WEBVTT - Baidu's CFO on How It Became a Full-Stack AI Player 0:00:02.720 --> 0:00:16.400 Bloomberg Audio Studios, Podcasts, Radio News. 0:00:18.480 --> 0:00:21.079 Hello and welcome to another episode of The Odd Laws podcast. 0:00:21.200 --> 0:00:23.360 I'm Joe Wisenthal and I'm Tracy Alloway. 0:00:23.480 --> 0:00:25.200 So we were in a Hong Kong recently. That was 0:00:25.200 --> 0:00:25.680 a lot of fun. 0:00:25.760 --> 0:00:26.600 I'd love to be back. 0:00:26.880 --> 0:00:29.920 Yeah, I haven't been back for four years. Not much 0:00:29.920 --> 0:00:32.159 has changed. Actually, I was kind of surprised and you 0:00:32.200 --> 0:00:35.280 expected LUs than I expected. But I am really glad 0:00:35.320 --> 0:00:37.800 we went back because obviously one of the big talking 0:00:37.840 --> 0:00:41.919 points in markets right now is competition between US versus 0:00:42.040 --> 0:00:46.040 Chinese AI. And we finally got a chance to talk 0:00:46.120 --> 0:00:49.880 to a couple of high level executives of Chinese tech 0:00:49.920 --> 0:00:53.520 companies who are actually making all the big capital allocation 0:00:53.720 --> 0:00:55.400 decisions when it comes to the AI race. 0:00:55.560 --> 0:00:57.800 Right, it felt like we're in this moment where there's 0:00:57.840 --> 0:00:59.680 been the I mean, the way I think about it, 0:01:00.000 --> 0:01:03.120 there's been Chinese internet giants, but they concentrated on China, right, 0:01:03.280 --> 0:01:05.680 there's been American internet giants that basically had the rest 0:01:05.680 --> 0:01:08.200 of the world. And whether we're talking about AI or 0:01:08.240 --> 0:01:10.920 self driving cars, we're going to see the first sort 0:01:10.920 --> 0:01:15.400 of like real head to head battle on internet companies specifically, 0:01:15.959 --> 0:01:18.160 and where they're like competing for playing the same game 0:01:18.280 --> 0:01:20.160 on some of the same markets. And of course we 0:01:20.240 --> 0:01:23.720 know American companies can use AI models built by China, 0:01:23.880 --> 0:01:26.040 et cetera, and so there are all kinds of options 0:01:26.080 --> 0:01:28.600 for people. So it's like really interesting to say, like, Okay, 0:01:28.840 --> 0:01:30.960 this clash is actually like it's happening. 0:01:31.040 --> 0:01:33.280 It's a good time to talk to a Chinese tech 0:01:33.319 --> 0:01:34.160 executive for. 0:01:34.080 --> 0:01:36.039 Sure, that's right. So the reason we were back in 0:01:36.080 --> 0:01:39.120 Hong Kong is because we were at the Bloomberg invest Conference. 0:01:39.160 --> 0:01:41.760 Though we also through an Odd Lots trivia night while 0:01:41.760 --> 0:01:42.440 we were in our. 0:01:42.560 --> 0:01:45.640 First non US overseas all Thoughts quiz night. 0:01:45.920 --> 0:01:47.760 Yeah, that was a lot of fun and I'm sure 0:01:47.800 --> 0:01:49.880 we'll come back and do that again. But we were 0:01:49.920 --> 0:01:52.320 at the Bloomberg invest conference and so we had the 0:01:52.440 --> 0:01:55.000 chance to speak with the CFO of by Do, Henry Hoo. 0:01:55.200 --> 0:01:57.920 So check it out. We truly have the perfect guest 0:01:58.000 --> 0:02:01.320 we're going to be speaking with by Do CFO Henry. So, Henry, 0:02:01.640 --> 0:02:03.600 thank you so much for coming on. 0:02:03.400 --> 0:02:05.200 Outlot, thanks for having me. 0:02:05.520 --> 0:02:08.079 And it's a great season, I think Hong Kong and 0:02:08.120 --> 0:02:10.440 it definitely is great to see in both Joan Tracy 0:02:10.520 --> 0:02:10.799 thank you. 0:02:11.160 --> 0:02:12.960 Very nice of you to say. So why do we 0:02:13.040 --> 0:02:16.040 start with this? You know, obviously I feel like half 0:02:16.080 --> 0:02:18.840 the conversations are probably about AI these days. But within AI, 0:02:19.680 --> 0:02:22.680 BYD is a full stack player, right You have cloud, 0:02:22.720 --> 0:02:25.000 you have the application layer, you have your own chips 0:02:25.000 --> 0:02:28.680 and of course your own model. As the CFO, you 0:02:28.800 --> 0:02:32.839 might have to think about prioritization, et cetera. Is there 0:02:32.880 --> 0:02:36.840 one layer of the stack that you feel is a 0:02:36.960 --> 0:02:41.000 must win for by do when you think about resource allocation? 0:02:41.160 --> 0:02:43.120 Is there a layer where it's like, okay, this is 0:02:43.160 --> 0:02:45.000 an area where we have to win. 0:02:45.320 --> 0:02:46.080 Thank you so much. 0:02:46.120 --> 0:02:48.920 And I think you'll probably put the tough question in end, 0:02:48.960 --> 0:02:51.160 but I think it's probably the most difficult question to 0:02:51.200 --> 0:02:54.519 start with. So I think the very unique thing today 0:02:54.600 --> 0:02:57.560 is I think the entire AI has been shifting from 0:02:57.680 --> 0:03:02.280 infrastructure to applications and from model to agents. I think 0:03:02.320 --> 0:03:05.120 that's actually the bad job. I think within that is, 0:03:05.120 --> 0:03:07.480 frankly speaking, right now, is very difficult to say at 0:03:07.480 --> 0:03:11.400 this moment which parties must have because in my view, 0:03:12.040 --> 0:03:15.200 the chap is infrastructures. You need to have a great 0:03:15.240 --> 0:03:18.920 model to bridge the capability. The cloud is a deployment 0:03:19.200 --> 0:03:22.600 of that capability, and obviously the monitization and all the 0:03:22.800 --> 0:03:25.440 ROI questions, especially for the people like me, I say, oh, 0:03:25.440 --> 0:03:28.720 we focus on that is on application layers. So without 0:03:28.800 --> 0:03:32.480 any of that, this ROI cannot work. So to answer 0:03:32.480 --> 0:03:34.760 a question, I think the key thing if I have 0:03:34.840 --> 0:03:38.120 to pick one, is Cloud, because Cloud at this moment 0:03:38.400 --> 0:03:41.440 is a platform. You cannot only hosting Arnie, which is 0:03:41.480 --> 0:03:43.520 our own model, but also I can work in a 0:03:43.600 --> 0:03:46.440 very open to hosting other models. And the my chip 0:03:46.600 --> 0:03:50.080 which connecting to my cloud platform can also help on 0:03:50.280 --> 0:03:53.040 inference because right now the play training is important, but 0:03:53.160 --> 0:03:56.000 eighty percent of the incremental demand today on a token 0:03:56.320 --> 0:03:59.200 are inference related. I think this part of the full 0:03:59.240 --> 0:04:01.440 picture is what I want to emphasize. But you know, 0:04:01.480 --> 0:04:03.280 given a tough question, if I want a big one 0:04:03.440 --> 0:04:05.920 as a student ABCD, I want big number C, which 0:04:05.960 --> 0:04:06.480 is my cloud. 0:04:07.560 --> 0:04:09.200 I'm going to ask you a question which I think 0:04:09.320 --> 0:04:11.960 is going to become standard for financial journalists in the 0:04:12.000 --> 0:04:16.200 same way we ask about headcount and expansion plans. What's 0:04:16.200 --> 0:04:20.600 your token budget? Is it bigger than Joe's? I mean 0:04:20.640 --> 0:04:23.880 the token the token budget for buy does or how 0:04:23.880 --> 0:04:26.159 do you measure? I'll ask it in a slightly different way. 0:04:26.200 --> 0:04:29.600 How do you measure what you're gaining from your token spend? 0:04:29.640 --> 0:04:31.040 How do you measure productivity? 0:04:31.160 --> 0:04:31.320 Yeah? 0:04:31.360 --> 0:04:36.440 Sure, I want to categorize probably two buckets. One, if 0:04:36.720 --> 0:04:41.040 we consume computing power to reach a higher level of 0:04:41.120 --> 0:04:45.520 technology standard you know, the ADI, the how would model 0:04:45.760 --> 0:04:48.919 form and also how harnys can be designed to deliver 0:04:49.000 --> 0:04:52.000 better results, So the are R and D efforts. However, 0:04:52.279 --> 0:04:55.280 as also a tech house, we also deliver those know 0:04:55.320 --> 0:04:59.039 how to our external clients with different verticals. I think 0:04:59.520 --> 0:05:03.520 if I'm measuring our internal consumption of the token, I 0:05:03.560 --> 0:05:06.440 really want to like how better and how efficient our 0:05:06.520 --> 0:05:10.000 technology can be developed. So that's on one side. However, 0:05:10.080 --> 0:05:12.520 on the other end, what I think the ROI is 0:05:12.760 --> 0:05:17.240 more relevant is how many real tasks that the open 0:05:17.279 --> 0:05:20.719 claw and for example our own application called do mate 0:05:20.839 --> 0:05:23.599 also is a real agency in human digits and other 0:05:23.640 --> 0:05:26.440 things can do the task. So I think these two 0:05:26.440 --> 0:05:28.960 different measurements are important in the way that right now 0:05:29.320 --> 0:05:33.159 there are two things are better than last year when 0:05:33.200 --> 0:05:36.160 it's a foundational model getting much better. And number two 0:05:36.480 --> 0:05:39.480 is the framework i e. Open Claw and other things 0:05:39.800 --> 0:05:42.600 can link up the foundation model capability to the rail 0:05:42.680 --> 0:05:46.560 world task from chatting on something to doing something and 0:05:46.720 --> 0:05:50.600 completing on something. I think completion part of the tokens 0:05:50.800 --> 0:05:54.080 is more important today, and I think consumption internally will 0:05:54.080 --> 0:05:56.840 actually encourage the people do that. But think about that, 0:05:56.880 --> 0:05:59.520 even last year or year before, everyone is actually beefing 0:05:59.560 --> 0:06:02.200 up budgets. I think that budget is all there, but 0:06:02.320 --> 0:06:03.960 the completion utask is more important. 0:06:03.960 --> 0:06:06.040 But let me just press you on this question a 0:06:06.040 --> 0:06:08.840 little bit further. So let's say Tracy and I are 0:06:09.240 --> 0:06:12.840 Let's say we worked for Baydo in the same department, 0:06:13.120 --> 0:06:16.479 I don't know, some department of yours. Would we have 0:06:17.040 --> 0:06:20.520 identical token budgets or would you have one way of saying, 0:06:20.680 --> 0:06:23.400 you know what, Joe, Tracy is actually finding ways to 0:06:23.400 --> 0:06:26.160 get more value out of AI than you are, So 0:06:26.200 --> 0:06:28.599 I'm going to increase her budget. Like do you make 0:06:28.680 --> 0:06:32.440 decisions like that? And do you have measurement techniques to 0:06:32.520 --> 0:06:35.520 see like this person really should get ten times the 0:06:35.520 --> 0:06:38.320 token budget of another person because they have figured out 0:06:38.320 --> 0:06:40.599 how to get a lot of juice from the squeeze, 0:06:40.640 --> 0:06:41.159 so to speak. 0:06:41.240 --> 0:06:43.560 You know, Joe, given the question asking I seeing pro 0:06:43.680 --> 0:06:47.279 next time, I give you multiple con Yeah, I think 0:06:47.600 --> 0:06:50.520 right now, the technology evolved very fast. I think that's 0:06:50.560 --> 0:06:53.000 the beauty part of AI. So we don't want to 0:06:53.040 --> 0:06:56.080 constrain by us out by before thinking through something, we 0:06:56.240 --> 0:06:58.520 just install certain policy by saying, you know, these are 0:06:58.560 --> 0:07:03.360 employees with defined the token number by the titles or similarities. 0:07:03.760 --> 0:07:06.120 I don't think that's the way it works. So I 0:07:06.160 --> 0:07:09.080 think we want to more open and more nimbo in 0:07:09.120 --> 0:07:13.000 a way that given enough token to the individuals to 0:07:13.120 --> 0:07:17.200 empower their internal RD efforts. But on the other end, 0:07:17.360 --> 0:07:19.760 we do have a lot of efforts to make sure 0:07:19.760 --> 0:07:22.800 the token costs become dramatic coming down. I think the 0:07:22.840 --> 0:07:26.040 cost is coming out very fast before you even think 0:07:26.040 --> 0:07:29.040 about getting a policy. Maybe the unit cost is coming 0:07:29.080 --> 0:07:31.440 down half in like a few weeks. So we need 0:07:31.480 --> 0:07:34.320 to think about the speed and the cost and output 0:07:34.320 --> 0:07:37.360 efficiency these three parameters as a package, not only on 0:07:37.400 --> 0:07:39.280 the number of tokens. That's when I say, on the 0:07:39.320 --> 0:07:42.040 other end, you feel very interesting facts. So right now 0:07:42.080 --> 0:07:44.520 you know we are recruiting a lot of younger talents 0:07:44.560 --> 0:07:47.160 to then even for Baydu, which is you know, twenty 0:07:47.240 --> 0:07:50.600 plus years listed a public company. So my seeing is 0:07:51.240 --> 0:07:54.840 the kids actually getting smarter than people expected, so they 0:07:54.880 --> 0:07:57.440 were not wasted the tokens you'll give to them, so 0:07:57.480 --> 0:08:00.760 they have a sensible judgment about what are the task 0:08:00.800 --> 0:08:03.800 they need to prioritize because they're working out agents and 0:08:03.840 --> 0:08:07.320 the models. The model actually helping them also prioritize autopaths 0:08:07.360 --> 0:08:09.960 they have. So I think the power of the technology 0:08:10.000 --> 0:08:12.160 today is it is not only at the tools. So 0:08:12.200 --> 0:08:15.120 that's actually the key concept on to mention it is 0:08:15.160 --> 0:08:18.120 not only a tool, it's a mindset, and the mindset 0:08:18.200 --> 0:08:21.480 become automatic and more intelligent. That if people can work 0:08:21.520 --> 0:08:24.040 on that well and have a new relationship with agents 0:08:24.120 --> 0:08:27.280 and with a model, some of the old questions we 0:08:27.400 --> 0:08:29.680 kind of struggle ourselves will be kind of diminished and 0:08:29.760 --> 0:08:30.400 less important. 0:08:30.600 --> 0:08:33.880 Okay, so no token maxing at BYDEO. But since you 0:08:33.960 --> 0:08:36.160 brought up talent, one thing I'm very curious about is 0:08:36.200 --> 0:08:39.439 we know the competition for the top engineers is so 0:08:39.520 --> 0:08:41.920 intense right now, and in the US we see these 0:08:41.960 --> 0:08:45.760 headlines where engineers are treated like sports stars. You know, 0:08:45.800 --> 0:08:49.520 they're being treated for millions of dollars or whatever. What 0:08:49.679 --> 0:08:53.640 is Bydeo's pitch to top talent, Like, if you're trying 0:08:53.640 --> 0:08:56.440 to attract someone to the company, what is it you say, 0:08:56.480 --> 0:08:58.079 to them that makes them want to work for buy 0:08:58.160 --> 0:09:00.640 do versus another tech. 0:09:00.480 --> 0:09:03.679 Firm's a great question, So let's bring a different perspective. 0:09:03.920 --> 0:09:07.400 I think, you know, we are a technology company. Previously, 0:09:07.679 --> 0:09:10.440 I think the priorities we empower our clients to be 0:09:10.520 --> 0:09:13.960 more intelligent. We give them more technology tools to help 0:09:14.000 --> 0:09:17.080 them to remove a move from the you know, the 0:09:17.120 --> 0:09:20.719 traditional it to the cloud environment, you know, such as that. 0:09:21.320 --> 0:09:24.440 But right now I think AI, especially for the big 0:09:24.480 --> 0:09:28.280 corporation like us, also change us as well. We need 0:09:28.320 --> 0:09:32.320 to think about the cultural change organization change not only 0:09:32.400 --> 0:09:35.839 as an organization and a company, but also how AI 0:09:35.920 --> 0:09:39.440 empower ourselves. So it's actually equally important to do something 0:09:39.480 --> 0:09:42.719 for client versus think about new tools affecting ourselves. 0:09:42.920 --> 0:09:43.800 So Trace is right. 0:09:44.040 --> 0:09:45.720 I think there are a few things we actually make 0:09:45.800 --> 0:09:49.400 a lot of different thinkings and some of the new initiatives. 0:09:49.440 --> 0:09:52.520 First of all, we're probably among a few companies in 0:09:52.600 --> 0:09:56.040 China still very open and even increasing the campus recruiting 0:09:56.200 --> 0:10:00.000 and the focus on the younger talents and number two recents. 0:10:00.200 --> 0:10:03.280 We also task the senior people not only look at 0:10:03.600 --> 0:10:06.840 you know, the current reporting structure, but also in the 0:10:06.920 --> 0:10:10.480 real mentor relationship with the younger, growing piece of the 0:10:10.559 --> 0:10:13.959 human capital in the company. But more importantly, I think 0:10:14.200 --> 0:10:17.280 it is really about giving people more autonomy to work 0:10:17.320 --> 0:10:20.600 in a company, so more trust and more autonomy and 0:10:20.760 --> 0:10:23.720 give them more real work. And you know, there's a 0:10:23.720 --> 0:10:26.520 one concept called one person company, right, so we're very 0:10:26.520 --> 0:10:29.080 happy to working with one person company because they actually 0:10:29.160 --> 0:10:31.360 use our AI to very nicely and willing to pay 0:10:31.400 --> 0:10:35.280 a lot of revenue to our products given the quality. However, 0:10:35.800 --> 0:10:38.720 within the company, we'll also encourage people to be the 0:10:38.720 --> 0:10:42.680 one person team so they can actually use the agents 0:10:42.760 --> 0:10:44.960 to work on a lot of internal tasks. So internally 0:10:44.960 --> 0:10:46.480 you have listen a little to call the doo Doo 0:10:46.760 --> 0:10:49.040 in Chinese is a very kind of nicky name, which 0:10:49.080 --> 0:10:52.360 is actually our internal kind of open clock, similar tools 0:10:52.400 --> 0:10:55.080 and which actually enhancing people's efficiency. And the true point 0:10:55.080 --> 0:10:58.280 I think give me more people, more autonomy, more trust 0:10:58.880 --> 0:11:01.559 and the more room to grow and attracting new talents 0:11:01.559 --> 0:11:04.360 I think equally all kind of very important to change yourselves. 0:11:04.360 --> 0:11:07.800 But also you know, reporting lines and organized instructure need 0:11:07.840 --> 0:11:09.640 to come with it to make sure that people can 0:11:09.720 --> 0:11:12.480 deliver the results. And the last note, I want to 0:11:12.480 --> 0:11:15.680 mention the key things that people see the application is important. 0:11:15.720 --> 0:11:17.439 They can work on a full stack in by DO, 0:11:17.520 --> 0:11:20.120 which is very unique value in the China tax space. 0:11:20.240 --> 0:11:23.440 You know, it just occurred to me. American companies are 0:11:23.480 --> 0:11:27.000 kind of becoming more Chinese in the sense that they're 0:11:27.200 --> 0:11:29.959 doing more vertical integration. Like that's sort of a long 0:11:30.200 --> 0:11:33.960 history here of sort of the whole thing, and now 0:11:33.960 --> 0:11:37.640 we see one of this phenomenon is that every American 0:11:37.679 --> 0:11:40.520 company like they want to even start designing and selling 0:11:40.520 --> 0:11:43.120 their own chips, their own silicon, which is something that 0:11:43.200 --> 0:11:45.160 you're doing your own business and you've had it for 0:11:45.200 --> 0:11:47.800 a while, and I'm trying to wrap my head, like 0:11:48.160 --> 0:11:51.320 what is the rationale? How much is it about just 0:11:51.760 --> 0:11:54.440 wanting to be able to control your own fate more 0:11:54.840 --> 0:11:57.720 and so wanting to control more of the supply chain 0:11:58.240 --> 0:12:03.800 versus having it that optimally aligns with the model that 0:12:03.840 --> 0:12:07.320 you're working on, because those are distinct priorities. So what 0:12:07.440 --> 0:12:10.319 is the real rationale for having custom Seligan? 0:12:10.559 --> 0:12:13.920 Yeah, I think thanks for the tech trend in the 0:12:14.000 --> 0:12:15.280 past kind of year and two. 0:12:15.600 --> 0:12:18.000 So if you look at the entire. 0:12:18.040 --> 0:12:21.800 Computing power consumed, for example, last year, most of the 0:12:21.840 --> 0:12:26.160 consumptions actually relating to the training of large foundation model, 0:12:26.200 --> 0:12:28.439 but right now you know many of them going to 0:12:28.520 --> 0:12:31.439 the influence and the complete tasks, and if you look 0:12:31.480 --> 0:12:34.080 at the different stacks, I think right now we are 0:12:34.080 --> 0:12:37.600 actually fitting to an incremental growing piece of the market 0:12:37.679 --> 0:12:40.360 which is well defined with a clear boundary, which is 0:12:40.400 --> 0:12:44.600 not focused on pre training for a very super scale 0:12:44.720 --> 0:12:48.760 foundation model. However, the influence application is important. So to 0:12:48.800 --> 0:12:51.960 a point, I think our cheir products appointing of cloud 0:12:52.200 --> 0:12:55.480 focusing on the inference and application is a unique way 0:12:55.520 --> 0:12:58.600 of we see the positive network effects. I think that's 0:12:58.840 --> 0:13:02.080 where the area will wants and also given the issue 0:13:02.080 --> 0:13:05.160 you mentioned, I think within that defined areas, we I 0:13:05.200 --> 0:13:07.880 feel pretty confident regarding all the issues you mentioned, and 0:13:08.040 --> 0:13:10.680 because on the supply and demand side, we can find 0:13:10.800 --> 0:13:14.000 the good match within the emergent market category, within the 0:13:14.000 --> 0:13:28.520 inference and application markets. 0:13:31.360 --> 0:13:35.320 So hypothetically you could try to do everything right the 0:13:35.440 --> 0:13:38.800 full stack, and I guess the capital investment you would 0:13:38.840 --> 0:13:42.800 need to do that is also hypothetically unlimited at this 0:13:42.880 --> 0:13:45.000 moment in time. And we hear these crazy numbers in 0:13:45.040 --> 0:13:48.400 the US about the hyperscalers spending hundreds of billions of 0:13:48.440 --> 0:13:52.240 dollars this year alone. But when you're looking at the 0:13:52.280 --> 0:13:54.240 different parts of the business. So you have a very 0:13:54.280 --> 0:13:57.960 mature internet search business, and then you have everything that 0:13:58.000 --> 0:14:01.760 you're doing with AI, including in structure, how are you 0:14:01.800 --> 0:14:05.360 actually allocating capital and then how are you actually I guess, 0:14:05.400 --> 0:14:09.839 balancing that with returning capital at the same time to shareholders. 0:14:10.600 --> 0:14:13.640 So I called it impossible triangle. So I kind of 0:14:13.880 --> 0:14:16.839 scratched my head every few months, every few quick weeks, 0:14:16.880 --> 0:14:19.960 depends on I also see the headline numbers brooke news 0:14:19.960 --> 0:14:22.400 with other peers as well, So sometimes I make a 0:14:22.440 --> 0:14:25.040 little bit kind of hesitating to make a statement, but 0:14:25.240 --> 0:14:27.400 I just want to tell the facts and tell the views. 0:14:27.440 --> 0:14:29.760 I want to separate them out. So in the recent 0:14:29.800 --> 0:14:32.400 quarter earnings, we you know, we mentioned we actually solve 0:14:32.520 --> 0:14:37.280 partially on these impossible triangles. One is our operating profit 0:14:37.520 --> 0:14:40.640 increase almost doubled on a Q on Q base and 0:14:40.720 --> 0:14:43.640 number two, our cloud revenue grew about seventy nine percent 0:14:43.680 --> 0:14:46.160 on a WILDWI basis, which is almost double off the 0:14:46.200 --> 0:14:49.160 wildwide growth rate for the cloud market in China. And 0:14:49.240 --> 0:14:52.680 number three is since Q three last year, our operating 0:14:52.760 --> 0:14:56.040 cash flow has turned positive. So positive operating cash flow, 0:14:56.280 --> 0:15:00.800 incremental operating profits and the higher growth and the market. However, 0:15:00.960 --> 0:15:03.720 my capbacks is not seeing kind of double even third 0:15:04.200 --> 0:15:07.080 multiple times. So I think the way of resolving that 0:15:07.280 --> 0:15:09.760 is as a CFO or as a management team of 0:15:09.800 --> 0:15:12.760 a heavy CABACS investor AI tech company right now need 0:15:12.800 --> 0:15:15.560 to find a way on one hand really drive the 0:15:15.600 --> 0:15:18.880 growth but also keep the density of the investment into 0:15:18.920 --> 0:15:22.560 air in a reasonable pacing. However, when you do that, 0:15:22.680 --> 0:15:25.360 you need to keep a conscious regarding OURI and our 0:15:25.560 --> 0:15:28.800 in mind to look at the entire cash cycle. For example, 0:15:28.840 --> 0:15:30.880 every dollar we spend today we probably need to wait 0:15:30.920 --> 0:15:33.480 for another probably twenty thirty forty months depends on the 0:15:33.520 --> 0:15:37.200 category to get full cash back. And during that frame, 0:15:38.040 --> 0:15:41.960 obviously there's a you know, price heights, memories, there's difficulty 0:15:41.960 --> 0:15:45.080 on IDC centers and a huge spending our service. So 0:15:45.120 --> 0:15:47.520 my point is as a SFO on every project, you 0:15:47.560 --> 0:15:49.520 need to look at the entire life cycle, not only 0:15:49.520 --> 0:15:52.080 at one time, but also the pacing important because the 0:15:52.120 --> 0:15:54.640 foundation model R and D always taking a few months. Right, 0:15:54.680 --> 0:15:56.440 So these are the things you need to keep our mind. 0:15:56.520 --> 0:15:59.400 But my statement today is as by do we want 0:15:59.480 --> 0:16:04.040 to invest probably in a more responsible manner to the shareholders, 0:16:04.440 --> 0:16:09.360 but do not diminish our ambitious to investment into AI. 0:16:09.600 --> 0:16:12.440 Keep the right density is important. But given the results 0:16:12.440 --> 0:16:14.320 for this quarter, I think we kind of resolve that 0:16:14.400 --> 0:16:16.480 at least for this quarter. So hopefully you can keep 0:16:16.520 --> 0:16:18.400 on working on that and maybe, you know, a half 0:16:18.480 --> 0:16:20.640 year later when we check on this point, we can 0:16:20.680 --> 0:16:22.600 still keep on the same pattern, you know, high growths, 0:16:22.920 --> 0:16:26.080 less dollars spend, but better II. I think that's probably 0:16:26.120 --> 0:16:27.120 the angle we want to achieve. 0:16:27.160 --> 0:16:29.560 Can I just mention something I'm very curious about? You know, 0:16:29.560 --> 0:16:34.480 I'd say the heads of the American AI labs maybe 0:16:34.640 --> 0:16:39.080 have varying degrees of AI psychosis. They have a lot 0:16:39.120 --> 0:16:43.520 of worries that the what they call alignment research, et cetera. 0:16:43.720 --> 0:16:46.480 Do you work on similar things or do you have 0:16:46.520 --> 0:16:48.520 the same concerns and do. 0:16:48.480 --> 0:16:51.320 You also have AI side, Like do you like you know, 0:16:51.440 --> 0:16:55.920 for speaking of like trying to make money, Like do 0:16:56.000 --> 0:16:58.720 you invest in or how much do you invest in 0:16:58.960 --> 0:17:02.240 what they would call AI safety or alignment and essentially 0:17:02.720 --> 0:17:06.159 making sure that the models that you're building don't go 0:17:06.359 --> 0:17:09.119 rogue and always work on behalf of human flourishing? 0:17:09.400 --> 0:17:11.399 Is that a thing that you allocate capital to? 0:17:11.600 --> 0:17:15.040 Yeah, that's great question. So there's an emerging area of 0:17:15.200 --> 0:17:18.719 example in this data sanity and all the kind of 0:17:18.920 --> 0:17:21.399 post training efforts need to work on that. You know, 0:17:21.440 --> 0:17:24.399 alignment obviously is one of that. But my point is 0:17:25.280 --> 0:17:27.840 if you look at this new concept of harness, right, 0:17:27.880 --> 0:17:31.160 it's not only about training and getting model on the leaderboard, 0:17:31.200 --> 0:17:34.600 but also more importantly to measure the robustness and all 0:17:34.640 --> 0:17:38.119 the things you mentioned. I think in the context in 0:17:38.240 --> 0:17:43.360 China tech sector, the engineering has to be and has 0:17:43.480 --> 0:17:47.600 been a good competitive advantage. So the harness from the 0:17:47.720 --> 0:17:50.879 data flightwell to the alignment, to the data quality and 0:17:50.920 --> 0:17:54.640 the labeling. I think the entire evil system has been robust. 0:17:54.640 --> 0:17:56.560 For if you think about even in the mobile internet 0:17:56.600 --> 0:17:59.679 work right, so as simple as data labeling to the 0:17:59.680 --> 0:18:02.560 alignment tracks and post training and SFT, I think this 0:18:02.840 --> 0:18:05.080 kind of the full chain of the capability in terms 0:18:05.080 --> 0:18:08.520 of the talents and the pool of resources and the 0:18:08.560 --> 0:18:11.840 cost of data sanity and all the tracks has been 0:18:12.040 --> 0:18:15.000 in my view a little bit kind of more efficient 0:18:15.200 --> 0:18:18.120 in a way that the ecosystem has been in place there, 0:18:18.280 --> 0:18:20.960 so the cost efficiency has been there. So my view 0:18:21.000 --> 0:18:26.760 is this is engineering, not theoretical quantum y right, So 0:18:26.800 --> 0:18:30.320 on that the engineering capability form. The China tech world 0:18:30.400 --> 0:18:33.600 and industry has been there with the key elements I 0:18:33.680 --> 0:18:37.320 mentioned right, talents, lower costs, more efficient I think these 0:18:37.320 --> 0:18:38.960 are the few things I just want to point out 0:18:39.040 --> 0:18:41.879 actually can help solve the issue. But as I mentioned it, 0:18:42.000 --> 0:18:44.159 the things are evolved very quickly, right, so you know, 0:18:44.200 --> 0:18:47.359 we don't worry too much about the issue you mentioned 0:18:47.400 --> 0:18:48.080 in local market. 0:18:48.240 --> 0:18:50.359 Yeah, so this is interesting. I'm curious. I want to 0:18:50.640 --> 0:18:53.880 press further on this because the American air I'm very 0:18:53.920 --> 0:18:57.960 anxious about this, and they publish these model reports and 0:18:58.160 --> 0:19:01.080 it says things that in the chain of thought, we 0:19:01.080 --> 0:19:03.720 were able to see that four percent of the time 0:19:04.119 --> 0:19:07.160 the model was able to identify that it was being tested, 0:19:07.640 --> 0:19:11.600 and therefore it changed its behavior in response to recognizing 0:19:11.640 --> 0:19:14.040 that it was tested. And this is a sign of 0:19:14.080 --> 0:19:18.000 potential misalignment. Are you doing the same sort of research 0:19:18.080 --> 0:19:23.000 and spending to establish that again, the models work for 0:19:23.119 --> 0:19:25.640 people and don't have a rogue goal, So. 0:19:26.080 --> 0:19:28.800 Yeah, sure, I think right now if you look at this, right. 0:19:28.920 --> 0:19:31.359 So, we are also part of the open source community, 0:19:31.520 --> 0:19:35.400 so many of the good model especially publishing recently, also 0:19:35.440 --> 0:19:38.320 will publish their salts. Makes sense, so we kind of 0:19:38.440 --> 0:19:41.359 follow the new sauts, but also doing our own pasts 0:19:41.400 --> 0:19:43.840 as well. So overall, I think people in the open 0:19:43.920 --> 0:19:46.439 source community today, in my view, is very collegial. So 0:19:46.480 --> 0:19:48.240 people still want to do a better model from kil 0:19:48.280 --> 0:19:50.639 model for everyone globally, not really on one country to 0:19:51.040 --> 0:19:52.160 different places yet. 0:19:52.760 --> 0:19:55.200 So actually, related to this, I'm going to ask something. 0:19:55.280 --> 0:19:58.120 Maybe it's slightly sensitive, but I think it's very important. 0:19:58.200 --> 0:20:01.359 So in the US, the AI companies, even as they 0:20:01.400 --> 0:20:04.920 talk about safety, they're basically self regulating, right Like they 0:20:05.080 --> 0:20:08.040 choose to put out these reports and judge their own 0:20:08.080 --> 0:20:11.040 models and things like that. In China, tell me if 0:20:11.040 --> 0:20:14.399 I'm wrong, but it feels very different. It feels like 0:20:14.440 --> 0:20:16.920 the government is more hands on when it comes to AI. 0:20:17.040 --> 0:20:19.639 China has been very explicit about this as an area 0:20:20.440 --> 0:20:24.880 national security, national strategy. So you're operating in an environment 0:20:24.920 --> 0:20:30.080 where you're firmly embedded in China's technological and industrial policy. 0:20:30.680 --> 0:20:34.360 How does that influence the development of your AI models 0:20:34.359 --> 0:20:35.520 and your broader tech. 0:20:35.720 --> 0:20:38.600 So obviously we're not inflation to commut of public policy, 0:20:38.640 --> 0:20:40.600 but I definitely happy to share some of our thoughts 0:20:40.640 --> 0:20:43.520 I have. I think in the world, in the China 0:20:43.560 --> 0:20:46.800 AI today, we believe we have a great group of 0:20:47.359 --> 0:20:51.280