00:00We are standing in front of the Kiwi K3 booth, it's the talk of the town here.
00:04How big is this going to be a shake-up for global frontier AI models?
00:08Thanks for having me, Miming.
00:10I think Kiwi's new model is definitely a surprise to both Wall Street and Silicon Valley.
00:18The biggest takeaway is now we have the debate heating up again
00:23over open-source models versus closed-source models.
00:26If Kiwi can produce at a much lower cost, and it's an open-weight model versus cloud and open AI,
00:34and then the question will be naturally, do we still need to stick to the closed-source model
00:39as all the US AI companies try to keep on?
00:44And the rest of the world will also look at China as like,
00:47okay, so now we have a more affordable solution.
00:49It's like a perfect plan B.
00:51So that's a clear challenger to the dominance of US AI models.
00:55I think one concern right now is that given that this is such a huge model, right,
00:592.8 trillion parameters, that means that running it requires a huge amount of accelerators.
01:05So given the shortage of GPUs, because of the export controls,
01:09is this going to limit the deployment of the Kiwi K3 model?
01:12You are really on the spot.
01:13I think, you know, this is also a question that will raise a lot of eyebrows in Washington, D.C.
01:19You know, after all the efforts, you know, to restrict like H200 and other chips for export,
01:24and, you know, Kimi now gives the result, you know, even without the H200,
01:28I think that they still use some H100 chips, plus some domestic chips,
01:32and they can still have the K3 models, you know, ready.
01:35And you just see Alibaba also last night announced their new model.
01:39I don't think they have, like, access to a large amount of H200, no matter what.
01:43So, you know, the question is whether the export control is really effective,
01:46and I think this will be also a key issue for the US-China AI dialogue.
01:50You know, both sides will, you know, rethink about the strategy.
01:52Is there a concern that because of the lack of those GPUs,
01:57even though they have come up with a new model now,
02:00the processing and the running and the usage rates could be constrained by how much compute they have?
02:06So far, it is.
02:08I think Alibaba was also very humble to say that new model is second to, you know,
02:13Cloud's, you know, Miso and Fable.
02:16So, you can see, you know, the Chinese companies know where the limit are,
02:21but this is just for now, right?
02:22If it's like five more years, you know, the story could be very different.
02:25And I was also very curious, you know, I saw Huawei also released the new AI class.
02:32So the logic is pretty simple.
02:33I can't beat you on the individual chips,
02:36but if I can have a better integrated solution,
02:39you know, that can also match the capacity of H200.
02:43So, we saw the tech route in Asia and the US since Friday.
02:47Do you think that is justified or is it an overreaction at this point?
02:51No, unfortunately, I think that the AI race, you know, between the U.S. and China, you know, now like
02:57a goal really to the next level.
02:58So, you know, we heard the President Xi's announcement of Wyco, the new international organization on AI.
03:05In my view, it's very much an answer to the Paxilica, you know, led by the U.S.
03:10So, my point is, you know, the competition between the U.S. and China on AI is not just between
03:15the two countries.
03:15Now, it's really about, you know, where to grab your partners and allies, you know, the competition will be driven
03:21by U.S. and China,
03:22but it's already on the water stage.
03:24And when you look at the two bifurcation in the tech ecosystem right now,
03:29you have the U.S. making leaps and bounds in the hardware side,
03:32and China making progress in the software side,
03:35so where do you think, you know,
03:37China and U.S. would stand in this race, you know, in the next three to five years?
03:40Right. I'm really glad you mentioned the software part.
03:44You know, we talk about those frontier models most of the time,
03:47but there are also very interesting, for example, like AI video generators, right?
03:54Chinese technology actually already dominating that space.
03:56You know, everybody knows, you know, if you do like short videos, production studios,
04:00they already know C-Dance, which is created by ByDance.
04:04So TikTok is already sort of a headache for the U.S.
04:09And now, you know, the U.S. has to deal with, like, you know, what do you do with, like,
04:13C-Dance.
04:14They are already, like, the most popular, you know, AI video generator.
04:17So I think that China will go, you know, try to go stronger and stronger in the software part.
04:22They know the hardware, the chips will take time to mature, but this is still on their radar.
04:27The Chinese government did approve a small amount of H200 chips import.
04:32Right.
04:32How do you read into this?
04:33Is it a reflection of the shortage we're seeing here?
04:36The shortage is for sure.
04:38How serious is it?
04:40So I understand, you know, based on my conversation with Chinese tech companies,
04:45they can deal with the situation.
04:48It's not ideal, especially for Alibaba, you know, ByDance,
04:51when they try to develop the next frontier models, you still need advanced chips.
04:56But for most, you know, regular R&D, the Huawei chips, you know, can do the work, you know, so
05:03far.
05:03I think that the, you know, the small quarter, you know, that will be allowed, you know, to be imported
05:09to China.
05:10This is a very much, you know, diplomatic, you know, gesture.
05:13Given that many of these Chinese AI models are being sold at just a fraction of U.S. pricing,
05:18how sustainable is this for the companies? Are they making losses and subsidizing consumers in the process?
05:25In a way, it is. You know, the Chinese government supplies is obvious, right?
05:29You know, we talked about chips, but you also saw the robots, right?
05:32You know, this year, you know, the WRC has, like, the entire floor door for robots.
05:37And then you go around, every robot company is like Beijing, Wuhan, Shanghai.
05:41So, like, you know, clearly government-backed.
05:44The chip industry is the same.
05:46I think, you know, from the top level, the message is at all costs, no matter what.
05:50In the five years, you have to, you know, develop a China-grown, China's homegrown chips that can match H200.
05:58So, everybody put the money in, including the private sector, you know, there will be some winners and losers.
06:05And given the latest progress over this weekend with the new AI models,
06:09do you foresee more actions being taken by the U.S. government to place curbs on any tech transfers to
06:14China?
06:15I believe so. I think the U.S. will, you know, as we just said,
06:19I think Washington will get more concerned, you know, to have a reflection on the export control policy.
06:25It will probably make Jensen-Huan more difficult to convince, you know, the commerce and the Congress
06:32to continue to explore more, you know, advanced models to China.
06:35Even, you know, Kimi, with the very limited supply of, you know, U.S. chips, they already got to K3,
06:41right?
06:41Can you imagine if you have, like, full access to H200, you know, give it, like, two years, you know,
06:46the story could be different. On the business side, I think Wall Street would also be a little bit concerned.
06:52Whether models like K3 would grab some customers, you know, which originally rely on U.S. models.
06:59But there's already a trend. A lot of companies began to use the hybrid model.
07:04They use the Chinese model for some basic research.
07:07They use the U.S. model for some of the monster research.
07:09This is not good news for companies like Azaropec and OpenAI because, you know,
07:12they consider they are the king of the world.
07:14But now the position is being challenged.
Comments