WEBVTT - Grace Shao on What the World Should Know About Chinese AI 0:00:02.720 --> 0:00:19.639 Bloomberg Audio Studios, Podcasts, radio News. Hello and welcome to 0:00:19.680 --> 0:00:21.960 another episode of The Odd Lots Podcast. 0:00:22.040 --> 0:00:24.480 I'm Joe Wasental and I'm Tracy Alloway. 0:00:24.720 --> 0:00:26.880 Tracy, I love being in Hong Kong. I love I 0:00:27.000 --> 0:00:29.040 hear so much. I love it here so much. I 0:00:29.080 --> 0:00:30.800 would like come here a few times a year if 0:00:30.840 --> 0:00:31.160 we could. 0:00:31.440 --> 0:00:32.400 I'm sure you would. 0:00:32.560 --> 0:00:36.320 I've lived here for like, I guess, almost four years, 0:00:36.400 --> 0:00:40.640 so it's kind of weird coming back. But Hong Kong 0:00:40.680 --> 0:00:43.880 has a lot of pluses, like great food, great weather 0:00:44.080 --> 0:00:47.200 for most of the year, beaches. I once heard someone 0:00:47.240 --> 0:00:50.159 describe it as Manhattan meets Maui, which I think is 0:00:50.240 --> 0:00:51.519 like pretty accurate. 0:00:51.760 --> 0:00:53.120 Oh it's so nice. 0:00:53.680 --> 0:00:56.200 The weather's shade, nugget, it's not great right now. 0:00:56.320 --> 0:00:58.520 Like this week, we've come during a I guess it's 0:00:58.520 --> 0:00:59.960 a monsoon season, right. 0:01:00.080 --> 0:01:03.000 Yeah, it's the rainy season. But oh well, I like thunderstorms, 0:01:03.000 --> 0:01:03.920 so I'm enjoying it. 0:01:04.040 --> 0:01:07.039 Yeah, I'm enjoying it too. Anyway. One thing that has 0:01:07.160 --> 0:01:09.680 changed since the last time you were in Hong Kong 0:01:10.400 --> 0:01:13.080 you loved in twenty twenty two, Yeah, we weren't doing 0:01:13.080 --> 0:01:14.240 as many AI episodes. 0:01:14.480 --> 0:01:18.000 No, we definitely weren't. In fact, So I remember one 0:01:18.080 --> 0:01:20.280 of the big stories when I was here in Hong 0:01:20.400 --> 0:01:24.399 Kong in twenty twenty was China's tech crackdown, right. 0:01:24.319 --> 0:01:26.360 That's right, that's right, right, And like there. 0:01:26.280 --> 0:01:28.960 Was all this concern about whether or not the crackdown 0:01:29.080 --> 0:01:32.720 was going to destroy China's entrepreneurial spirit. I'm doing air 0:01:32.840 --> 0:01:35.800 quotes on a podcast. I don't know why, but fast 0:01:35.840 --> 0:01:42.360 forward six years and there's entrepreneurism basically everywhere, and we 0:01:42.440 --> 0:01:45.800 talk a lot about how China is producing all these 0:01:45.880 --> 0:01:47.440 new AI models. 0:01:47.640 --> 0:01:51.320 Okay, can I like say, like, I know very little. 0:01:51.680 --> 0:01:53.720 I mean, I know very little about AI, but I 0:01:53.760 --> 0:01:56.600 know even less about Chinese AI. But here are some 0:01:56.720 --> 0:02:01.120 of my general impressions, which is, it seems like there's 0:02:01.200 --> 0:02:04.320 so many open source Okay, so I know they're largely 0:02:04.360 --> 0:02:07.720 open source. It seems like every random company you see, 0:02:07.760 --> 0:02:11.160 like some toothpaste company, and they'll have produced an LLM. 0:02:11.240 --> 0:02:14.560 So I'm very curious, like how they're making money on it. 0:02:15.200 --> 0:02:16.960 I also get the impression and like, do you know 0:02:17.000 --> 0:02:19.720 the heads of American AI labs speak in these like 0:02:19.760 --> 0:02:23.560 sort of quasi mystical terms, et cetera. It doesn't feel 0:02:23.840 --> 0:02:26.080 quite the same here where it feels like a bit 0:02:26.120 --> 0:02:28.920 more of like yet another technology. But I'm glad you 0:02:28.960 --> 0:02:31.360 brought up the point about the tech crackdown because at 0:02:31.400 --> 0:02:34.000 the time, the whole story was like, oh, there needs 0:02:34.040 --> 0:02:36.160 to be less focus on sort of digital tech and 0:02:36.200 --> 0:02:39.280 more focused on card tech, which has been done extremely 0:02:39.680 --> 0:02:43.440 that's been an extraordinarily successful endeavor. And then my last 0:02:43.440 --> 0:02:46.320 impression though, is that since the release of JADGPT in 0:02:46.400 --> 0:02:49.440 late twenty twenty two, that was the moment it's like, no, 0:02:49.520 --> 0:02:53.760 we really have to also compete on sort of this 0:02:53.919 --> 0:02:57.720 next era of software and sort of consumer facing tech 0:02:57.880 --> 0:02:59.280 tech breakthroughs. Yeah. 0:02:59.320 --> 0:03:02.360 Oh, for all that AI scene in China feels much 0:03:02.360 --> 0:03:06.639 more utilitarian to me. It's more about like the big companies, 0:03:06.720 --> 0:03:09.520 the ten cents, the Ali Baba is sort of using 0:03:09.680 --> 0:03:15.160 AI for their existing business models rather than this existential thing, 0:03:15.200 --> 0:03:17.960 which it is in the US, where like AI is 0:03:18.160 --> 0:03:19.720 the business that's just. 0:03:19.680 --> 0:03:23.040 It, right, Yeah, that's exactly. AI is sort of weird, 0:03:23.120 --> 0:03:24.880 like it sort of sits in the middle of what 0:03:24.919 --> 0:03:27.560 you would call like software and hard tech because we 0:03:27.600 --> 0:03:30.919 consume it through the browser, right, sort of the same way, 0:03:31.080 --> 0:03:33.600 or in many cases through the browser the same way 0:03:33.639 --> 0:03:35.840 that we would go to an Amazon or online gaming 0:03:35.960 --> 0:03:38.360 or something like that. But it's clearly, you know, it's 0:03:38.400 --> 0:03:41.840 a scientific endeavor, and so it's sort of is this 0:03:41.960 --> 0:03:44.280 blend And then you have to figure China's so far 0:03:44.320 --> 0:03:46.000 ahead of the US when it comes to things like 0:03:46.120 --> 0:03:49.320 robotics and evs and batteries, and one thing I don't 0:03:49.360 --> 0:03:52.520 know anything about is the degree to which that melding 0:03:52.680 --> 0:03:57.280 of hardware capabilities with AI capabilities, how that influences the 0:03:57.360 --> 0:04:00.000 direction of the developments of the AI. 0:04:00.440 --> 0:04:03.680 Yeah, I'm also very interested in, like the capital stack 0:04:03.800 --> 0:04:06.280 for Chinese companies, because over in the US, we all 0:04:06.320 --> 0:04:09.160 know that people are flinging money at anything with the 0:04:09.200 --> 0:04:13.240 word AI in it, but in China it's very different. 0:04:13.520 --> 0:04:15.920 I get the impression that it's like much harder to 0:04:16.040 --> 0:04:19.760 raise enormous sums of capital, and so I'm very curious 0:04:19.800 --> 0:04:24.360 how that limited capital actually influences the development of these 0:04:24.400 --> 0:04:25.280 models and the tech. 0:04:25.600 --> 0:04:27.440 I think it's safe to say that both of us 0:04:27.720 --> 0:04:31.960 have a lot of impressions. Yes, right, impression Manytimes that 0:04:32.000 --> 0:04:33.760 its intro is like, I get the impression, but I 0:04:33.760 --> 0:04:36.360 actually have no idea. So that is a good reason 0:04:36.640 --> 0:04:39.120 to actually bring in our guest, someone who is more 0:04:39.160 --> 0:04:42.600 than quote impressions unquote about the AI tech scene. We're 0:04:42.640 --> 0:04:44.960 going to be speaking to someone whose newsletter I'm a 0:04:44.960 --> 0:04:47.280 big fan of and everyone should read. Are going to 0:04:47.320 --> 0:04:51.039 be speaking to the perfect guest, Grace Shao. She's an independent, 0:04:51.080 --> 0:04:55.120 a researcher, and she has a great substack called AI Prome, 0:04:55.200 --> 0:04:57.760 and she joined us here in our Hong Kong office. 0:04:57.760 --> 0:05:00.120 So Grace, thank you so much for coming on od Lave. 0:05:00.360 --> 0:05:02.040 Thank you so much for having me, Joe and Tracy. 0:05:02.200 --> 0:05:05.159 How did we do on our impression? Quote? 0:05:05.360 --> 0:05:07.320 Those are pretty accurate impressions? I think? 0:05:07.320 --> 0:05:10.560 Okay, good, that was the episode. Like, let's start at 0:05:10.600 --> 0:05:13.600 a basic level. So the big impression, the one that 0:05:13.640 --> 0:05:17.919 everyone knows is Chinese models are open source versus the 0:05:18.000 --> 0:05:23.360 closed frontier models of the US. Why did it develop 0:05:23.440 --> 0:05:23.840 that way? 0:05:24.720 --> 0:05:28.640 Yeah, I think people like to think of these myscal reasons, 0:05:29.080 --> 0:05:31.480 but really it was a very pragmatic business reason to 0:05:31.520 --> 0:05:34.160 start with. To start with, a lot of the labs 0:05:34.160 --> 0:05:38.520 have cited that, you know, for Western companies or Western 0:05:38.560 --> 0:05:41.680 developers to trust them, they needed to open source their 0:05:41.720 --> 0:05:44.600 models to build that trust and credibility. So in many ways, 0:05:44.640 --> 0:05:48.200 it's a branding decision. Then on top of that, I 0:05:48.240 --> 0:05:50.640 think you can see it as a philosophical drive. You know, 0:05:50.720 --> 0:05:53.680 the founder of deep Seagalmen Fung, has openly said he 0:05:53.720 --> 0:05:56.400 wants open source his most frontier research to really help 0:05:56.400 --> 0:05:59.040 propel the whole industry as a whole, and that kind 0:05:59.040 --> 0:06:02.039 of R and D sharing has now formed a layer 0:06:02.480 --> 0:06:05.000 for the whole ecosystem where each of the labs kind 0:06:05.040 --> 0:06:08.320 of integrate each other's kind of breakthroughs. You know, you 0:06:08.360 --> 0:06:10.840 see them congratulate each other even on X when they 0:06:10.880 --> 0:06:13.600 have new models announced, so you can say it's a 0:06:13.640 --> 0:06:17.000 bit more collegial. I wouldn't say they're not competing though, however, 0:06:17.400 --> 0:06:20.239 because of the compute constraint, they're faced with, talent constraint 0:06:20.279 --> 0:06:23.040 and the capital constraint even mentioned, they are a lot 0:06:23.080 --> 0:06:25.440 more conscious with where they want to put their money, 0:06:25.600 --> 0:06:27.279 where they want to put their time in R and 0:06:27.320 --> 0:06:30.200 D and all of that forms the basis of a 0:06:30.240 --> 0:06:32.160 strong open source ecosystem. 0:06:32.760 --> 0:06:36.920 Is the culture as pro sharing and pro open source 0:06:37.279 --> 0:06:41.600 as it was even two years ago now. The Deep 0:06:41.640 --> 0:06:45.080 Seek moment was right around Trump's inauguration in early twenty 0:06:45.120 --> 0:06:47.680 twenty five, so about a year and a half ago. 0:06:48.480 --> 0:06:52.760 Since then, has the culture stayed the same or has 0:06:52.880 --> 0:06:56.640 that sort of competition bug, that intense competition bug that 0:06:56.680 --> 0:07:01.240 we know among American AI labs, has it spread to 0:07:01.320 --> 0:07:02.640 the Chinese labs at all. 0:07:03.400 --> 0:07:07.520 I think the sharing is an unintentional result rather than 0:07:07.720 --> 0:07:10.440 you know, an intentional effort in the beginning to even 0:07:10.480 --> 0:07:13.680 start with. They are for sure extremely competitive, and you know, 0:07:13.720 --> 0:07:17.200 we all know the word involution, so like China Ai 0:07:17.520 --> 0:07:20.280 is do adding as well, that means like there's evolution 0:07:20.360 --> 0:07:23.640 in this ecosystem as well. However, I think bringing up 0:07:23.640 --> 0:07:26.360 deep sea deep sea plays a very interesting role in 0:07:26.400 --> 0:07:30.000 the whole ecosystem. Like you mentioned, V three, propel the 0:07:30.040 --> 0:07:32.840 whole industry forward. Everyone kind of start taking China Ai 0:07:32.920 --> 0:07:36.160 more seriously. You know, it brought a lot of interests 0:07:36.200 --> 0:07:39.520 from investors globally back into the internet companies that Tracy mentioned, 0:07:39.560 --> 0:07:41.280 you know, project that there was a bit of a 0:07:41.320 --> 0:07:45.560 slump for three to five years. However, you know, Jiarpoolzai 0:07:45.640 --> 0:07:47.880 is now public listed in Hong Kong, Mini Max is 0:07:47.880 --> 0:07:50.520 pubably listed in Hong Kong. Moonshot is you know, in 0:07:50.600 --> 0:07:54.840 preparation to go public next year. They are competing with 0:07:54.920 --> 0:07:58.800 each other to capture market share, to capture developer mind share. 0:07:59.200 --> 0:08:01.400 But deep Sea plays an interesting role. I want to 0:08:01.400 --> 0:08:05.200 bring it back to DEEPCV four, so V four, you know, 0:08:05.320 --> 0:08:07.800 on the surface, you know, people said, Okay, it wasn't 0:08:07.840 --> 0:08:11.240 as maybe impressive on evails and performance, They didn't catch 0:08:11.320 --> 0:08:12.400 up with the most front of your labs in the 0:08:12.480 --> 0:08:16.800 US whatnot, Right, But it was a very interesting move 0:08:16.920 --> 0:08:19.600 because what I heard from researchers on the ground in 0:08:19.640 --> 0:08:24.160 Beijing was that the lab actually delayed their release for 0:08:24.200 --> 0:08:26.440 about three to four months because they wanted to re 0:08:26.520 --> 0:08:30.120 engineer a lot of the inference onto Huawei. So I'm 0:08:30.120 --> 0:08:34.600 not saying this completely replaces in video or Kuda, not 0:08:34.760 --> 0:08:37.120 at all, because if you ask any developers, they still 0:08:37.160 --> 0:08:40.400 want to use Kuda if they can. However, it was 0:08:40.440 --> 0:08:43.800 the first effort to really I kind of like did 0:08:43.840 --> 0:08:46.200 one for the team, like they kind of like hunt them. 0:08:46.480 --> 0:08:49.800 It's supposed to be like a signal basically, yeah, we're 0:08:49.840 --> 0:08:52.480 doing this all on like a Chinese stack. Yeah. 0:08:52.520 --> 0:08:54.520 They were like, look, guys, like you can actually do this, 0:08:54.679 --> 0:08:58.320 and they became a shared foundation layer for China's model ecosystem. So, 0:08:58.640 --> 0:09:02.160 because again everything is open source and open weight, other 0:09:02.320 --> 0:09:05.120 labs were able to study what they did to actually 0:09:05.280 --> 0:09:09.120 start inferencing on Huawei stack, and I think that was 0:09:09.160 --> 0:09:12.200 the first step, whether it's signaling or actually you know, 0:09:12.440 --> 0:09:17.440 very pragmatic reason to start shifting some reliance on you know, 0:09:17.520 --> 0:09:18.880 the China AI. 0:09:18.679 --> 0:09:20.600 Stack aside from deep Sea. 0:09:20.679 --> 0:09:24.520 Can you kind of describe the differences or what China 0:09:24.640 --> 0:09:27.679 is trying to do on the actual frontier side, because 0:09:27.679 --> 0:09:29.480 there are there are some I. 0:09:29.720 --> 0:09:31.520 Think if you really have to have to look at 0:09:31.559 --> 0:09:34.840 the ecosystem, we can kind of put side the big 0:09:34.880 --> 0:09:37.600 tech for now, but looking at the maybe the foremost 0:09:37.640 --> 0:09:42.200 relevant startup labs Deep Seak, Moonshot, who has Kimmi, Zia 0:09:42.280 --> 0:09:46.319 who has GLM, and then Mini Max. They are still 0:09:46.360 --> 0:09:50.720 probably the most committed to frontier research. However, because the 0:09:51.040 --> 0:09:53.760 constraint we mentioned that they face, whether it's compete when 0:09:53.800 --> 0:09:57.160 it is capital or even friendly talent, they have decided 0:09:57.280 --> 0:09:59.959 out of necessity to basically each focus on a different 0:10:00.240 --> 0:10:03.640 vertical in capturing a different kind of business share. So 0:10:04.160 --> 0:10:08.080 Zi is very focused on coding capabilities. So if anything, 0:10:08.120 --> 0:10:10.440 their gl AND plan is much more similar to maybe 0:10:10.440 --> 0:10:12.840 what you think of Claude cloud co wor cloud code 0:10:12.840 --> 0:10:15.760 et cetera, a codex that kind of product. And then 0:10:15.760 --> 0:10:20.679 you look at Minimax, they're really focused on the multimodality capabilities, Moonshot, 0:10:20.720 --> 0:10:23.560 they're really focused on agents and deep sea. 0:10:23.679 --> 0:10:24.000 Again. 0:10:24.200 --> 0:10:27.200 Really is just focused on pushing the frontier and trying 0:10:27.240 --> 0:10:30.520 to play catch up and push the Chinese ecosystem as 0:10:30.559 --> 0:10:31.440 fast as possible. 0:10:32.160 --> 0:10:34.960 It's really crazy to look at some of the ones 0:10:35.080 --> 0:10:38.040 that they have already gone public here and just put 0:10:38.080 --> 0:10:42.240 in So Minimax is public and in US dollar terms, 0:10:42.800 --> 0:10:46.040 it's a twenty billion dollar company. I mean there are 0:10:46.160 --> 0:10:49.319 people in the US who have done nothing but publish 0:10:49.360 --> 0:10:52.520 a paper on archive dot org, who do not even 0:10:52.559 --> 0:10:56.160 have a product yet, who have probably VC backed applied 0:10:56.280 --> 0:10:59.400 valuations of our twenty billion dollars. How do they make money? 0:10:59.480 --> 0:11:02.480 You know again and open source? Okay? Like in these 0:11:02.520 --> 0:11:05.600 four models that you name, do they have different thoughts 0:11:05.720 --> 0:11:08.920 on how they plan to make money or different business models? 0:11:09.480 --> 0:11:09.800 Yeah? 0:11:09.880 --> 0:11:14.160 So China's VC space in general has not been that vibrant, 0:11:14.200 --> 0:11:17.080 frankly since Internet crackdown and a lot of USD funds 0:11:17.080 --> 0:11:19.840 did exit you know, three to five years ago, with 0:11:19.920 --> 0:11:22.439 Saquoia being maybe the most like high profile, right, like 0:11:22.480 --> 0:11:25.800 we all remember that. Now, people forget even in twenty 0:11:25.800 --> 0:11:27.640 twenty two, a lot of these labs that we just 0:11:27.679 --> 0:11:30.560 talked about, they were struggling to even raise money, you know, 0:11:30.640 --> 0:11:33.240 raise capital, and a lot of them spun out of 0:11:33.320 --> 0:11:36.920 academic institutions. You know, you mentioned their value anywhere roughly 0:11:36.920 --> 0:11:39.120 between twenty to thirty billion right now, but they went 0:11:39.160 --> 0:11:41.880 public between like six to eight billion. That's like kind 0:11:41.920 --> 0:11:46.280 of tiny compared to American valuations. Right now, however, they 0:11:46.320 --> 0:11:49.800 are actually making money, you know, the publicly disclosed information 0:11:49.960 --> 0:11:52.320 I think from mini max and DRUP who indicates that 0:11:52.440 --> 0:11:56.200 they were making just as much like in their last month. 0:11:56.280 --> 0:11:58.559 They've made the same amount of money last month as 0:11:58.559 --> 0:12:01.679 they did last year essentially, and their end of your 0:12:02.000 --> 0:12:04.719 AR projection is anywhere between one to one point two 0:12:04.760 --> 0:12:05.560 billion right now. 0:12:05.960 --> 0:12:07.240 So they are making money. 0:12:07.240 --> 0:12:10.600 And how well, just because their open source doesn't mean 0:12:10.600 --> 0:12:12.760 they don't make money. I think people forget. You know, 0:12:12.920 --> 0:12:15.880 we had open source softwares before as well. People are 0:12:15.880 --> 0:12:19.880 paying for managed services, and when you're paying for an 0:12:19.880 --> 0:12:24.240 API through that AI or minimax whatnot, you basically don't 0:12:24.240 --> 0:12:25.600 have to self host, you don't have to get your 0:12:25.600 --> 0:12:27.280 own GPU. You don't have to get you figure out 0:12:27.280 --> 0:12:29.160 your gown compute. You don't have to figure out your 0:12:29.200 --> 0:12:31.960 own guardrails or deployment, your security or monitoring whatnot. 0:12:32.040 --> 0:12:36.679 Right, So, just to be clear, you can self host 0:12:36.760 --> 0:12:40.360 all of these models, but for the most part, they 0:12:40.400 --> 0:12:43.200 do offer the inference part of the stack, and that 0:12:43.440 --> 0:12:44.960 is a profit center for them. 0:12:45.280 --> 0:13:02.679 Yes, exactly, got it all right. 0:13:02.720 --> 0:13:07.360 So there seem to be two major constraints on Chinese AI. 0:13:07.520 --> 0:13:09.720 Maybe energy is a constraint as well, we should talk 0:13:09.760 --> 0:13:14.679 about that. But there's the capital issue, so not as 0:13:14.760 --> 0:13:18.480 much capital available or people aren't flinging it at AI 0:13:18.600 --> 0:13:20.840 companies the way they are in the US. And then secondly, 0:13:20.880 --> 0:13:23.760 there are the export controls on chips and we talked 0:13:23.760 --> 0:13:26.280 about that a little bit, but can you describe how 0:13:26.360 --> 0:13:31.120 those controls are actually I guess influencing the development of 0:13:31.200 --> 0:13:34.439 the models themselves and like I guess optimization. 0:13:35.120 --> 0:13:37.480 So some of these big tech are actually even buying 0:13:37.480 --> 0:13:40.160 out the contracts that data centers have with some of 0:13:40.160 --> 0:13:42.600 these labs, and you know they are taking over compute. 0:13:42.600 --> 0:13:47.720 So these labs essentially are now optimizing for the highest 0:13:47.800 --> 0:13:50.319 quality inference amount, if that makes sense. They don't actually 0:13:50.400 --> 0:13:53.400 have enough even supply, they don't have enough compute power 0:13:53.400 --> 0:13:56.600 to even like meet the demand that's coming through. So 0:13:56.720 --> 0:13:59.720 that's on like how they're servicing clients, how they're changing 0:13:59.760 --> 0:14:02.120 I guess, or how they're optimizing O their training is 0:14:02.120 --> 0:14:04.680 that a lot of them are really focused on the 0:14:04.720 --> 0:14:08.400 post training. And this goes back to so you know 0:14:08.440 --> 0:14:10.480 how open ai has like three buckets where they check 0:14:10.520 --> 0:14:12.800 money at there's like the R and D, there's pre training, 0:14:12.800 --> 0:14:15.240 this post training R and D. A lot of times 0:14:15.280 --> 0:14:17.480 a lot of money is spent, but say like one 0:14:17.520 --> 0:14:19.680 out of ten things stick, but you need a lot 0:14:19.680 --> 0:14:22.560 of compute and resource some people to be figuring out 0:14:22.600 --> 0:14:25.520 where to go. For a lot of these labs in China, 0:14:25.720 --> 0:14:28.840 they frankly don't have that luxury. So they've even given 0:14:28.840 --> 0:14:31.400 me a metaphor and said it's kind of like knowing 0:14:31.440 --> 0:14:34.840 what the answer to the homeworks and working backwards. So 0:14:34.880 --> 0:14:37.320 they will wait till the frontier labs to come out 0:14:37.360 --> 0:14:40.600 with where the right direction is for the next frontier model, 0:14:40.800 --> 0:14:43.240 and they will work backwards and actually focus all their 0:14:43.280 --> 0:14:46.680 resources on post training. So with post training, they will 0:14:46.680 --> 0:14:49.880 optimize a lot of times the data they collect. For example, 0:14:50.000 --> 0:14:53.720 if a data provider like Mercore provides a very very 0:14:53.800 --> 0:14:57.560 niche set of data set for like an open AI whatnot, 0:14:58.240 --> 0:15:01.040 Maybe they will charge them ten twenty mille dollars. The 0:15:01.160 --> 0:15:05.200 Chinese lab will weigh out that exclusivity contract three to 0:15:05.240 --> 0:15:07.600 six months time, let's say, and then pay a fraction 0:15:07.720 --> 0:15:09.520 if not like a tenth of that price the same 0:15:09.600 --> 0:15:11.880 data set, and that kind of plays into that like 0:15:11.960 --> 0:15:14.480 six to nine month leg that we hear about as well. 0:15:14.680 --> 0:15:18.120 That's really interesting. Let's talk about energy then, because the 0:15:18.200 --> 0:15:21.240 story in the US is that electricity is really the 0:15:21.240 --> 0:15:24.280 big constraint on AI use, and you know, you got 0:15:24.320 --> 0:15:27.400 to find a data center that has an electricity hookup, 0:15:27.440 --> 0:15:29.480 and it has to be reliant and all of that. 0:15:30.000 --> 0:15:32.560 It seems to be in short supply. Is it a 0:15:32.640 --> 0:15:34.640 similar story in China? 0:15:34.720 --> 0:15:38.240 Honestly, energy is probably not the biggest bottleneck right now 0:15:38.280 --> 0:15:40.880 in China. And I think people like to say, well, 0:15:40.920 --> 0:15:43.320 some people like to say, oh, somehow the Chinese government 0:15:43.360 --> 0:15:46.880 had foresight on the AI boom, driving like the energy consumption, 0:15:46.960 --> 0:15:50.680 but definitely not. I think people forget that China's economic 0:15:50.760 --> 0:15:52.920 growth over the last three to four decades also meant 0:15:53.040 --> 0:15:56.200 a rise of urbanization and a lot of the cities 0:15:56.240 --> 0:15:58.080 that you know, we are visiting these days, like at 0:15:58.120 --> 0:16:00.720 least westerns are visiting, like Beijing, China High Engine with 0:16:00.760 --> 0:16:03.800 all these robots and evs or whatnot. These were all 0:16:03.920 --> 0:16:08.240 really urbanized within the last two three decades. And because 0:16:08.240 --> 0:16:10.720 of that, the grid is very new, and because of that, 0:16:10.920 --> 0:16:13.720 the government already foresaw that there was going to be 0:16:13.880 --> 0:16:17.000 a increase in energy demand, and you know, so a 0:16:17.040 --> 0:16:20.880 lot of the energy plants, you know, the solar plants, 0:16:20.960 --> 0:16:23.760 hydro plants, whatnot, were actually built out in you know, 0:16:23.800 --> 0:16:28.000 anticipation for that. Now obviously this has coincided with now 0:16:28.040 --> 0:16:30.960 the AI boom, and it's really helped out beyond that. 0:16:31.040 --> 0:16:33.920 You know, China has an advantage in the fact that 0:16:33.960 --> 0:16:37.840 they can actually drive top down mandates and provincial governments 0:16:37.880 --> 0:16:40.840 will follow suit. This is something quite unique to China 0:16:40.920 --> 0:16:43.640 because it's not like decided about each state. So when 0:16:43.680 --> 0:16:46.160 they pushed out the East State to West compute where 0:16:46.320 --> 0:16:49.240 it's basically a top down initiative where they built a 0:16:49.360 --> 0:16:54.200 ton of renewable energy for cheap in rural mountainous areas 0:16:54.280 --> 0:16:59.280 in guat Zo Province like even Sinzong, Inner Mongolia Suchwan 0:16:59.360 --> 0:17:03.720 you know, those were like very easily executed, frankly, and 0:17:03.720 --> 0:17:06.840 then ninety percent of the population actually sit on the 0:17:06.880 --> 0:17:10.760 eastern coastal lines, like we think about Beijing, Tianjing, Shanghai, 0:17:10.760 --> 0:17:13.040 sh Engine, that's all on east, so that's where the 0:17:13.160 --> 0:17:17.240 data comes from. So that kind of optimization has also 0:17:17.320 --> 0:17:20.000 really helped them, you know, with the low that is 0:17:20.359 --> 0:17:22.159 the demand right now, I want to. 0:17:22.119 --> 0:17:24.680 Get back to something you said. So first of all, 0:17:24.840 --> 0:17:28.920 just to clarify, you mentioned companies like Mercore that sell 0:17:29.040 --> 0:17:33.360 proprietary data that they are able to collect and manufacture 0:17:33.440 --> 0:17:36.080 in various ways. Then they sell it to an open 0:17:36.119 --> 0:17:38.720 aira So a company like Merkcore will hire a bunch 0:17:38.800 --> 0:17:41.640 of people to say, build power points, and then they'll 0:17:41.640 --> 0:17:43.720 collect the data on how they do that, and that 0:17:43.840 --> 0:17:46.399 is fresh data that they can sell. So those have 0:17:47.160 --> 0:17:50.679 exclusivity windows, after which they can then sell them to anyone. 0:17:50.960 --> 0:17:54.240 I'm not saying Marcos specifically, but supposedly there are these 0:17:54.280 --> 0:17:57.080 data providers that do the cell and they have exclusivity windows, 0:17:57.119 --> 0:17:59.520 and then the Chinese labs kind of weigh that out 0:17:59.600 --> 0:18:02.679 so they can pay like maybe a million dollars versus 0:18:02.720 --> 0:18:04.399 like time and folk same data side. 0:18:04.480 --> 0:18:07.840 So This gives to something generally speaking, which is that 0:18:08.080 --> 0:18:12.640 people are around the world correctly like quite impressed by 0:18:12.960 --> 0:18:16.200 how high quality the Chinese models are, even if they're behind. 0:18:16.680 --> 0:18:18.359 But then you have things like that, and then you 0:18:18.440 --> 0:18:21.720 also have accusations from the likes of Anthropic that they're 0:18:21.720 --> 0:18:25.640 distilling models and that they're finding ways to collect the 0:18:25.680 --> 0:18:31.560 outputs of American models for training. So then you could say, well, yeah, sure, 0:18:31.760 --> 0:18:34.040 that's like, it's great this open source model and it 0:18:34.040 --> 0:18:36.800 can stay close to the edge. But then the encounter 0:18:36.920 --> 0:18:40.359 is that they can only be so advanced because there's 0:18:40.480 --> 0:18:44.399 this extremely capital intensive closed source model in the US 0:18:44.880 --> 0:18:49.359 that's really establishing the frontier, and that these Chinese companies 0:18:49.359 --> 0:18:52.160 wouldn't be anywhere near where they were if they weren't 0:18:52.200 --> 0:18:54.600 sort of I guess you would say drafting off the 0:18:54.640 --> 0:18:55.479 American labs. 0:18:56.040 --> 0:18:58.600 Yeah, I think the compute constraint on the capital constraint 0:18:58.640 --> 0:19:01.840 Israel and frankly, like no one's hiding that or pretending 0:19:01.880 --> 0:19:04.040 that that's not an issue for them right now, like 0:19:04.119 --> 0:19:07.040 deep Sea has openly said they even were struggling right 0:19:07.080 --> 0:19:09.919 like they needed more compute. I think on the distillation 0:19:10.400 --> 0:19:14.080 allegations or accusation it is quite interesting. Like recently, I've 0:19:14.080 --> 0:19:16.560 been thinking about this a lot and thinking about what 0:19:16.600 --> 0:19:19.000 it means for distillation and what it means for the 0:19:19.160 --> 0:19:22.280 models to catch up. Right. So, there was this one 0:19:22.600 --> 0:19:25.080 quote from y'all showing you who is a Google Deep 0:19:25.119 --> 0:19:29.480 Mind researcher. He said, there're smart distillation and dumb disolation. 0:19:29.880 --> 0:19:32.840 Dumb distillation is something I think most of us were 0:19:32.880 --> 0:19:35.280 frankly non technical think about. It's like, okay, you take 0:19:35.320 --> 0:19:38.119 like a thousand queries, you take the answers of whatever 0:19:38.200 --> 0:19:40.080 claud gives you, right, and then you kind of force 0:19:40.160 --> 0:19:43.520 copy that into your said model, and then you forcefully 0:19:43.720 --> 0:19:46.760 make them basically like get the exact same answer. Smart 0:19:46.760 --> 0:19:49.960 distillation is like you using the frontier model almost as 0:19:50.000 --> 0:19:52.720 a partner to help you with the judgment, for the 0:19:52.800 --> 0:19:56.400 valuation and even the data labeling itself. So you're using 0:19:56.480 --> 0:19:59.359 it as almost a teacher for your own model. It 0:19:59.480 --> 0:20:03.000 guides it little bit versus really copy pasting the answer, 0:20:03.080 --> 0:20:06.240 if that makes sense, And that part of it is 0:20:06.320 --> 0:20:09.960 frankly not that unethical or like you know that frown 0:20:10.080 --> 0:20:12.399 upon right now, because that is what enterprises do when 0:20:12.400 --> 0:20:14.879 they're fine to me, So it's all kind of a 0:20:14.880 --> 0:20:16.200 bit of a mercury area. 0:20:15.960 --> 0:20:16.480 To be honest. 0:20:16.560 --> 0:20:19.119 Okay, so you mentioned data, just then talk to us 0:20:19.119 --> 0:20:22.440 about what the Chinese data set actually looks like, because 0:20:22.480 --> 0:20:25.320 I imagine, if you're a ten cent I mean you've 0:20:25.320 --> 0:20:28.160 got we chat, right, that must be a whole load 0:20:28.200 --> 0:20:31.479 of data on which to actually like build your AI. 0:20:32.119 --> 0:20:34.000 But on the other hand, I imagine, like there are 0:20:34.000 --> 0:20:38.359 some restrictions around the Internet. Obviously, what does it actually 0:20:38.400 --> 0:20:39.000 look like here? 0:20:39.520 --> 0:20:41.280 So I actually split that into two parts. 0:20:41.320 --> 0:20:44.560 On the data itself, people often think China is so 0:20:44.800 --> 0:20:47.840 data intensive and you just have mass amount of data 0:20:47.840 --> 0:20:51.080 to use for AI training. However, actually people forget again 0:20:51.280 --> 0:20:55.640 China's enterprise build out or you know whatever. The knowledge 0:20:55.640 --> 0:20:59.960 work economy is very new and not as sophisticated frankly 0:21:00.080 --> 0:21:02.800 as American ecosystem or the Western ecosystem, if you have 0:21:02.840 --> 0:21:05.560 to put it that way. So data is often unstructured, 0:21:05.840 --> 0:21:08.480 and data thus a lot of the specific needs for 0:21:08.680 --> 0:21:11.040 you know, the kind of training we're seeing today is 0:21:11.080 --> 0:21:14.320 not as vibrant, or the data ecosystem is not as 0:21:14.359 --> 0:21:18.359 sophisticated as what American data providers kind of can provide, 0:21:18.400 --> 0:21:21.320 such as mercre like we just mentioned now on the 0:21:21.359 --> 0:21:24.240 big tech side, it's bit interesting. So I'm glad you 0:21:24.280 --> 0:21:27.280 brought up Tencent, because Tencent actually just announced last week 0:21:27.320 --> 0:21:30.439 that they are working in the works of creating an 0:21:30.480 --> 0:21:32.880 agent that can be plugged into we Chat. This has 0:21:32.920 --> 0:21:35.560 been very controversial and it has actually had a lot 0:21:35.560 --> 0:21:40.760 of pushback even internally, because WeChat's product manager, Alan Jung 0:21:40.880 --> 0:21:44.520 has been famously or notoriously known to be kind of 0:21:44.560 --> 0:21:46.840 hard to work with if you want to push something 0:21:46.840 --> 0:21:49.720 with a WeChat because he's so protective of that user experience. 0:21:50.000 --> 0:21:53.480 It's his baby, right, And Tencent, like you said, has 0:21:53.600 --> 0:21:56.959 we Chat, which is a super app, has more than 0:21:57.000 --> 0:22:00.720 one point four billion MAU globally, like mostly one point 0:22:00.720 --> 0:22:02.840 three million people in China and the Chinese ask for 0:22:03.040 --> 0:22:07.280 globally or people who work with China. That is immense value. 0:22:07.600 --> 0:22:11.919 But the risk and compliance risk of potentially of agent 0:22:12.000 --> 0:22:16.119 going rogue within that chatbot, or that of agent going 0:22:16.200 --> 0:22:19.240 rogue in executing you know, whether it's a purchase whatnot, 0:22:19.320 --> 0:22:22.120 is that risk is very high. So they've been working 0:22:22.200 --> 0:22:25.320 on that. And on top of that, Tencent itself has 0:22:25.400 --> 0:22:28.119 been lagging behind compared to other big tech in their 0:22:28.160 --> 0:22:31.280 own proprietary models, and they've really been trying to really 0:22:31.280 --> 0:22:31.959 play catch up. 0:22:32.240 --> 0:22:33.360 They actually last. 0:22:33.200 --> 0:22:36.480 Year poached someone from open Ai who is a researcher 0:22:36.560 --> 0:22:38.520 called y'all shoam U as well has the same name 0:22:38.560 --> 0:22:40.800 as the other researcher we just mentioned, to lead this 0:22:40.840 --> 0:22:43.800 whole initiative, and their whole goal is to basically build 0:22:43.840 --> 0:22:47.560 a ten cent agent native model, and that is their 0:22:47.600 --> 0:22:49.600 biggest goal because end of the day, like you said 0:22:49.600 --> 0:22:52.719 in the very beginning, their goal is to optimize their 0:22:52.760 --> 0:22:56.600 existing businesses already and bring AI to the mass consumer 0:22:56.720 --> 0:22:57.520 as fast as they can. 0:22:57.880 --> 0:23:01.520 You know, you mentioned poaching a research from open Ai, 0:23:01.800 --> 0:23:04.600 and it's like the way I see it, AR will 0:23:04.600 --> 0:23:07.720 definitely be built by the Chinese. The question is whether 0:23:07.760 --> 0:23:10.080 it be built by the Chinese working the American labs 0:23:10.160 --> 0:23:12.520 or whether it will be built by Chinese working and 0:23:12.640 --> 0:23:15.840 Chinese labs. Has there been a gathering of steam of 0:23:15.880 --> 0:23:20.440 researchers from that had been working at American labs going 0:23:20.920 --> 0:23:23.320 to Chinese labs or are they still sort of one 0:23:23.359 --> 0:23:24.359 off and somewhat rare. 0:23:24.920 --> 0:23:27.000 I think even during the Internet era, we saw a 0:23:27.000 --> 0:23:29.920 lot of Chinese nationals or Chinese I think people returning 0:23:29.920 --> 0:23:32.919 to China right. I think this it's easy to blanket 0:23:32.960 --> 0:23:37.399 statement as geopolitical headwinds. People are scared, but realistically, I 0:23:37.440 --> 0:23:40.280 think most people are just trying to take care of 0:23:40.320 --> 0:23:42.800 their families and live a good life, right, yea. I 0:23:42.840 --> 0:23:44.800 hate to sound so crass about it, but you know, 0:23:44.920 --> 0:23:49.200 sometimes it's what your package look like and over generalize. 0:23:49.240 --> 0:23:52.359 I've heard from many researchers say, look, if my wife 0:23:52.400 --> 0:23:54.320 is a lawyer in China, my wife is a nurse 0:23:54.400 --> 0:23:56.840 in China, my wife is a teacher in China, that 0:23:56.960 --> 0:24:00.000 kind of employment opportunity is very, very hard to actually 0:24:00.119 --> 0:24:01.000 you transfer to. 0:24:00.960 --> 0:24:01.760 A new market. 0:24:01.880 --> 0:24:03.800 They're like, if I can get a similar package and 0:24:03.840 --> 0:24:06.600 a growth opportunity in one of the big labs in China, 0:24:06.920 --> 0:24:09.399 I would pick that over living in the US. And 0:24:09.440 --> 0:24:11.160 on top of that, I think that something is lost 0:24:11.200 --> 0:24:14.359 in the nuance is my parents immigrated to North America 0:24:14.440 --> 0:24:17.160 thirty plus years ago. It was a very clean cut, 0:24:17.359 --> 0:24:21.080 like quality of life. It's just like objectively better in 0:24:21.200 --> 0:24:23.920 any city in North America compared to any city in China. 0:24:24.440 --> 0:24:27.280 Now that's kind of a personal debate, right, because it 0:24:27.320 --> 0:24:29.800 depends on what you really value. If you want to 0:24:29.840 --> 0:24:32.240 be close to the city center. You want that fast paced, 0:24:32.280 --> 0:24:37.520 like techno evy futuristic lifestyle, China actually gives that to you. 0:24:37.560 --> 0:24:38.960 And then on top of that, if you want to 0:24:38.960 --> 0:24:41.399 be close to your family, it's a very personal reason. 0:24:41.520 --> 0:24:44.080 So I've met a lot of research I actually decided 0:24:44.160 --> 0:24:47.040 to come back to China or this part of the 0:24:47.080 --> 0:24:49.320 world simply because they wanted to do it for personal 0:24:49.359 --> 0:24:50.080 family reasons. 0:24:50.920 --> 0:24:53.119 Are they paid as much as they are in the US? 0:24:53.240 --> 0:24:55.479 Because we get headlines all the time about you know, 0:24:55.520 --> 0:24:59.480 so and so is joining whatever company, and people treat 0:24:59.560 --> 0:25:04.320 that new whose like sports stars, right, like teams trading 0:25:04.800 --> 0:25:07.160 their best players. Is it a similar thing here? 0:25:07.680 --> 0:25:09.960 I think you definitely get less of that sports star 0:25:10.400 --> 0:25:14.960 vibe or mentality. Here. They're still getting paid like hafty amounts. 0:25:15.040 --> 0:25:18.240 How much they don't usually disclose, but at least even 0:25:18.240 --> 0:25:21.159 in the Internet era, like a byte dance product manager 0:25:21.280 --> 0:25:24.040 can make just as much as a meta product manager. Similarly, 0:25:24.440 --> 0:25:27.360 if you're like an average AI researcher, you're probably making 0:25:27.440 --> 0:25:31.000 a similar amount. Although the star star players, like the 0:25:31.040 --> 0:25:33.800 ones that are signing one hundred million bonuses, I don't 0:25:33.840 --> 0:25:37.080 know if we had anything like that big like in China, 0:25:37.640 --> 0:25:40.440 but look, they made their money with the IPOs, they 0:25:40.480 --> 0:25:44.280 made the money recently with all this AI boom. It's 0:25:44.320 --> 0:25:46.720 just on a maybe slightly smaller scale. Doesn't mean that 0:25:46.720 --> 0:25:49.720 they're not making much more than the frankly average person. 0:25:50.040 --> 0:25:52.320 Tracy, Can I say something that might be sort of 0:25:52.359 --> 0:25:56.280 a sacrilege for a podcast host to say? Uh, Okay, 0:25:57.520 --> 0:26:01.240 I'm only speaking for myself here, not necessarily speaking for 0:26:01.280 --> 0:26:04.399 the team, But it occurred to me like, I'm not 0:26:04.560 --> 0:26:08.240 really sure if I'd be that interested in, say, getting 0:26:08.280 --> 0:26:11.240 the CEO of like an American AI lab on the 0:26:11.280 --> 0:26:13.160 podcast as a guest. I don't know what I would 0:26:13.160 --> 0:26:15.840 ask them, because like, like do I really want to 0:26:15.880 --> 0:26:20.280 hear like Sam Altman or Dabi Sasabi's or whatever like 0:26:20.480 --> 0:26:23.399 the future of work and all this stuff or like 0:26:23.600 --> 0:26:26.760 these all the big you know, I love doing AI episodes. 0:26:27.160 --> 0:26:30.240 I just feel like at that level I would rather 0:26:30.359 --> 0:26:32.919 talk to us or like someone in actually the engine 0:26:33.000 --> 0:26:35.679 room rather than this sort of big picture a person 0:26:35.720 --> 0:26:39.320 who may have some degree of AI psychosis and just 0:26:39.359 --> 0:26:42.760 like as on a speaks in like the biggest generalities. 0:26:42.960 --> 0:26:46.600 Okay, well, now Sam Altman needs to like invite himself 0:26:46.600 --> 0:26:49.280 on the show just to test your your commitment to 0:26:49.400 --> 0:26:51.080 not having AI CEOs. 0:26:51.359 --> 0:26:53.199 No, I would, I would do it, But I like, 0:26:53.480 --> 0:26:56.080 let's just agree here that if we ever get one 0:26:56.080 --> 0:26:59.200 of the really big like lab CEOs, let's just ask 0:26:59.240 --> 0:27:03.280 the very like sort of mundane questions about operations and now, 0:27:03.400 --> 0:27:05.800 like what are we all going to do? And you know, 0:27:05.960 --> 0:27:07.959 is what is the meaning of life going to be 0:27:08.040 --> 0:27:09.800 when we don't have jobs? Because I'm so sick of 0:27:09.840 --> 0:27:13.439 those conversations. They may be important at some point, but Grace, 0:27:13.480 --> 0:27:17.960 I'm sort of curious from your perception, like it does 0:27:18.040 --> 0:27:20.639 feel like the heads of the American AI labs have 0:27:20.840 --> 0:27:24.520 some degree of AI psychosis themselves. Either they talk about 0:27:24.920 --> 0:27:27.800 all white color employment is going to disappear, or that 0:27:27.800 --> 0:27:30.720 they're going to build a monster that if done wrong, 0:27:31.000 --> 0:27:33.800 is going to be out of control and that they're 0:27:33.840 --> 0:27:36.359 not you know, it'll escape, that it'll escape the sandbox. 0:27:36.720 --> 0:27:41.679 Is there the same sort of existential discourse in the 0:27:41.760 --> 0:27:43.040 Chinese AI community? 0:27:43.600 --> 0:27:48.040 Yeah? I think to start with in the AI community themselves, 0:27:48.200 --> 0:27:50.600 I would say people are a lot more pragmatic and 0:27:50.640 --> 0:27:52.760 I think recently I was talking to Nathan Lambert, who 0:27:52.760 --> 0:27:55.160 is open source researcher who just came to China visited 0:27:55.160 --> 0:27:57.520 all the labs. He said, Look, I was shocked to 0:27:57.560 --> 0:28:00.840 see majority of labs are so as in like a 0:28:00.880 --> 0:28:04.399 lot of the researchers are still students, a lot of 0:28:04.400 --> 0:28:07.040 them are interns, and the core research teams are maybe 0:28:07.080 --> 0:28:09.720 led by a handful of people. And then these people 0:28:09.760 --> 0:28:13.360 are academics by training. So maybe they're a bit less commercial. 0:28:13.560 --> 0:28:14.160 Maybe you can. 0:28:14.080 --> 0:28:17.720 Say they're less like sophisticated to make manimally the market, 0:28:17.760 --> 0:28:20.560 whatever you want to call them. So I definitely feel 0:28:20.600 --> 0:28:23.120 like there's less of that kind of psychosis or high 0:28:23.280 --> 0:28:24.679 level narrative going around. 0:28:25.000 --> 0:28:26.480 However, I would. 0:28:26.200 --> 0:28:30.560 Say that there is obviously anxiety from the public in 0:28:30.600 --> 0:28:34.399 some degree, not as much of a back pushback. But 0:28:35.240 --> 0:28:37.880 recently there was a very interesting court case in Hongjo, 0:28:38.080 --> 0:28:40.640 which is you know, home to Ali Baba and a 0:28:40.680 --> 0:28:44.560 lot of these al abs. Basically, a company try to 0:28:45.240 --> 0:28:48.480 like lay off a person saying you are being replaced 0:28:48.480 --> 0:28:51.080 with AI, and the court literally rules say that is 0:28:51.120 --> 0:28:54.040 not allowed and no company can use AI as an 0:28:54.080 --> 0:28:57.000 excuse to lay off or replace or even cut short 0:28:57.280 --> 0:29:00.440 their contract time. So that was a really swift reaction 0:29:00.520 --> 0:29:04.040 from regulators, and I think it really did serve as a. 0:29:04.040 --> 0:29:05.560 Calming factor for the public. 0:29:05.960 --> 0:29:08.280 Obviously, I also think I want to preface the fact 0:29:08.280 --> 0:29:10.680 that the knowledge worker economy makes. 0:29:10.480 --> 0:29:12.840 Up less of the whoverall. 0:29:12.280 --> 0:29:15.400 Economy in China as well, so that kind of fear 0:29:15.560 --> 0:29:19.280 maybe doesn't feel as imminent, but that conversation is being had. 0:29:19.600 --> 0:29:21.800 But I do think in Asia in general, not only 0:29:21.800 --> 0:29:24.840 in China. You look at South Korea, Singapore, all these 0:29:24.880 --> 0:29:28.560 countries are approaching AI in a very pragmatic way, and 0:29:28.680 --> 0:29:30.959 the tire moms are trying to train up the kids 0:29:31.000 --> 0:29:34.000 to be AI native, the students are trying to train 0:29:34.040 --> 0:29:36.480 themselves up to be AA native. People are preparing for 0:29:36.520 --> 0:29:39.480 the future versus pushing back on the future. 0:29:39.920 --> 0:29:43.280 That's interesting. You know, in the US, obviously, companies announced 0:29:43.320 --> 0:29:46.640 that they're laying people off, and they cite AI even 0:29:46.640 --> 0:29:49.360 frequently when there's no evidence that AI had anything to 0:29:49.400 --> 0:29:51.400 do with it. So it's interesting they're not allowed to 0:29:51.440 --> 0:29:55.239 do it. In defense of American AI labs, most of 0:29:55.280 --> 0:29:58.040 them lose a lot of money, and yet they actually 0:29:58.120 --> 0:30:01.200 do spend, at least the one spend there quite a 0:30:01.240 --> 0:30:06.440 decent amount on so called like alignment research, safety research, 0:30:06.520 --> 0:30:10.080 making sure the AIS don't go rogue et cetera. How 0:30:10.080 --> 0:30:13.160 big of a part of the Chinese labs, how much 0:30:13.200 --> 0:30:15.080 did they spend on I guess what they would you 0:30:15.120 --> 0:30:17.200 would put but the American labors would put into the 0:30:17.200 --> 0:30:18.160 safety bucket. 0:30:18.360 --> 0:30:20.360 I do have to say, I'm not a policy expert, 0:30:20.400 --> 0:30:22.080 so I don't work with a lot of the safety 0:30:22.120 --> 0:30:25.880 people as much. But there are organizations in China that 0:30:26.000 --> 0:30:29.360 are definitely working, like the regulators as well as the 0:30:29.680 --> 0:30:33.959 private sector working together. And for some context, in China, 0:30:34.080 --> 0:30:37.040 there are various moving parts in the government. There's MIT, 0:30:37.240 --> 0:30:41.640 the CAC, et cetera. These agencies basically some are to 0:30:41.680 --> 0:30:45.840 propel economic development. So in this case AI diffusion, the 0:30:45.880 --> 0:30:49.280 whole idea of AI plus AI plus every single sector 0:30:49.320 --> 0:30:52.640 you can think of, some act more like a guardrail 0:30:53.000 --> 0:30:55.480 as a protector, so they are working. 0:30:55.200 --> 0:30:55.920 Hand in hand. 0:30:56.000 --> 0:31:00.000 And on top of that, every single AI jen AI 0:31:00.040 --> 0:31:03.080 application as well as LM company have to go through 0:31:03.120 --> 0:31:07.320 the National Registry in China, so they actually disclose what 0:31:07.520 --> 0:31:10.320 is being trained, what is you know, the potential risk 0:31:10.680 --> 0:31:13.560 that said, I think right now, you know, no one 0:31:13.680 --> 0:31:17.120 really knows what the real impact of AI will be 0:31:17.160 --> 0:31:20.560 on economy, but you know definitely that fear mongering narrative 0:31:20.720 --> 0:31:36.040 is not mainstream in China. 0:31:37.960 --> 0:31:41.000 How would you describe I guess the approach of the 0:31:41.080 --> 0:31:44.720 Chinese government to AI in general, because it feels like 0:31:44.800 --> 0:31:47.640 the trade off for maybe not being on the cutting 0:31:47.760 --> 0:31:51.200 edge with frontier models is well, you're further along with 0:31:51.640 --> 0:31:54.560 sovereign AI and the government maybe has like a better 0:31:54.640 --> 0:31:57.360 handle on what the labs are actually doing. 0:31:58.200 --> 0:32:01.520 The Chinese government probably sees this in a much more 0:32:01.600 --> 0:32:05.959 pragmatic way. You know, just like how there was an 0:32:05.960 --> 0:32:10.040 Internet plus policy fifteen years ago, there's now an AI 0:32:10.080 --> 0:32:13.200 plus policy. When deep seak moment took off, you know, 0:32:13.240 --> 0:32:16.760 there was a frenzy of private companies even in home appliances, 0:32:16.800 --> 0:32:19.760 trying to embed deep seat, integrate deep sea. I'm just like, 0:32:20.200 --> 0:32:22.520 what is a AI vacuum going to do for you? 0:32:22.640 --> 0:32:25.720 Or AI like electric truth rush for you? It was wild, right, 0:32:26.120 --> 0:32:28.520 but the government picked that up. And I think what 0:32:28.640 --> 0:32:30.920 the Chinese government, going back to what we talked about earlier, 0:32:30.960 --> 0:32:33.920 is that they have the advantage of having the ability 0:32:33.920 --> 0:32:36.640 to push things down from top down and at the 0:32:36.800 --> 0:32:40.200 very high level, they're seeing AI as an economic driver 0:32:40.760 --> 0:32:45.600 to propel maybe efficiency, to address some of the labor 0:32:45.600 --> 0:32:48.480 shortage that is coming as the population could continues to 0:32:48.520 --> 0:32:51.520 age and decline. It also addresses a lot of issues 0:32:51.560 --> 0:32:53.520 where a lot of the young people don't want to 0:32:53.560 --> 0:32:56.760 work in manufacturing roles. They want to be automated actually, 0:32:56.920 --> 0:32:59.000 and they want to work in urban areas. So they 0:32:59.040 --> 0:33:02.800 have that now, and then each of the provincial governments 0:33:02.880 --> 0:33:05.760 will take that as kind of like a KPI. They're like, 0:33:05.760 --> 0:33:08.880 all right, let's go ask like vcs essentially and go 0:33:09.560 --> 0:33:13.240 find you know, future deep seeks and fund them. However, 0:33:13.320 --> 0:33:15.720 how much these companies want to take government money is 0:33:15.760 --> 0:33:18.560 a different discussion. A lot of them will then even 0:33:18.720 --> 0:33:22.840 support them by providing you know, infrastructure like buildings, offices, 0:33:23.000 --> 0:33:25.720 even like some of them heard like dormitory for these 0:33:25.760 --> 0:33:28.840 young entrepreneurs, and then give them money and capital try 0:33:28.840 --> 0:33:31.840 things out. There's also these AI pilot zones being rolled 0:33:31.840 --> 0:33:34.200 out across the country. I think now about eleven or 0:33:34.200 --> 0:33:36.640 twelve of them, where you know, people can try out 0:33:36.680 --> 0:33:40.480 new AI products. I met with the largest AI developer 0:33:40.520 --> 0:33:42.960 community founder in China a couple of weeks ago, and 0:33:43.000 --> 0:33:45.760 she was saying, there's more than a couple hundred thousand 0:33:45.840 --> 0:33:48.800 developers in this ecosystem and they are working with. 0:33:48.840 --> 0:33:50.680 Regulators and the private sector. 0:33:50.920 --> 0:33:52.920 So if say bitdowns have a new product, they might 0:33:52.960 --> 0:33:54.680 go to them first and say can you try this out? 0:33:55.040 --> 0:33:58.320 And then they will report and debug and see what's happening, 0:33:58.560 --> 0:34:01.600 and then tell the whoever low government that's funding them 0:34:01.600 --> 0:34:04.560 with providing with the infrastructure, say, hey, this product might 0:34:04.600 --> 0:34:06.640 come out. Do you want to be part of it? 0:34:06.720 --> 0:34:08.160 Do you want to give it money? Do you want 0:34:08.200 --> 0:34:10.920 to provide it with whatever resource you want to? So 0:34:11.000 --> 0:34:14.360 there is kind of this like cohesive ecosystem where they 0:34:14.480 --> 0:34:18.360 kind of all dance together. How much these companies actually 0:34:18.440 --> 0:34:22.840 want to take state money, I think that's debatable. However, 0:34:23.000 --> 0:34:26.000 as AI and robot has become morem when sensitive and 0:34:26.040 --> 0:34:30.080 being recognized as not only an economic driver but potentially 0:34:30.120 --> 0:34:34.160 a military use or geopolitical I guess talking point. At 0:34:34.160 --> 0:34:36.920 this point, it is becoming more and more nationalized, not 0:34:36.960 --> 0:34:39.560 only in China but globally in the US and so on. 0:34:40.440 --> 0:34:44.240 Robotics is obviously an area where China is just straight 0:34:44.320 --> 0:34:47.120 up ahead of the United States, or at least according 0:34:47.160 --> 0:34:49.720 to all the videos on my Twitter and Instagram feed 0:34:49.880 --> 0:34:53.920 of humanoid robots and so forth. How much does you 0:34:53.920 --> 0:34:56.680 know when we were all kids when we thought of 0:34:56.800 --> 0:34:59.480 like AI, I think we thought of robots, right, we 0:34:59.520 --> 0:35:02.480 thought about fifteen one thousand or some version of it. 0:35:02.760 --> 0:35:04.600 And now when we think of AI, most people think 0:35:04.600 --> 0:35:07.600 of chatbots. But that's just one aspect of AI. For 0:35:07.840 --> 0:35:11.160 the advanced labage, whether we're talking about the deep sea 0:35:11.280 --> 0:35:14.040 so then Mini Max's or GLM and so forth, like, 0:35:14.800 --> 0:35:18.200 are they actively working hand in hand with some of 0:35:18.239 --> 0:35:21.919 the unit trees and advanced robotics companies to figure out 0:35:22.160 --> 0:35:25.320 how you can actually have that the true AI robot 0:35:25.360 --> 0:35:26.000 of the future. 0:35:26.520 --> 0:35:30.480 Yeah, I think China, having been the manufacturing hub of 0:35:30.560 --> 0:35:33.200 like literally everything in the sun over the last like 0:35:33.239 --> 0:35:37.480 three decades, has definitely an advantage of owning all the 0:35:37.480 --> 0:35:40.040 supply chain, right. And it's like not only just owning 0:35:40.080 --> 0:35:42.520 the supply chain, but you know, there are literally regions 0:35:42.560 --> 0:35:46.040 where that whole supply from raw material to like the 0:35:46.280 --> 0:35:49.359 result end product from OEM is all within like say, 0:35:49.400 --> 0:35:53.120 fifty kilometers of each other. So what we're seeing is 0:35:53.320 --> 0:35:56.920 a lot of American investors and entrepreneurs coming into China 0:35:56.960 --> 0:35:58.799 to kind of get a sense of that. And like 0:35:58.880 --> 0:36:01.759 you mentioned, because we've been so fixed in software, I 0:36:01.800 --> 0:36:05.520 think China having a very strong hardware background is now 0:36:05.600 --> 0:36:08.520 thinking about how can we actually integrate the software into 0:36:08.560 --> 0:36:12.200 the hardware. How ready that is to the mass market. 0:36:12.320 --> 0:36:15.040 I frankly don't think it's really there yet. So recently 0:36:15.040 --> 0:36:18.080 I just met with some robotic companies. They actually can't 0:36:18.080 --> 0:36:20.560 just plug in a Mini Max. You know, that's like 0:36:20.880 --> 0:36:24.279 for them, they need to actually get physical data. They 0:36:24.360 --> 0:36:26.400 This is where like you know, now all the hype 0:36:26.440 --> 0:36:29.239 is on world models physical AI. You know, that is 0:36:29.280 --> 0:36:32.960 a complete different set of kind of technology essentially, where 0:36:33.560 --> 0:36:37.800 without the three D data that these models need right now, 0:36:38.120 --> 0:36:41.160 the bottomneck right now is that you know, these hardware 0:36:41.200 --> 0:36:44.560 is these humanoids, quadrupids, dogs, whatever you want to call them, 0:36:44.719 --> 0:36:46.800 they cannot be powered by lms. 0:36:47.040 --> 0:36:47.880 That's number one. 0:36:48.080 --> 0:36:51.520 Number two is despite that China being very strong on hardware, 0:36:51.840 --> 0:36:54.200 the bottleneck is actually a lot of times in the 0:36:54.280 --> 0:36:57.880 integration as well as the battery solutions. You know, you 0:36:57.920 --> 0:37:01.000 think of China having very strong battery solutions, but most 0:37:01.040 --> 0:37:03.920 of these gadgets can't last more than like say two hours, 0:37:04.400 --> 0:37:06.640 and there's no one that's really come out with a 0:37:06.680 --> 0:37:09.840 better solution so far. What I've seen the most creative 0:37:09.840 --> 0:37:11.879 things so far is like you know, those glasses you wear, 0:37:11.960 --> 0:37:14.600 like the metaglasses, they kind of die within two hours. 0:37:14.840 --> 0:37:18.000 But China, like I Fly Tech or Rocket that's kind 0:37:18.040 --> 0:37:22.840 of a newer player startup, they created these battery capsules 0:37:22.840 --> 0:37:25.960 where you can just like stick on to your glasses. 0:37:26.239 --> 0:37:29.359 It's very lightweight, doesn't really affect your user experience, and 0:37:29.400 --> 0:37:31.600 that's actually able to kind of extend it by a 0:37:31.640 --> 0:37:34.719 few hours. So to go back to your question, is 0:37:34.800 --> 0:37:36.680 China trying to do physical AI? 0:37:36.840 --> 0:37:37.400 Definitely? 0:37:37.680 --> 0:37:39.920 What is their edge? I think it's still in manufacturing. 0:37:40.080 --> 0:37:42.359 Is their software good enough? I don't think anyone really 0:37:42.360 --> 0:37:44.239 has good enough software right now as of now. 0:37:44.680 --> 0:37:46.040 Wait, I'm just going to press you on this. 0:37:46.200 --> 0:37:49.600 So if we fast forward ten years, like, what would 0:37:49.640 --> 0:37:53.279 you say is most likely to be China's comparative advantage. 0:37:53.360 --> 0:37:57.600 Is it like the cheap, open source super optimized models. 0:37:57.880 --> 0:38:01.920 Is it software AI software that's like integrated with like 0:38:02.040 --> 0:38:05.680 industry and existing business or is it robotics and the 0:38:05.719 --> 0:38:07.200 sort of hardware side of AI. 0:38:07.840 --> 0:38:09.440 I think any. 0:38:11.360 --> 0:38:13.000 But we want to challenge here. 0:38:14.360 --> 0:38:16.040 A lot of these companies that exist ten years ago 0:38:16.120 --> 0:38:16.960 or not even five years. 0:38:17.040 --> 0:38:20.800 That's fair, okay, I can shorten the timeframe in three years. 0:38:20.880 --> 0:38:24.200 All right, So don't chase me down if I'm wrong 0:38:24.239 --> 0:38:27.480 in three years. But I think, you know, there's two parts. 0:38:27.520 --> 0:38:29.480 One is I think I agree with the hardware side. 0:38:29.560 --> 0:38:32.400 China is definitely going to have I think, more breakthroughs 0:38:32.680 --> 0:38:35.000 and have a lot of edge. Not only is the 0:38:35.040 --> 0:38:38.719 supply chain all domestically there, I think something overlooked by 0:38:38.760 --> 0:38:41.120 people is the fact that a lot of the know 0:38:41.200 --> 0:38:45.200 how is also there, and that's not easy to transfer overnight. 0:38:45.239 --> 0:38:47.879 You know Patrick McGee's book recently in his Apple book 0:38:47.960 --> 0:38:50.399 saying how Apple tried to move this whole supply chain 0:38:50.440 --> 0:38:53.040 to India. The biggest bottom neck is actually these like 0:38:53.120 --> 0:38:58.080 highly skilled laborist jobs that actually are so technical that 0:38:58.160 --> 0:39:01.880 cannot be even trained in one generation. It took decades 0:39:01.920 --> 0:39:04.960 to really train up the local community labor force whatnot. 0:39:05.320 --> 0:39:08.320 So that's still there now because of that ecosystem. 0:39:08.600 --> 0:39:09.360 A lot of these. 0:39:09.320 --> 0:39:13.400 Robots home appliance is whatnot. These tech gadgets are produced 0:39:13.520 --> 0:39:16.239 at less than fifty percent of the costs of where 0:39:16.280 --> 0:39:18.279 you could produce that anywhere else in the world. They 0:39:18.320 --> 0:39:22.399 are also extremely innovative. I've talked to people at ev 0:39:22.520 --> 0:39:26.000 companies just for example, to ship out a new model 0:39:26.440 --> 0:39:31.279 from ideation to production to hitting the floors that takes 0:39:31.320 --> 0:39:34.640 maybe less than fifteen months. Wow, But for traditional OEM 0:39:34.680 --> 0:39:36.600 like way for at least three to five years. Right, 0:39:36.840 --> 0:39:41.360 So there's the hardware side. I think another very underappreciated 0:39:41.440 --> 0:39:43.920 fact on the Chinese open source model is that people 0:39:44.000 --> 0:39:46.920 don't realize. So a year ago when I spoke to 0:39:47.120 --> 0:39:50.080 startups as semlic Valley, they were the most cost conscious, 0:39:50.080 --> 0:39:55.680 frankly less compliance conscious and gives very little the care 0:39:55.680 --> 0:39:59.120 about geopolitics. They were building on top of QUID. Now 0:39:59.200 --> 0:40:02.040 it's actually all a lot of Ai Native American enterprises 0:40:02.080 --> 0:40:05.239 building on Chinese open source because we're seeing headlines on 0:40:06.040 --> 0:40:09.960 the ROI is not like showing it's extremely expensive for 0:40:10.000 --> 0:40:16.000 these token maxing projects whatnot. So Harvey Cursor, they've talked 0:40:16.000 --> 0:40:19.759 about using a hybrid model where they will build majority 0:40:19.760 --> 0:40:22.440 on GLM or KIMMI, but kind of like what we 0:40:22.440 --> 0:40:26.399 talked about earlier where they use like Opus to act 0:40:26.440 --> 0:40:29.120 as a judge or a guidance. So I think that's 0:40:29.160 --> 0:40:32.759 something where we're continue to see and these companies are 0:40:32.840 --> 0:40:35.560 generating a lot of revenue, and going back to the 0:40:35.600 --> 0:40:37.520 fact that like a lot of them don't even have 0:40:37.640 --> 0:40:40.279 enough capability to support the demand that's coming through. 0:40:40.600 --> 0:40:42.879 That's such a fascinating idea, and it makes a lot 0:40:42.920 --> 0:40:46.520 of sense that at the application layer that probably a 0:40:46.560 --> 0:40:48.799 lot of different models can go into it. 0:40:48.880 --> 0:40:49.040 You know. 0:40:49.160 --> 0:40:50.480 By the way, I was talking to someone out a 0:40:50.520 --> 0:40:53.560 dinner recently and he said he thinks that AI writing 0:40:53.600 --> 0:40:56.799 will get a lot better when AI is embedded in 0:40:56.920 --> 0:41:00.799 humanoid robots, because then the will have this sort of 0:41:00.920 --> 0:41:03.560 groundedness in the real world where I have no idea 0:41:03.680 --> 0:41:05.960 this is true, he said. The reason his theory was 0:41:06.000 --> 0:41:08.040 that the reason why AI writing is still so weird 0:41:08.120 --> 0:41:11.759 is because it's in this disembodied data centers, and that is, 0:41:12.080 --> 0:41:14.759 as soon as they're really in robots, then they'll have 0:41:15.040 --> 0:41:17.200 a sort of real world groundedness. A right, have one 0:41:17.239 --> 0:41:20.880 last question, you know, after open Claw came out, I 0:41:20.920 --> 0:41:24.879 started seeing again my entire Twitter feed and answergram feed. 0:41:25.000 --> 0:41:27.400 I have done this to myself, but all I do 0:41:27.440 --> 0:41:31.040 all day is consume Chinese propaganada. But I started seeing 0:41:31.040 --> 0:41:33.520 all these videos, like all these grandmothers and stuff like 0:41:33.840 --> 0:41:36.279 setting up their claws and stuff like that. And I 0:41:36.320 --> 0:41:38.400 saw these videos and I said to myself, I just 0:41:38.440 --> 0:41:40.719 don't believe this. I think this is fake news. I 0:41:40.760 --> 0:41:43.759 do not actually believe that there's all these eighty year 0:41:43.800 --> 0:41:47.040 old Grammas or whatever really excited about setting up their 0:41:47.080 --> 0:41:49.800 open claw or whatever. Are those real? Like, what's the 0:41:49.880 --> 0:41:50.399 deal with that? 0:41:51.120 --> 0:41:51.279 Yeah? 0:41:51.320 --> 0:41:53.160 I think that was definitely a bit of a hype. 0:41:53.320 --> 0:41:53.839 I knew it. 0:41:54.040 --> 0:41:56.799 No, No, but I will say there were Grammas lining 0:41:56.920 --> 0:41:57.680 up to get it done. 0:41:57.800 --> 0:42:01.520 Okay, I'll answer this twofold on the kind of the 0:42:01.600 --> 0:42:02.560 surface is. 0:42:02.640 --> 0:42:08.120 I think Chinese aunties, uncles whatnot. They are just much 0:42:08.160 --> 0:42:11.440 more open to technology because you know, you go around, 0:42:11.480 --> 0:42:14.160 whether it's by force or by nature. You know, you 0:42:14.200 --> 0:42:17.920 can't really navigate modern Chinese life without being on alipay. 0:42:18.040 --> 0:42:18.799 No, you really can't. 0:42:18.800 --> 0:42:22.080 You can't like buy Starbucks in Beijing without like having 0:42:22.160 --> 0:42:22.720 we paid. 0:42:22.800 --> 0:42:25.640 After COVID, I went back to I think Shanghai for 0:42:25.680 --> 0:42:28.120 the first time, and I was sitting at a restaurant 0:42:28.200 --> 0:42:30.680 just like waiting for someone to help me order food. 0:42:30.840 --> 0:42:33.200 No one came because they're just like, why are you 0:42:33.360 --> 0:42:35.480 so Like, why are you a caveman? Don't you know 0:42:35.520 --> 0:42:38.680 how to like scan the QR code on your table? Okay, 0:42:38.680 --> 0:42:43.160 So that Tangent's side, I think the overall optimism around. 0:42:43.239 --> 0:42:47.799 Technology is very different from the West because in the 0:42:47.880 --> 0:42:50.920 last twenty thirty years, a lot of rural eras in 0:42:51.000 --> 0:42:56.080 China literally could not access resources, information, goods, whatever that, 0:42:56.480 --> 0:42:59.680 you know, like big cities could not until these super 0:42:59.719 --> 0:43:02.319 apps came about. So a lot of people don't have 0:43:02.360 --> 0:43:04.280 TVs in their homes and they live in a village 0:43:04.280 --> 0:43:06.759 and there may be annual household income is like one 0:43:06.800 --> 0:43:09.480 thousand dollars, but they will have a smartphone and that 0:43:09.640 --> 0:43:12.560 smartphone will be able to actually enable them to get 0:43:12.680 --> 0:43:16.200 micro loans to purchase goods, you know, to help their 0:43:16.280 --> 0:43:20.279 kids access information online whatever that is. So technology is 0:43:20.480 --> 0:43:24.440 very much kind of accepted and respected and actually like 0:43:24.480 --> 0:43:26.759 big tech is loved, like if you work for one 0:43:26.800 --> 0:43:29.520 of big tech, you are like a pride of the family. 0:43:30.000 --> 0:43:33.279 So there's that very culture aspect of it. Then is 0:43:33.360 --> 0:43:36.080 like going back to the super apps. So I think 0:43:36.120 --> 0:43:39.120 the open Claw frenzy was interesting because some people say 0:43:39.200 --> 0:43:42.400 it was the first agent that you know, Chinese people 0:43:42.440 --> 0:43:44.320 could get their hands on, like a Western agent that 0:43:44.360 --> 0:43:47.160 they can get their hands on because you know, anthropic 0:43:47.200 --> 0:43:49.279 and open it doesn't actually operated in China. You can't 0:43:49.280 --> 0:43:53.520 access that. So when Tensan and Ali Baba try to 0:43:53.960 --> 0:43:58.719 embed open claw products into their own like series of 0:43:58.800 --> 0:44:02.560 products or business whatever offerings. It got people really excited, 0:44:02.960 --> 0:44:06.239 and because of these super app models that they have, 0:44:06.719 --> 0:44:10.000 it was a very natural way for people to access them. 0:44:10.000 --> 0:44:12.759 There's a functional adjacency to you know, the search bar 0:44:13.160 --> 0:44:15.560 and then opening up an open claw and then trying 0:44:15.600 --> 0:44:18.320 to like run like you know, manage your mini programs 0:44:18.320 --> 0:44:21.200 within ten cent we chat and then try to order something. 0:44:21.239 --> 0:44:23.600 So all of that kind of took off. But that said, 0:44:24.080 --> 0:44:28.080 actually the Chinese government again regulators acted very swiftly in 0:44:28.160 --> 0:44:31.680 the beginning. I think local government, like we see try 0:44:31.719 --> 0:44:36.040 to even encourage local businesses to embrace open claw. But 0:44:36.239 --> 0:44:39.839 the Beijing government immediately said, guys, actually be very very 0:44:39.840 --> 0:44:43.120 careful of your data privacy. Your security banks didn't do 0:44:43.120 --> 0:44:45.560 this as so you shouldn't do this. Be mindful of 0:44:45.920 --> 0:44:48.600 what you're doing with this technology. And then the big 0:44:48.640 --> 0:44:50.719 tech kind of rolled back a bit of their marketing 0:44:51.000 --> 0:44:54.880 and you can see, actually hilariously there were advertisements for 0:44:55.080 --> 0:44:58.840 helping these aunties and uncles how to uninstall these open claw. 0:44:58.640 --> 0:45:01.719 On their gadget. So anyway, that's a bit of a 0:45:01.840 --> 0:45:02.839 kind of background on that. 0:45:03.239 --> 0:45:05.880 All right, Graceha, thank you so much for coming on 0:45:05.920 --> 0:45:07.960 odd lot. It's great to connect with you here in 0:45:08.000 --> 0:45:11.200 Hong Kong. And we'll have you back in three years, no, 0:45:11.280 --> 0:45:14.720 hopefully before then. We'll have you back at least certainly 0:45:14.719 --> 0:45:17.080 in three years to see how your predictions held up. 0:45:17.320 --> 0:45:30.880 Thanks so much, Tracy. 0:45:30.920 --> 0:45:33.600 That was fun. There was a lot of interesting ideas 0:45:33.719 --> 0:45:36.840 in a fairly short conversation. But one thing specifically and 0:45:36.840 --> 0:45:39.920 then that like sort of stands out to me is 0:45:40.000 --> 0:45:44.279 thinking about some of these application companies, how much it 0:45:44.360 --> 0:45:46.959 makes sense for them to sort of Yeah, they'll use 0:45:47.080 --> 0:45:49.759 like a state of the art American closed model for 0:45:49.840 --> 0:45:53.319 like some of the work, but then other just you know, 0:45:53.680 --> 0:45:56.960 almost as capable models underneath, so you have like a 0:45:57.200 --> 0:46:00.600 legal AI app like a Harvey or something. Sure, some 0:46:00.640 --> 0:46:03.680 of these open source models work well for some tasks 0:46:03.719 --> 0:46:06.640 within that context, And how much it makes sense to 0:46:06.719 --> 0:46:10.240 sort of combine them under one app layer. 0:46:10.480 --> 0:46:14.560 Yeah, Well, you don't need the cutting edge model for everything, right, 0:46:14.840 --> 0:46:17.960 but like if you can get some of the cutting 0:46:18.080 --> 0:46:22.000 edge model combined with like the cheapness of the open 0:46:22.000 --> 0:46:24.959 source thing, like that seems like a pretty good deal 0:46:25.200 --> 0:46:26.200 for a lot of companies. 0:46:26.440 --> 0:46:30.200 Yeah, definitely. I also think like, how is the massive 0:46:30.320 --> 0:46:34.279 manufacturing edge that China has not going to just keep 0:46:34.320 --> 0:46:38.480 compounding itself. I mean, this is like the multi trillion 0:46:38.520 --> 0:46:42.279 dollar question of the entire world. But it really does. 0:46:42.400 --> 0:46:44.680 I mean, when you think about, Okay, there's a sort 0:46:44.680 --> 0:46:49.160 of presumably natural synergy, there's all this real world physical 0:46:49.280 --> 0:46:54.439 data that Chinese manufacturing companies can theoretically get from their 0:46:54.560 --> 0:46:58.480 various robotic vacuum cleaners and so forth, and then feed 0:46:58.520 --> 0:47:02.600 those into certain model then we call AI. That really 0:47:02.640 --> 0:47:05.400 does seem like a potential leg up that you know, 0:47:05.520 --> 0:47:09.520 just done the data collection alone from all those physical things, 0:47:09.560 --> 0:47:12.040 like a huge potential edge over the next several years. 0:47:12.120 --> 0:47:14.799 Yeah, although it was very interesting, Grace was saying that 0:47:14.840 --> 0:47:17.600 it's not as developed or as structured as it is 0:47:17.640 --> 0:47:20.520 in the US, because I had thought the same thing, like, 0:47:20.560 --> 0:47:23.960 if everyone's talking over we chat, if everyone's paying over 0:47:24.040 --> 0:47:27.280 we pay, you must have oodles and oodles of data. 0:47:27.400 --> 0:47:31.239 But yeah, that was interesting. The ecosystem thing, I think 0:47:31.360 --> 0:47:33.960 is really important, and it just seems like really hard 0:47:34.040 --> 0:47:37.640 to get an ecosystem kind of going from scratch. And 0:47:37.719 --> 0:47:41.000 I remember maybe it was with Dan Wong, but someone 0:47:41.080 --> 0:47:44.000 describing how like, if you go to shen Zen, you 0:47:44.040 --> 0:47:48.920 can basically start like an entire company manufacturing a physical 0:47:49.040 --> 0:47:52.640 thing because every single supplier is there. You just go 0:47:52.719 --> 0:47:57.040 from like one storefront to another storefront to another storefront. 0:47:57.600 --> 0:48:00.000 I don't think that can be replicated anywhere in the US. 0:48:01.040 --> 0:48:02.640 I mean, this is a bit of a tangent. But 0:48:02.719 --> 0:48:06.279 you know this is economists talk about quote agglomeration all 0:48:06.320 --> 0:48:09.880 the time. The advantage is exactly that of having everything 0:48:09.960 --> 0:48:12.600 there and then you build these deep networks. One thing 0:48:12.600 --> 0:48:14.360 I find to be a little strange, and again this 0:48:14.520 --> 0:48:19.279 is a tangent, is how San Francisco concentrated AI is, 0:48:19.760 --> 0:48:22.120 even though at the American level it is sort of 0:48:22.160 --> 0:48:25.600 pure like desk work right Like, it's not the sort 0:48:25.600 --> 0:48:29.959 of we need the bolts manufacturer, we need the servos manufacturer, 0:48:30.040 --> 0:48:34.200 we need robovision manufacturer, and yet it's still so agglomerated. 0:48:34.760 --> 0:48:37.040 Or the fact that finance is so agglomerated in New 0:48:37.120 --> 0:48:40.279 York City. I find that or conglomerated. I find that 0:48:40.320 --> 0:48:43.440 to be a little odd. So, but yeah, I don't 0:48:43.600 --> 0:48:47.720 conglomerated is a conglomerated I don't know whether it's conglomerated 0:48:47.800 --> 0:48:50.799 or agglomerated. I'm just gonna say glomerated. We know that 0:48:50.880 --> 0:48:53.359 we are certain industries in the US that are quote 0:48:53.360 --> 0:48:57.080 conglomerated one way or another. But yeah, I did find 0:48:57.120 --> 0:49:00.400 that to be And man, these like Mini Mac billion 0:49:00.440 --> 0:49:03.640 dollar company that's like nothing compared to the valuations that 0:49:03.680 --> 0:49:06.719 we see with American companies. Pretty wild stuff. Also, the 0:49:06.719 --> 0:49:09.440 point we should do more, I mean, there's a million 0:49:09.520 --> 0:49:11.960 we gotta do we have and there are plenty of 0:49:12.000 --> 0:49:14.160 non a I things to do, so you can't just 0:49:14.239 --> 0:49:16.880 keep saying we should do an episode on X or Y. 0:49:17.360 --> 0:49:20.880 But data markets and like the idea of like, okay, 0:49:20.920 --> 0:49:26.719 these Chinese companies can buy very pricey proprietary data after 0:49:26.760 --> 0:49:29.839 some exclusivity window. There's some interesting stuffing to down there. 0:49:29.920 --> 0:49:32.680 Yeah, the monetization was really interesting to hear. 0:49:32.800 --> 0:49:34.319 Okay, shall we leave it there for now? 0:49:34.360 --> 0:49:35.120 Let's leave it there. 0:49:35.360 --> 0:49:37.920 This has been another episode of the All Thoughts podcast. 0:49:38.040 --> 0:49:39.080 I'm Tracy Alloway. 0:49:39.160 --> 0:49:41.560 You can follow me at Tracy Alloway. 0:49:41.320 --> 0:49:43.640 And I'm sure wasn't Thal. You can follow me at 0:49:43.680 --> 0:49:47.160 the Stalwart. Follow our guest Grace Show. She's at Grace 0:49:47.440 --> 0:49:51.120 MZ show and check out her subtck ai prom Follow 0:49:51.160 --> 0:49:55.080 our producers Carmen Rodriguez at Carman armand Dashel Bennett at Dashbot, 0:49:55.160 --> 0:49:58.880 Calebrooks at Cale Brooks and Kevin Lozano at Kevin Lloyd 0:49:58.920 --> 0:50:02.040 Lozano and from our Oddlots content. Go to Bloomberg dot 0:50:02.040 --> 0:50:04.600 com slash odd loss for the daily newsletter and all 0:50:04.640 --> 0:50:06.839 of our episodes, and you can chat about all these 0:50:06.880 --> 0:50:10.640 topics twenty four seven in our discord Discord dot gg 0:50:10.680 --> 0:50:11.640 slash odlines. 0:50:11.760 --> 0:50:13.920 And if you like odd Lots, if you want us 0:50:13.960 --> 0:50:16.440 to do an episode on data markets, then please leave 0:50:16.520 --> 0:50:20.240 us a positive review on your favorite podcast platform. And remember, 0:50:20.280 --> 0:50:22.920 if you are a Bloomberg subscriber, you can listen to 0:50:23.040 --> 0:50:26.359 all of our episodes absolutely ad free. All you need 0:50:26.400 --> 0:50:29.200 to do is find the Bloomberg channel on Apple Podcasts 0:50:29.280 --> 0:50:32.120 and follow the instructions there. Thanks for listening.