 --> 0:00:27.880 Tracy, let's talk about AI some more. 0:00:29.960 --> 0:00:32.760 Okay, Well, we could just have AI write the script 0:00:32.760 --> 0:00:34.360 for us. We could give ourselves some time. 0:00:34.479 --> 0:00:36.360 No, I don't think the technology is there yet, you 0:00:36.400 --> 0:00:39.680 know what, so I can't say who. Maybe I was 0:00:39.720 --> 0:00:43.000 trying to a professor recently and she said something really 0:00:43.040 --> 0:00:45.720 interesting to me. I'm not supposed I'm not sure it's fine, 0:00:45.720 --> 0:00:48.279 I think, And she said, like, you know, there's all 0:00:48.280 --> 0:00:51.879 this anxiety about you know, kids cheating on their essays 0:00:51.960 --> 0:00:55.279 or having Chad g GBT write their essays for them, 0:00:55.320 --> 0:00:58.120 and like, you know, supposedly professors are like tearing their 0:00:58.120 --> 0:01:00.400 hair out trying to figure out what to do about this, 0:01:00.480 --> 0:01:03.200 and the AI detector don't really work all that well particularly, 0:01:03.560 --> 0:01:06.399 but apparently like it seems like the solution is to 0:01:06.560 --> 0:01:09.800 just grade them as regular essays no matter what. And 0:01:09.840 --> 0:01:12.600 it sounds like at this point all the chat GPT 0:01:13.040 --> 0:01:16.160 essays are basically solid c essays, and so if you 0:01:16.200 --> 0:01:18.039 just sort of like take them at phase value, even 0:01:18.080 --> 0:01:20.920 if you think they might be AI generated at least 0:01:20.959 --> 0:01:23.160 at this point, it doesn't seem like a way at 0:01:23.200 --> 0:01:25.960 least for college students to write good essays. 0:01:25.760 --> 0:01:29.520 Yet, so our baseline, our average is chat GPT. 0:01:29.680 --> 0:01:32.080 Now, yeah, that's basically can you beat the bot? 0:01:32.240 --> 0:01:35.560 Did you see the thing someone was tweeting about one 0:01:35.640 --> 0:01:39.200 of the tells for AI generated words is if you 0:01:39.400 --> 0:01:42.119 use the word delve. Yes, I saw that, which, as 0:01:42.160 --> 0:01:45.600 someone who I'm sure has used delve numerous times on 0:01:45.640 --> 0:01:47.640 this podcast and in my writing, I thought was a 0:01:47.680 --> 0:01:50.320 little unfair. It is a little it's a cliche, but 0:01:50.520 --> 0:01:52.520 that doesn't mean it's from AI. 0:01:52.720 --> 0:01:55.960 Although being said, there's more to AI than just chat 0:01:56.000 --> 0:01:59.160 GPT obviously in chatbots, and this sort of has come 0:01:59.240 --> 0:02:02.840 up on a couple episodes recently, but only like very tangentially. 0:02:03.240 --> 0:02:06.840 People talk about the use of like AI and industrial applications, 0:02:06.880 --> 0:02:08.280 and I've seen a lot of stuff. There have been 0:02:08.280 --> 0:02:12.760 a couple Bloomberg articles about some startups that sort of say, like, okay, well, 0:02:12.800 --> 0:02:15.800 like what if we trained robots the same way we 0:02:15.880 --> 0:02:18.480 trade large language models, where you feed them just like 0:02:18.680 --> 0:02:21.280 tons and tons and tons of real world data, and 0:02:21.360 --> 0:02:23.360 so that yeah, like you sure you still have to 0:02:23.360 --> 0:02:26.600 solve the mechanical engineering part. But then what if that 0:02:26.760 --> 0:02:30.080 allows them, like all this training data to do more 0:02:30.280 --> 0:02:34.760 advanced industrial things like I don't know, like make a 0:02:34.800 --> 0:02:38.560 pizza or you know, be a more powerful humanless assembly 0:02:38.600 --> 0:02:39.079 line or. 0:02:39.040 --> 0:02:39.840 Something like that. 0:02:40.240 --> 0:02:42.800 Where it's like we see all these oppressive robots and 0:02:42.880 --> 0:02:45.680 videos like Boston Dynamics, but I never know like if 0:02:45.720 --> 0:02:48.000 any of this is like quite there yet in terms 0:02:48.000 --> 0:02:49.000 of having its value. 0:02:49.160 --> 0:02:52.679 Yeah. So the robotics aspect of AI is something that's 0:02:52.800 --> 0:02:55.799 incredibly interesting to me. It kind of makes me think 0:02:55.960 --> 0:02:58.760 about the world that we want to see. So it 0:02:58.800 --> 0:03:02.280 would be great if if we had physical robots who 0:03:02.320 --> 0:03:04.799 are able to do stuff like clean up a house 0:03:05.120 --> 0:03:08.480 or take care of an elderly family member or something 0:03:08.560 --> 0:03:11.320 like that. It's not so great if all of our 0:03:11.320 --> 0:03:15.959 technological prowess basically goes into writing satirical lyrics. Yes, via 0:03:16.080 --> 0:03:19.000 chat GPT, like that's fun. I can do that myself, 0:03:19.400 --> 0:03:22.600 But what I really need is someone to vacuum or 0:03:22.680 --> 0:03:23.600 dust the house. 0:03:23.520 --> 0:03:24.320 Do the laundry. 0:03:24.440 --> 0:03:26.760 Yeah, exactly, really nice. All right, Well, I just want 0:03:26.760 --> 0:03:29.400 to jump right into it because we really do have 0:03:29.680 --> 0:03:31.800 the perfect guest. We're going to be speaking to someone 0:03:31.800 --> 0:03:34.960 who has been investing in AI for a long time. 0:03:35.280 --> 0:03:39.040 A lot of vcs like started investing in AI last 0:03:39.120 --> 0:03:42.560 year obviously, but this is someone who has been investing 0:03:42.560 --> 0:03:45.360 in AI for quite some time before it became the 0:03:45.360 --> 0:03:48.360 hot new thing. We spoke to him last July and 0:03:48.440 --> 0:03:50.560 had a great conversation about what he was seeing in 0:03:50.600 --> 0:03:53.400 the space. So I'm really pleased to welcome back on 0:03:53.440 --> 0:03:55.240 the show. We're going to be speaking with Josh Wolfe, 0:03:55.560 --> 0:03:59.360 co founder and managing partner of Lux Capital. Josh, thank 0:03:59.360 --> 0:04:01.120 you so much for coming back on ODLTS. 0:04:01.120 --> 0:04:03.440 Great to be on. I feel like I should say 0:04:03.480 --> 0:04:04.920 hello and like a robot voice. 0:04:06.560 --> 0:04:08.840 So what's interesting to you these days? What are you 0:04:08.840 --> 0:04:10.360 seeing out there that gets you excited? 0:04:10.720 --> 0:04:12.440 Well, you know, you guys started this off with AI 0:04:12.720 --> 0:04:15.960 and with an AI. Look, we've had what I would 0:04:15.960 --> 0:04:19.120 call maybe a little bit pejorative, a little bit looting crude. 0:04:19.440 --> 0:04:21.880 We've had chips everybody knows that we talked last time 0:04:21.880 --> 0:04:25.320 about in AMD. We've got chatbots. You already have some 0:04:25.360 --> 0:04:27.479 of these guys that are starting to fail. They've raised 0:04:27.480 --> 0:04:30.680 billions of dollars and in some cases just you know, 0:04:30.800 --> 0:04:35.720 relatively undifferentiated, big debate between open source that is approaching 0:04:35.760 --> 0:04:39.480 the ASSYMP tote of achievement that the big private models 0:04:39.520 --> 0:04:42.799 have and then being a little bit lout you've got chicks. 0:04:42.880 --> 0:04:44.719 What do I mean by that? Most of the applications 0:04:44.720 --> 0:04:47.760 in AI are the mundane and passe on the one side, 0:04:47.760 --> 0:04:51.520 which would be like customer service and basic call center 0:04:52.000 --> 0:04:55.400 supplementation or substitution. And then at the other end you 0:04:55.440 --> 0:04:57.400 have people that are spending tens of thousands of dollars 0:04:57.440 --> 0:05:00.599 a month in some case on AI girlfriends and people 0:05:00.640 --> 0:05:02.800 that are doing what they often do with technology which 0:05:02.800 --> 0:05:05.840 is used for prurient interests. So that to me the 0:05:05.880 --> 0:05:08.400 two barbel extremes in AI, where people are actually making 0:05:08.440 --> 0:05:12.640 money and profits serving demand for basic human instincts I 0:05:12.640 --> 0:05:16.680 guess and needs. That's interesting. Overall, you're seeing a big 0:05:16.680 --> 0:05:19.719 shift from the compute piece to the energy piece, meaning 0:05:19.960 --> 0:05:24.400 people now recognize this bottleneck in AI is not going 0:05:24.440 --> 0:05:25.880 to be so much about the chips. We also talked 0:05:25.880 --> 0:05:28.800 about this months ago when we said, look, you don't 0:05:28.839 --> 0:05:31.960 necessarily need these in video chips for inference, the part 0:05:32.000 --> 0:05:33.960 that most people do when they're querying all these models. 0:05:33.960 --> 0:05:37.000 You do need them for training, but the power levels 0:05:37.040 --> 0:05:39.599 on these things are just enormous. There was a Dell 0:05:39.720 --> 0:05:42.960 earnings call that they think sort of accidentally leaked that 0:05:43.040 --> 0:05:45.720 this in VideA B one hundred Blackwell chip is going 0:05:45.760 --> 0:05:48.320 to be a thousand wide draw of power, which is 0:05:48.360 --> 0:05:51.039 like forty fifty percent more than these eight one hundred chips. 0:05:51.560 --> 0:05:53.840 Why does that matter because now you got to figure 0:05:53.880 --> 0:05:56.279 out how do you supply the energy for that. And 0:05:56.320 --> 0:05:59.880 a tidbit which I thought was interesting was Amazon acquired 0:06:00.120 --> 0:06:03.520 a nuclear powered data center in Pennsylvania. They've spent six 0:06:03.560 --> 0:06:06.080 hundred and fifty million dollars, they got about a gig 0:06:06.200 --> 0:06:08.680 lot of power. And I think that that is going 0:06:08.720 --> 0:06:09.880 to be a trend. I think it's actually going to 0:06:09.960 --> 0:06:11.840 usher in a wave of what I like to call 0:06:11.880 --> 0:06:14.880 out elemental energy, but demand for nuclear power to power 0:06:14.920 --> 0:06:16.240 these big AI data centers. 0:06:16.480 --> 0:06:18.160 Yeah, it's kind of funny when you think about it. 0:06:18.200 --> 0:06:20.920 I don't think anyone expected uranium to end up being 0:06:20.960 --> 0:06:23.680 an AI play, but here we are. I want to 0:06:23.720 --> 0:06:26.280 go back to something you said, and you actually brought 0:06:26.279 --> 0:06:28.960 it up the last time we spoke to you last year, 0:06:29.240 --> 0:06:31.800 and it was the idea of I guess the novelty 0:06:32.160 --> 0:06:36.200 of some of these more public facing AI projects, And 0:06:36.360 --> 0:06:39.119 I think you pointed out last year it's really fun 0:06:39.200 --> 0:06:42.080 to use some of these things, like generate a bunch 0:06:42.080 --> 0:06:45.160 of cartoon versions of yourself or whatever, but it might 0:06:45.200 --> 0:06:47.480 not be a sustainable business model, and it might end 0:06:47.560 --> 0:06:51.920 up being a functionality that is eventually incorporated into another 0:06:52.000 --> 0:06:55.520 platform or a different project. Have you seen any example, 0:06:55.680 --> 0:07:01.080 specific examples of more public facing novelty AI start to 0:07:01.320 --> 0:07:04.039 like go away? I think you mentioned a few failures recently. 0:07:04.400 --> 0:07:06.280 Yes, You've had a whole bunch of companies that basically 0:07:06.320 --> 0:07:09.200 took chat, GPTA, GPT three or four and put a 0:07:09.200 --> 0:07:12.440 wrapper around it, basically meaning give the average user who 0:07:12.480 --> 0:07:14.720 doesn't know how to use these or even do prompts, 0:07:14.760 --> 0:07:17.080 some means to interact with it. And those things raised 0:07:17.120 --> 0:07:18.720 a bunch of money. They made these things accessible, and 0:07:18.720 --> 0:07:21.400 they've sort of gone away. The foundation models themselves that 0:07:21.440 --> 0:07:24.560 are undergirding all of this themselves are also starting to 0:07:24.560 --> 0:07:27.640 be relatively commoditized away. And some of these things have 0:07:27.760 --> 0:07:29.560 prominent people and have raised a lot of money, but 0:07:29.600 --> 0:07:32.640 they are what I would consider failing. Take inflection. You've 0:07:32.600 --> 0:07:35.360 got Mustapha Solimon, super smart guy, co founder of deep 0:07:35.400 --> 0:07:37.640 mind raised I think a billion and a half for 0:07:37.680 --> 0:07:39.480 that company. You know, I'm gonna be careful here because 0:07:39.520 --> 0:07:43.000 Microsoft has been probably the savviest actor in this entire game, 0:07:43.320 --> 0:07:45.880 figuring out that they can acquire things by doing things 0:07:45.920 --> 0:07:48.480 in a clever way, skirting FTC and DJ oversight. They 0:07:48.480 --> 0:07:51.000 effectively control open ai. As I think we also talked 0:07:51.000 --> 0:07:53.480 about heard, particularly going into the end of last year 0:07:53.520 --> 0:07:55.960 when there was all the drama around open Ai. Sadia said, look, 0:07:56.160 --> 0:07:58.120 if open ai went out of business, we own it, 0:07:58.200 --> 0:08:00.600 control it. We've got all the data up left, right center, 0:08:00.680 --> 0:08:03.320 you know, all around them. Same thing with this company Inflection. 0:08:03.400 --> 0:08:04.840 They did I think a six hundred and fifty six 0:08:04.920 --> 0:08:07.840 hundred and seventy five million dollar license for the technology 0:08:08.240 --> 0:08:10.480 that basically was a payment above and beyond what the 0:08:10.640 --> 0:08:13.360 venture investors had made. Venture investors made a little bit 0:08:13.400 --> 0:08:16.320 of money, not a lot. Key management went over to Microsoft. 0:08:16.360 --> 0:08:19.000 But Microsoft has been very clever. So back to your question, 0:08:19.440 --> 0:08:20.920 I think the big are going to get bigger and 0:08:20.960 --> 0:08:25.120 are going to be most of the beneficiaries here, Microsoft, Adobe, Amazon, 0:08:25.280 --> 0:08:28.000 Amazon themselves. Coming up on the one year anniversary of Bedrock, 0:08:28.040 --> 0:08:30.240 they're going to announce that they've got the best performing 0:08:30.360 --> 0:08:33.320 model with Anthropic, which they've made billions of dollars of 0:08:33.360 --> 0:08:36.200 investments in now sort of competitor to Opening Eye CHATGPT. 0:08:36.280 --> 0:08:38.400 They're also going to renounce something with one of our 0:08:38.440 --> 0:08:41.240 companies that hasn't yet been publicly disclosed in biology. That's 0:08:41.280 --> 0:08:44.000 going to be one of the two biggest waves next 0:08:44.040 --> 0:08:47.719 biology and what you started the conversation with robotics. If 0:08:47.720 --> 0:08:51.319 you just take robotics as an example, I think in 0:08:51.960 --> 0:08:53.839 one of our companies, Hugging Face, which is one of 0:08:53.880 --> 0:08:56.880 the main repositories for all these open source models, there's 0:08:56.920 --> 0:09:01.440 something like sixty thousand text generations models. You know, it's 0:09:01.440 --> 0:09:03.720 like fifty nine thousand, seven hundred something now, but just 0:09:03.760 --> 0:09:06.040 an enormous number of text generation models. This is in 0:09:06.120 --> 0:09:08.520 like two or three. It's like everybody's doing this. Everybody's 0:09:08.520 --> 0:09:10.280 trying to do it all, basically trying to predict the 0:09:10.280 --> 0:09:13.760 next word based on the prior one. What this transformer technology, 0:09:14.160 --> 0:09:17.520 which was invented at Google, ended up parlaying into guess 0:09:17.520 --> 0:09:20.240 how many robotic models there are? Fifty nine thousand in 0:09:20.280 --> 0:09:22.200 text generation? Guess how many robotic models there are? 0:09:22.520 --> 0:09:23.120 A fraction. 0:09:24.559 --> 0:09:27.600 I mean, I'm obviously leading with my phone are but 0:09:28.360 --> 0:09:31.319 nineteen robotic models. Okay, so you got that to me. 0:09:31.960 --> 0:09:34.040 As a venture investor, we're just always looking at where's 0:09:34.040 --> 0:09:39.200 their abundance, where's their scarcity? Their scarcity of robotic models? Now, why, Well, 0:09:39.240 --> 0:09:42.920 it's relatively easy to train on the open Internet. You've 0:09:42.920 --> 0:09:45.400 got Wikipedia, you've got YouTube videos. You know, whether you're 0:09:45.400 --> 0:09:47.120 not you're supposed to be doing that or not. Like 0:09:47.120 --> 0:09:50.000 the woman who was asked it, soa hey, how did 0:09:50.040 --> 0:09:51.560 you train these things? Was it on YouTube? And she 0:09:51.600 --> 0:09:54.280 gave it and you probably saw that, So there's gonna 0:09:54.280 --> 0:09:56.400 be all kinds of copyright stuff on that. Robots are hard. 0:09:56.440 --> 0:09:59.400 Why Most of the robotic stuff that has been out 0:09:59.400 --> 0:10:01.480 in the world, like you talked about in your intro, 0:10:02.200 --> 0:10:06.320 is constrained in manufacturing facilities, in work cells on an 0:10:06.320 --> 0:10:11.600 assembly line, very specific, parametrically constrained, so very few degrees 0:10:11.600 --> 0:10:13.960 of freedom of what they're actually doing. The robots themselves 0:10:14.040 --> 0:10:16.520 might have multi access scripts and controllers, but they're not 0:10:16.559 --> 0:10:19.880 moving around very freely. You have exceptions with like Amazon 0:10:20.120 --> 0:10:24.200 which acquired Kiva, moving the warehouse inventory stuff around, but 0:10:24.280 --> 0:10:30.360 again relatively xyz access not unstructured environments. You and I 0:10:30.840 --> 0:10:33.880 and our listeners. We all thrive every day in unstructured 0:10:33.960 --> 0:10:38.240 environments and that is where you need enormous training data. 0:10:38.440 --> 0:10:40.280 You can't search the internet for that, So how do 0:10:40.320 --> 0:10:43.079 you do it. There's a few things that have emerged, 0:10:43.080 --> 0:10:45.920 and you mentioned some articles. We funded a company that 0:10:45.960 --> 0:10:48.400 recently came out of STELL called Physical Intelligence stead of 0:10:48.480 --> 0:10:51.559 artificial intelligence, Physical Intelligence, and it is the krem Dela 0:10:51.600 --> 0:10:54.320 Creme team from Stanford and Berkeley. You've got some open 0:10:54.360 --> 0:10:57.240 ai folks, You've got Google Deep Mind folks. They took 0:10:57.240 --> 0:10:59.600 investment from open ai US and a bunch of other 0:10:59.679 --> 0:11:03.560 vcs and they are just twenty four to seven training 0:11:03.640 --> 0:11:07.360 robots doing all kinds of crazy things like folding, laundry, pouring. 0:11:07.440 --> 0:11:12.240 Determined but to let these robots encounter unstructured environments and 0:11:12.280 --> 0:11:15.480 then be able to thrive in them. The next thing 0:11:15.480 --> 0:11:18.480 that you're going to see are visual models where you're 0:11:18.520 --> 0:11:21.760 effectively giving, like an ikea sketch or you're drawing something, 0:11:21.840 --> 0:11:24.800 and you're being able to instruct the robot to have 0:11:24.840 --> 0:11:27.439 a sense of intuitive physics of how the world works 0:11:27.440 --> 0:11:29.559 and how things might connect to each other and then 0:11:30.120 --> 0:11:33.360 learn from that. And then we're also training these robots 0:11:33.360 --> 0:11:36.560 with simple verbal cues. So there's a video you can 0:11:36.600 --> 0:11:39.440 see online from some of the researchers where they are 0:11:39.480 --> 0:11:42.560 picking nuts and m and ms and separating them, you know, 0:11:42.640 --> 0:11:44.440 just as a task of being able to sort and 0:11:44.480 --> 0:11:47.480 filter with precision and dexterity, and if they picked a 0:11:47.480 --> 0:11:50.120 wrong one, you can actually instead of physically grabbing the things, 0:11:50.120 --> 0:11:53.760 say stop grab the Eminem's not the nuts. And now 0:11:53.800 --> 0:11:57.040 it knows that. So I think that we're about to 0:11:57.120 --> 0:12:01.000 unleash in robotics what will become a chat GPT like 0:12:01.080 --> 0:12:03.439 moment where people are so used to seeing robots and 0:12:03.480 --> 0:12:05.679 they see the arms, and they've seen West World and 0:12:05.720 --> 0:12:08.600 this kind of stuff, and suddenly something happens that just 0:12:08.679 --> 0:12:11.240 blows your mind. And I think that's coming soon. 0:12:27.120 --> 0:12:29.960 That's pretty exciting because, like I said, you know, for 0:12:30.120 --> 0:12:33.000 like at least a decade, I've been watching those Boston 0:12:33.480 --> 0:12:36.440 whatever online, those youtubes, and at this point I'm sort 0:12:36.440 --> 0:12:38.960 of convinced that it's like basically a content generator because 0:12:38.960 --> 0:12:41.280 it never seems like they're like crazy robot dogs or 0:12:41.320 --> 0:12:44.040 anything like ever become commercial. But maybe this is the 0:12:44.040 --> 0:12:46.880 missing link. But you've brought up like three different avenues 0:12:46.920 --> 0:12:48.400 we could go on and I want to sort of 0:12:48.400 --> 0:12:51.120 eventually hit on any of them. Here's a specific question, 0:12:51.120 --> 0:12:53.920 and then we can maybe get back to robotics. This 0:12:54.080 --> 0:12:57.760 element where there is such a shortage of advanced, cutting 0:12:57.880 --> 0:12:59.840 edge talent, people who really know how to do this. 0:13:00.200 --> 0:13:02.960 And you mentioned that guy that got hired by Microsoft 0:13:03.000 --> 0:13:07.440 from his other companies as an investor in AI or 0:13:07.559 --> 0:13:11.160 robotics startups. Is this a dynamic that's different than in 0:13:11.280 --> 0:13:15.400 other software or other tech investing. Basically, this sort of 0:13:15.440 --> 0:13:18.520 like highly skilled tech qman risk. 0:13:18.600 --> 0:13:22.640 Basically yes, in that you always are looking for what's 0:13:22.640 --> 0:13:25.040 scarce and you want scarce talent. If anybody could do this, 0:13:25.120 --> 0:13:27.920 it's just not that valuable companies would get funded. Vcs 0:13:27.920 --> 0:13:30.440 would fund forty of them at the same time, or 0:13:30.520 --> 0:13:33.120 maybe four hundred of them. Contrast to things where it's 0:13:33.240 --> 0:13:35.840 very web based or the old groupons of yesterday. This 0:13:35.920 --> 0:13:40.160 is highly technical, often PhD scientists. The vast majority of 0:13:40.200 --> 0:13:43.080 the founding teams that we've backed at companies like Covariant 0:13:43.360 --> 0:13:46.760 or format or this new company Physical Intelligence, they're all 0:13:46.840 --> 0:13:50.319 PhDs that are coming out of Stanford, Carnegie, Mel and MIT. 0:13:50.800 --> 0:13:53.640 Some of the best robotics programs in the world, and 0:13:53.840 --> 0:13:57.160 there's lineage of these great professors. Many of them have passed, 0:13:57.160 --> 0:13:59.320 but for example, there's this one guy, Hans Morvec used 0:13:59.320 --> 0:14:00.679 to be a carnegiemail and I got to meet him 0:14:00.679 --> 0:14:02.199 when he was still alive, but he was one of 0:14:02.240 --> 0:14:04.880 the very early pioneers in robotics. And he's got this 0:14:05.000 --> 0:14:09.199 paradox that insiders in the robotics world called the Morvek paradox, 0:14:09.520 --> 0:14:12.320 which is this weird counterintuitive phenomenon which is basically like 0:14:12.679 --> 0:14:15.120 all the stuff that we think is really hard is 0:14:15.120 --> 0:14:17.920 actually pretty easy for AI, and all the stuff that 0:14:17.960 --> 0:14:20.760 we find totally intuitive and easy, like riding a bike, 0:14:21.000 --> 0:14:24.440 that's really hard for robots. So there's this great paradox 0:14:24.480 --> 0:14:26.560 that some of the most brilliant researchers are working on, 0:14:27.400 --> 0:14:29.040 which is how do we do the kind of stuff 0:14:29.040 --> 0:14:31.280 that a four year old can do very intuitively with 0:14:31.360 --> 0:14:33.560 these very complex, expensive machines. And there's all kinds of 0:14:33.560 --> 0:14:35.760 considerations we could talk about about where are these arms 0:14:35.840 --> 0:14:39.200 coming from? The acquisitions that China has been making from 0:14:39.240 --> 0:14:41.520 what historically was a lot of German companies. I mean 0:14:41.520 --> 0:14:43.760 when I say arms meeting the robotic arms they can 0:14:43.840 --> 0:14:46.760 move thing. And then there's this great philosophical debate that 0:14:46.760 --> 0:14:48.600 it hasn't yet come to four. But I believe will 0:14:49.000 --> 0:14:51.120 and investors are sort of lining up on this. I'm 0:14:51.160 --> 0:14:55.680 on the opposite side of some people are funding humanoid robots. 0:14:56.120 --> 0:14:57.600 And the reason that I say I'm on the opposite 0:14:57.600 --> 0:14:59.640 side of it is I don't really believe in them. Yes, 0:15:00.040 --> 0:15:01.480 you would want somebody to help take care of your 0:15:01.520 --> 0:15:04.920 grandma and maybe provide some companionship. But this idea of 0:15:04.960 --> 0:15:08.960 the movies of these ex machina kind of robots that 0:15:09.160 --> 0:15:13.800 embody a human form. We know that engineering is better 0:15:13.920 --> 0:15:16.840 than evolution. If we were inventing a car tomorrow, it 0:15:16.840 --> 0:15:19.400 would be a terrible idea to take Fred Flintstone and 0:15:19.520 --> 0:15:22.280 use his feet, you know, to power these stone wheels. 0:15:22.320 --> 0:15:24.640 We know that an actuator and the axle and an 0:15:24.640 --> 0:15:27.680 engine or just better, and evolution didn't create that. Why 0:15:27.680 --> 0:15:31.480 would we create these humanoid hands where if I'm twisting 0:15:31.520 --> 0:15:33.640 the cap off of a bottle, you know I can 0:15:33.720 --> 0:15:35.440 only I have to turn my hand like seven times 0:15:35.440 --> 0:15:38.600 to do that. Whereas if I was just designed the 0:15:38.600 --> 0:15:40.800 perfect robot, I'd have a little suction cup that would 0:15:40.800 --> 0:15:42.640 go on top. It'd have like a drill bit mechanism. 0:15:42.680 --> 0:15:44.720 It would quickly twist it off and then it would 0:15:44.920 --> 0:15:47.360 Swiss army knife, you know, swap out for the next 0:15:48.280 --> 0:15:52.280 technical gripper capability. So I think that people are misguided 0:15:52.800 --> 0:15:54.600 and they're basically going to end up doing things for 0:15:54.680 --> 0:15:59.040 like prosthetics or you know, something that's sort of Westworld like. 0:15:59.480 --> 0:16:02.120 But I think the practical robots that we're going to 0:16:02.200 --> 0:16:03.680 all be using in our homes are going to look 0:16:03.680 --> 0:16:05.160 nothing like these humanoid robots. 0:16:05.360 --> 0:16:08.320 This is funny. This is very reminiscent of a weird 0:16:08.400 --> 0:16:10.920 conversation I used to have with my dad. He had 0:16:11.040 --> 0:16:16.040 like some sort of bugbear around the shape of aliens hands, 0:16:16.400 --> 0:16:18.640 and he was like, why are they always shown or 0:16:18.720 --> 0:16:24.040 depicted in these illustrations as having like human like hands 0:16:24.160 --> 0:16:26.760 or sometimes even three fingers, Like why wouldn't they have 0:16:26.960 --> 0:16:32.520 just evolved to the next level of very very efficient physiology. Anyway, 0:16:32.800 --> 0:16:34.920 One thing I wanted to ask, and I'm trying to 0:16:34.920 --> 0:16:37.560 think how to frame this question or what the right 0:16:37.600 --> 0:16:43.320 word is, but how open source is robotics in the 0:16:43.360 --> 0:16:46.720 sense of, like how much of the technology is shareable 0:16:47.000 --> 0:16:49.600 or replicable? Because I feel like one of the reasons, 0:16:49.640 --> 0:16:51.360 and you touched on it earlier, but one of the 0:16:51.360 --> 0:16:55.280 reasons we have seen this boom in AI is because 0:16:55.320 --> 0:16:58.880 you can go to places like hugging Face and download 0:16:59.000 --> 0:17:02.160 a bunch of code, open source code, and build off 0:17:02.200 --> 0:17:05.919 of it and it sort of multiplies around itself. But 0:17:06.080 --> 0:17:09.159 is there any aspect of that all in robotics or 0:17:09.200 --> 0:17:11.119 is it just much more proprietary. 0:17:11.520 --> 0:17:15.040 The hardware piece has historically been very proprietary, although there's 0:17:15.080 --> 0:17:18.120 lots of knockoff things. There is a Chinese company which 0:17:18.160 --> 0:17:20.480 is increasingly dominating the field. A lot of people don't 0:17:20.480 --> 0:17:23.960 know the name called Unitry Uni Tree Unitry that is 0:17:24.000 --> 0:17:27.200 sort of copying the Boston Dynamics robots that Joe was 0:17:27.240 --> 0:17:29.720 talking about and that you see in Black Mirror episodes 0:17:29.880 --> 0:17:33.560 and those kinds of things. On the software side, it 0:17:33.720 --> 0:17:38.000 really because it has the same kernel, the same origins 0:17:38.400 --> 0:17:41.840 as many of the AI software things that led to 0:17:41.920 --> 0:17:45.600 large language models and came from transformers academic roots and 0:17:45.640 --> 0:17:48.080 academics like to share and publish, and of course you 0:17:48.119 --> 0:17:50.880 can patent certain things, but by and large, the early 0:17:51.720 --> 0:17:55.560 systems something called ROSS, as you might guess, robot operating system. 0:17:55.840 --> 0:17:58.159 People that were doing something called r duino, which is 0:17:58.520 --> 0:18:02.119 sort of for hobby programmers with hardware and software at 0:18:02.160 --> 0:18:06.119 the intersection Poppy. There's a handful of these things. But again, 0:18:06.200 --> 0:18:08.480 now you're in this mode where you need to find 0:18:08.560 --> 0:18:11.840 training data and you need to do the work and 0:18:11.920 --> 0:18:14.560 spend the time, and that costs money. So you will 0:18:14.600 --> 0:18:17.280 have a mix of open and closed models. And if 0:18:17.320 --> 0:18:19.760 you take a company like Physical Intelligence, their mo their 0:18:19.800 --> 0:18:23.200 motive is is we want to build the operating system 0:18:23.200 --> 0:18:25.639 that any robot can basically use to navigate the world. 0:18:25.640 --> 0:18:28.800 They want to build the brains for the robots as 0:18:28.800 --> 0:18:32.800 opposed to the robots themselves. And there's an interesting philosophical 0:18:32.840 --> 0:18:36.800 and scientific tangent which Barbara Tavski, who is a friend 0:18:36.840 --> 0:18:39.879 and she was the partner of now the late Danny Connoman, 0:18:39.920 --> 0:18:42.439 who was also a friend, her work and she was 0:18:42.600 --> 0:18:44.880 far less famous, but I think actually more important work 0:18:45.240 --> 0:18:49.359 is all about motor function, and her hypothesis is that 0:18:49.440 --> 0:18:52.520 the entire reason for the existence of the brain, like 0:18:52.560 --> 0:18:56.640 the sole purpose of it is to actually produce movement, 0:18:56.880 --> 0:19:00.239 movement towards food or a mate, or to run from 0:19:00.280 --> 0:19:02.119 a prey or predator, which in turn is doing the 0:19:02.160 --> 0:19:04.800 same kind of thing, and I think that some of 0:19:04.840 --> 0:19:11.680 the most interesting philosophical questions about consciousness and memory, spatial perception, 0:19:12.160 --> 0:19:16.600 embodied cognition, gesturing. I mean, I'm wildly gesticulating while I'm 0:19:16.600 --> 0:19:18.600 talking now with my hands. It's just an innate thing 0:19:18.760 --> 0:19:21.040 being able to do mental simulations, like when we think 0:19:21.080 --> 0:19:24.640 about human brains and machine brains merging. I actually think 0:19:24.680 --> 0:19:28.800 it's really going to be very revelatory as these robotic 0:19:28.800 --> 0:19:31.320 systems advance and us understanding that's a lot of the 0:19:31.400 --> 0:19:34.879 purpose of thinking and intelligence is actually just about moving. 0:19:35.359 --> 0:19:37.720 And so that's a pretty cool side effect that I 0:19:37.720 --> 0:19:39.400 think is going to come from all the commercial and 0:19:39.640 --> 0:19:42.400 speculative ventures stuff that we do and funding these things. 0:19:42.720 --> 0:19:46.080 You keep saying things that like jog me onto something 0:19:46.080 --> 0:19:47.560 else that I was meaning to ask you, but you 0:19:47.680 --> 0:19:49.639 mentioned this impulse. 0:19:49.240 --> 0:19:52.360 That academics like to share their work, and. 0:19:52.320 --> 0:19:54.479 That reminded me of something that someone was telling me. 0:19:54.640 --> 0:19:57.280 So you know, maybe I'm sure you are on this 0:19:57.320 --> 0:19:59.679 site a lot. But for listeners, there's this site archived 0:19:59.720 --> 0:20:03.320 dot org where people like publish research and sort of 0:20:03.359 --> 0:20:06.560 like an open source, sort of ungated manner about all 0:20:06.640 --> 0:20:09.760 kinds of scientific and computer things and just today on 0:20:09.840 --> 0:20:12.879 the artificial intelligence page, there's like fifteen new papers and 0:20:12.880 --> 0:20:16.520 they have headlines like autonomous evaluation and Refinement of digital 0:20:16.560 --> 0:20:20.520 agents or a modular benchmark framework to measure progress and 0:20:20.640 --> 0:20:23.760 improve LLM agents. And this guy was who I was 0:20:23.760 --> 0:20:25.920