WEBVTT - The Architect of China's AI Breakthrough

0:00:02.720 --> 0:00:07.200
<v Speaker 1>Bloomberg Audio Studios, podcasts, radio news.

0:00:08.600 --> 0:00:11.640
<v Speaker 2>Earlier this year, a new product from the Chinese AI

0:00:11.720 --> 0:00:15.400
<v Speaker 2>startup Deep Seek, shocked the world and rattled Wall Street.

0:00:16.079 --> 0:00:20.040
<v Speaker 2>China's deep Seek is freaking out the AI world right now.

0:00:20.079 --> 0:00:23.279
<v Speaker 2>Techtoks tumbled as It's absurd to the top of the

0:00:23.360 --> 0:00:27.920
<v Speaker 2>download charts. But despite the global attention, very little is

0:00:28.000 --> 0:00:32.160
<v Speaker 2>known about the man behind deep Seek. Chinese entrepreneur Liang Wu.

0:00:33.120 --> 0:00:35.839
<v Speaker 1>Liang Wenfeng is certainly a mystery figure.

0:00:36.400 --> 0:00:40.080
<v Speaker 2>Bloomberg crit the rye covers artificial intelligence in Asia.

0:00:40.680 --> 0:00:44.240
<v Speaker 1>He's certainly one of the most inaccessible and low key

0:00:44.840 --> 0:00:48.319
<v Speaker 1>tech entrepreneurs that I've come across. Just to illustrate how

0:00:48.360 --> 0:00:52.920
<v Speaker 1>private he is, we weren't able to find any pictures

0:00:52.960 --> 0:00:56.600
<v Speaker 1>of him on the internet when we scoured through his

0:00:56.760 --> 0:01:00.520
<v Speaker 1>website and all of that. But finally appear in a

0:01:00.600 --> 0:01:05.520
<v Speaker 1>really high profile meeting with President Sheijeng Peing, and that

0:01:05.560 --> 0:01:08.440
<v Speaker 1>picture got out into the world and he was everywhere.

0:01:09.240 --> 0:01:11.120
<v Speaker 2>What does this man of mystery look like?

0:01:11.920 --> 0:01:18.440
<v Speaker 1>He is slim, wears glasses, but doesn't talk much. Baby faced, Yes,

0:01:18.680 --> 0:01:20.319
<v Speaker 1>I think we could describe him like that.

0:01:25.240 --> 0:01:29.119
<v Speaker 2>Deep Seek rarely answers questions about Leang citing his privacy.

0:01:29.640 --> 0:01:32.399
<v Speaker 2>But Sarita and her colleagues were curious about this man

0:01:32.560 --> 0:01:35.480
<v Speaker 2>whose AI systems turned the tech world on its head.

0:01:36.040 --> 0:01:38.839
<v Speaker 2>So they spoke with dozens of people familiar with his work,

0:01:39.040 --> 0:01:43.040
<v Speaker 2>from former employees and fellow researchers to investors and insiders

0:01:43.040 --> 0:01:43.760
<v Speaker 2>in the industry.

0:01:44.360 --> 0:01:47.760
<v Speaker 1>And what we found is that, yes, he is extraordinarily

0:01:47.840 --> 0:01:54.880
<v Speaker 1>low key, very shy, but extraordinarily driven and talented and passionate.

0:01:55.240 --> 0:01:58.640
<v Speaker 1>And I think he has somewhat taken on Deep Seek

0:01:58.680 --> 0:02:03.000
<v Speaker 1>as a sort of a mission to establish China in AI,

0:02:03.560 --> 0:02:07.480
<v Speaker 1>trying to make sure that China is the force to

0:02:07.560 --> 0:02:09.320
<v Speaker 1>reckon with in AI.

0:02:12.680 --> 0:02:15.920
<v Speaker 2>Welcome to the Big Take Asia from Bloomberg News. I'm wanha.

0:02:16.600 --> 0:02:18.960
<v Speaker 2>Every week we take you inside some of the world's

0:02:18.960 --> 0:02:22.920
<v Speaker 2>biggest and most powerful economies and the markets, tycoons and

0:02:23.040 --> 0:02:27.520
<v Speaker 2>businesses that drive this ever shifting region. Today, on the show,

0:02:27.680 --> 0:02:31.079
<v Speaker 2>who is LNG One Fun, we learn about the mysterious

0:02:31.120 --> 0:02:34.040
<v Speaker 2>tech founder who led Deep Seek to the frontline of

0:02:34.120 --> 0:02:38.680
<v Speaker 2>AI advances. Plus what is the company's rapid rise? Tell

0:02:38.760 --> 0:02:45.840
<v Speaker 2>us about the US China artificial intelligence race. Sriita, thanks

0:02:45.840 --> 0:02:49.240
<v Speaker 2>for joining us. I'm fascinated by AI. You guys did

0:02:49.280 --> 0:02:52.040
<v Speaker 2>such an interesting job with a story. I wonder if

0:02:52.040 --> 0:02:54.800
<v Speaker 2>we can start with who is Lang? One fun? What

0:02:54.840 --> 0:02:56.320
<v Speaker 2>do we know about his roots?

0:02:56.800 --> 0:03:01.040
<v Speaker 1>So? Liang is about forty years old in a small

0:03:01.120 --> 0:03:06.760
<v Speaker 1>village called Middling in the Guangdong province. His parents were

0:03:07.000 --> 0:03:12.960
<v Speaker 1>school teachers, mainly taught primary school kids. He was extremely

0:03:13.160 --> 0:03:18.280
<v Speaker 1>bright and went on to study at Xijiang University and

0:03:18.360 --> 0:03:20.519
<v Speaker 1>then also did his master's there.

0:03:21.040 --> 0:03:24.800
<v Speaker 2>At Jujung University. Liang and his friends immersed themselves in

0:03:24.919 --> 0:03:30.600
<v Speaker 2>all things tech, machine learning, signal processing, electronic engineering. They

0:03:30.639 --> 0:03:34.680
<v Speaker 2>even developed programs to trade stocks during the financial crisis.

0:03:35.280 --> 0:03:39.160
<v Speaker 2>After graduating, Liang joined forces with two of his classmates

0:03:39.360 --> 0:03:43.080
<v Speaker 2>and set up a quantitative hedge fund called High Fire Management.

0:03:44.160 --> 0:03:48.360
<v Speaker 1>So quant funds basically work with mathematical models and statistical

0:03:48.400 --> 0:03:53.120
<v Speaker 1>analysis to do stock trading. Humans are not involved in

0:03:53.160 --> 0:03:57.680
<v Speaker 1>taking the decisions. At its peak, High Flyer Management was

0:03:57.760 --> 0:04:02.240
<v Speaker 1>managing something like fourteen billion in assets, so it was

0:04:02.480 --> 0:04:07.000
<v Speaker 1>quite a sizeable fund, and in its most successful runs,

0:04:07.080 --> 0:04:12.480
<v Speaker 1>it was providing annualized returns averaging thirty five percent to

0:04:12.560 --> 0:04:15.040
<v Speaker 1>its investors, so I would say that it was doing

0:04:15.120 --> 0:04:15.520
<v Speaker 1>very well.

0:04:15.560 --> 0:04:20.360
<v Speaker 2>Indeed, according to former employees, high Flyer had a geeky

0:04:20.400 --> 0:04:24.480
<v Speaker 2>startup culture. It's early job postings moosted of attracting top

0:04:24.560 --> 0:04:27.440
<v Speaker 2>talent from Google and Facebook and said they were looking

0:04:27.480 --> 0:04:30.760
<v Speaker 2>for math and coding geeks with quirky brilliants.

0:04:32.120 --> 0:04:37.359
<v Speaker 1>The early job postings also referred to Sheldon, who is

0:04:37.560 --> 0:04:43.919
<v Speaker 1>this very awkward main character in the prominent American sitcom

0:04:44.000 --> 0:04:45.520
<v Speaker 1>called The Big Bang Pory.

0:04:46.240 --> 0:04:48.960
<v Speaker 2>For example, I cry because others are stupid and it

0:04:48.960 --> 0:04:49.599
<v Speaker 2>makes me sad.

0:04:51.480 --> 0:04:56.680
<v Speaker 1>Sheldon has a legend of fans and it's extraordinarily funny

0:04:56.680 --> 0:05:00.520
<v Speaker 1>without meaning to be. So, you know, the whole culture

0:05:00.600 --> 0:05:04.360
<v Speaker 1>of deep Seek in the early days revolved around recreating

0:05:04.400 --> 0:05:08.520
<v Speaker 1>some of that geeky nerdy culture. They were free snacks,

0:05:08.760 --> 0:05:12.880
<v Speaker 1>poker game nights. Everybody was dressed in T shirts and slippers.

0:05:13.120 --> 0:05:14.920
<v Speaker 2>Sounds like a great place to work. Yeah.

0:05:14.960 --> 0:05:19.760
<v Speaker 1>It was really a very unorthodox startup culture, unlike what

0:05:19.839 --> 0:05:22.839
<v Speaker 1>you'll probably see in the big tech companies in China

0:05:22.920 --> 0:05:24.560
<v Speaker 1>such as the Aliba Bay and the Tencent.

0:05:25.080 --> 0:05:28.479
<v Speaker 2>And how did Liang transition from doing quant financials into

0:05:28.600 --> 0:05:30.120
<v Speaker 2>AI and building deep seak.

0:05:30.800 --> 0:05:37.720
<v Speaker 1>Liang was always extraordinarily interested in machine learning and artificial intelligence,

0:05:38.360 --> 0:05:41.640
<v Speaker 1>and then a few months after open AI. You know

0:05:41.760 --> 0:05:47.640
<v Speaker 1>launched chat GPT, the chatbot that became an overnight global success.

0:05:47.960 --> 0:05:51.000
<v Speaker 1>It was then the spring of twenty twenty three, a

0:05:51.040 --> 0:05:54.120
<v Speaker 1>few months had passed after the launch of chat GPT,

0:05:55.080 --> 0:06:00.839
<v Speaker 1>and Liang then announced that deep Seek would be set up.

0:06:01.480 --> 0:06:06.400
<v Speaker 1>In its early manifesto, deep Seek talked about shunning mediocrity

0:06:07.080 --> 0:06:12.000
<v Speaker 1>and tackling the big challenges in AI, and of course,

0:06:12.120 --> 0:06:15.640
<v Speaker 1>ultimately cracking artificial general intelligence.

0:06:16.400 --> 0:06:20.440
<v Speaker 2>The manifesto also laid out deep Seek's ambition to position

0:06:20.640 --> 0:06:23.440
<v Speaker 2>China as a leader in cutting edge technologies.

0:06:24.320 --> 0:06:27.960
<v Speaker 1>You know, Liang has given two interviews, rare as they

0:06:28.000 --> 0:06:32.080
<v Speaker 1>may be. In both those interviews, he's talked about bringing

0:06:32.200 --> 0:06:36.760
<v Speaker 1>China's AI ecosystem to the forefront of where the world is.

0:06:37.160 --> 0:06:40.279
<v Speaker 1>You know, China has been accused constantly of being a copycat.

0:06:40.560 --> 0:06:43.719
<v Speaker 1>He wanted in AI China to chart a different path.

0:06:44.720 --> 0:06:48.719
<v Speaker 2>Deep Seek worked fast since twenty twenty three, it's released

0:06:48.720 --> 0:06:51.919
<v Speaker 2>over half a dozen AI models and helped pioneer a

0:06:51.960 --> 0:06:55.880
<v Speaker 2>technique called sparsity, which enabled those models to train and

0:06:55.960 --> 0:07:01.400
<v Speaker 2>run more efficiently. Developers started to take note then earlier

0:07:01.480 --> 0:07:03.000
<v Speaker 2>this year. Back to that top.

0:07:02.839 --> 0:07:06.480
<v Speaker 1>Story, now deep Seek shaking up global tech lower technology

0:07:06.480 --> 0:07:09.680
<v Speaker 1>co host when they released their reasoning model R one

0:07:10.280 --> 0:07:14.560
<v Speaker 1>that cost such an upheaval in the industry and coused

0:07:15.000 --> 0:07:18.880
<v Speaker 1>a trillion dollar stock market meltdown. That's when the world

0:07:19.200 --> 0:07:24.880
<v Speaker 1>really started paying attention to this secretive AI entrepreneur in.

0:07:24.800 --> 0:07:33.040
<v Speaker 2>China and Trita. What is show groundbreaking about deep seeks

0:07:33.120 --> 0:07:33.880
<v Speaker 2>are one model.

0:07:34.520 --> 0:07:39.240
<v Speaker 1>The AI industry until recently was always about billions of

0:07:39.280 --> 0:07:43.440
<v Speaker 1>dollars spent in building the infrastructure, the data centers, the

0:07:43.920 --> 0:07:47.640
<v Speaker 1>graphics processing units in the data centers that would train

0:07:47.720 --> 0:07:52.480
<v Speaker 1>these models. But what Deepseek did was show that its

0:07:52.600 --> 0:07:56.880
<v Speaker 1>models could match or even outdoor on some benchmark measures

0:07:57.480 --> 0:08:02.400
<v Speaker 1>what the latest open AI or anthropic models were doing,

0:08:03.000 --> 0:08:08.440
<v Speaker 1>and with far less computational power, with far less resources,

0:08:08.480 --> 0:08:14.080
<v Speaker 1>and as deep Sea claimed, with far less capital as well.

0:08:14.360 --> 0:08:16.960
<v Speaker 2>So how did Liang and his team manage to achieve

0:08:17.120 --> 0:08:20.080
<v Speaker 2>true innovation at what it says is a fraction of

0:08:20.120 --> 0:08:23.320
<v Speaker 2>the cost? And what does Deep seek success say about

0:08:23.320 --> 0:08:27.480
<v Speaker 2>the AI race between China and the US. That's after

0:08:27.520 --> 0:08:41.040
<v Speaker 2>the break. For much of the last decade, the US

0:08:41.080 --> 0:08:45.920
<v Speaker 2>has tried to restrict China's access to semiconductors. Tensions reached

0:08:45.920 --> 0:08:48.920
<v Speaker 2>a fever pitch in twenty twenty two. In the following year,

0:08:49.160 --> 0:08:53.200
<v Speaker 2>when Washington targeted Beijing with two rounds of chip export

0:08:53.240 --> 0:08:55.640
<v Speaker 2>controls in Vidia and shares.

0:08:55.679 --> 0:08:59.520
<v Speaker 1>Some semiconductor companies have slumped today after the Biden administration

0:08:59.679 --> 0:09:03.400
<v Speaker 1>said it would tighten restrictions on exports of AI chips

0:09:03.400 --> 0:09:04.800
<v Speaker 1>to China.

0:09:05.360 --> 0:09:08.920
<v Speaker 2>That limited sales from American firms like Nvidia, whose cutting

0:09:09.000 --> 0:09:11.800
<v Speaker 2>edge chips are used by tech companies to help train

0:09:11.960 --> 0:09:15.960
<v Speaker 2>their AI models. The move presented a significant challenge for

0:09:16.040 --> 0:09:19.680
<v Speaker 2>developers in China, but as Bloomberg's read the rises, it

0:09:19.840 --> 0:09:22.559
<v Speaker 2>also forced them to develop workarounds.

0:09:23.120 --> 0:09:27.160
<v Speaker 1>Necessity is always the mother of innovation. This has been

0:09:27.320 --> 0:09:33.680
<v Speaker 1>proven by AI developers in China. Never mind the export cubs.

0:09:34.120 --> 0:09:37.200
<v Speaker 1>They've still gone on to build good models that have

0:09:37.400 --> 0:09:39.840
<v Speaker 1>benchmarked with the best around the world.

0:09:40.559 --> 0:09:43.360
<v Speaker 2>And one of the most innovative approaches from Deepseek is

0:09:43.360 --> 0:09:45.439
<v Speaker 2>a sparsity technique we mentioned earlier.

0:09:47.120 --> 0:09:51.120
<v Speaker 1>Now. Sparsity is something to do with building a model

0:09:51.160 --> 0:09:55.400
<v Speaker 1>without having the high end computational power. It's when a

0:09:55.559 --> 0:09:59.120
<v Speaker 1>large language model doesn't have to be entirely harnessed to

0:09:59.160 --> 0:10:02.960
<v Speaker 1>give a nant to a query. Instead, Liang and his

0:10:03.400 --> 0:10:10.000
<v Speaker 1>fellow developers tried to apportion the expertise of the model

0:10:10.080 --> 0:10:14.120
<v Speaker 1>into smaller expert groups and then only harness those groups

0:10:14.160 --> 0:10:17.840
<v Speaker 1>that required to be used. So in doing that, they

0:10:18.240 --> 0:10:21.920
<v Speaker 1>made it much more computation efficient and also much more

0:10:21.960 --> 0:10:22.760
<v Speaker 1>cost efficient.

0:10:23.679 --> 0:10:26.679
<v Speaker 2>Is it basically, instead of using your whole brain, are

0:10:26.720 --> 0:10:30.240
<v Speaker 2>you using just certain parts of your brain to do

0:10:30.400 --> 0:10:31.679
<v Speaker 2>that computation.

0:10:31.880 --> 0:10:35.560
<v Speaker 1>That's exactly right on? You know, instead of entirely using

0:10:35.640 --> 0:10:39.400
<v Speaker 1>every little gray cell in your brain, it only fires

0:10:39.520 --> 0:10:42.880
<v Speaker 1>up those neurons or little portions in your brain that

0:10:42.960 --> 0:10:47.560
<v Speaker 1>contained that particular field of expertise and then bring that

0:10:47.880 --> 0:10:50.280
<v Speaker 1>to you know, respond to a query or give an

0:10:50.280 --> 0:10:53.599
<v Speaker 1>answer to a particular question, whether it's a command or

0:10:53.640 --> 0:10:54.520
<v Speaker 1>a coding question.

0:10:56.720 --> 0:11:00.560
<v Speaker 2>The sparsity breaks through impressed Deep seeks competitors, but its

0:11:00.640 --> 0:11:05.040
<v Speaker 2>price point is what ultimately made headlines. Deep Seek said

0:11:05.040 --> 0:11:07.760
<v Speaker 2>it cost them just five point six million dollars to

0:11:07.840 --> 0:11:11.360
<v Speaker 2>train its V three model. That's far less than they

0:11:11.440 --> 0:11:14.800
<v Speaker 2>estimated one hundred million dollars Open Ai spent on its

0:11:14.800 --> 0:11:17.120
<v Speaker 2>most advanced version of chat GPT.

0:11:18.000 --> 0:11:21.439
<v Speaker 1>Now, there is definitely a whole lot of skepticism around

0:11:21.880 --> 0:11:27.400
<v Speaker 1>that number because just the intrastructure, the training of the model,

0:11:27.520 --> 0:11:31.280
<v Speaker 1>the talent, and the time it takes all of it

0:11:31.280 --> 0:11:34.040
<v Speaker 1>adds to quite a sizable sum of money, so the

0:11:34.160 --> 0:11:39.240
<v Speaker 1>skepticism is warranted. People have estimated that there was no

0:11:39.360 --> 0:11:41.679
<v Speaker 1>way Deep Seak could have pulled that off without at

0:11:41.760 --> 0:11:43.280
<v Speaker 1>least a billion dollars or more.

0:11:44.679 --> 0:11:47.760
<v Speaker 2>Also in deep Seak's favor is that AI startups like

0:11:47.800 --> 0:11:51.160
<v Speaker 2>it have a staunch ally in China's government and present.

0:11:51.200 --> 0:11:56.160
<v Speaker 2>She Jinkin Sriita says she sees generative AI, robotics and

0:11:56.160 --> 0:11:59.640
<v Speaker 2>other high tech ambitions as beneficial to the state's agenda,

0:12:00.120 --> 0:12:03.760
<v Speaker 2>part of a larger push for self reliance in key technologies,

0:12:04.360 --> 0:12:07.480
<v Speaker 2>and Deep seax success has spurred much bigger rivals such

0:12:07.480 --> 0:12:10.720
<v Speaker 2>as Ali, Baba, ten Cent, and byte Edance to release

0:12:10.760 --> 0:12:15.800
<v Speaker 2>their own AI models. Srida deep six model is entirely

0:12:15.840 --> 0:12:18.880
<v Speaker 2>open sourced at this point. That means any individual or

0:12:18.960 --> 0:12:23.319
<v Speaker 2>company could incorporate deep Six's algorithms into their own programs.

0:12:24.040 --> 0:12:26.760
<v Speaker 2>Why did the company choose this approach and why is

0:12:26.760 --> 0:12:29.040
<v Speaker 2>that important open source?

0:12:29.440 --> 0:12:34.240
<v Speaker 1>On one level, you could say that it is democratizing

0:12:34.400 --> 0:12:37.960
<v Speaker 1>AI and taking it out into the world, But let's

0:12:38.040 --> 0:12:43.560
<v Speaker 1>not forget that China's AI models would have otherwise found

0:12:43.640 --> 0:12:47.800
<v Speaker 1>fewer takers around the world if they were proprietary models

0:12:47.840 --> 0:12:51.439
<v Speaker 1>and were on par in terms of what they cost

0:12:51.600 --> 0:12:55.400
<v Speaker 1>with Western companies such as open ai. By making it

0:12:55.480 --> 0:12:59.400
<v Speaker 1>cheap and by making it open source, China allowed people

0:12:59.440 --> 0:13:01.680
<v Speaker 1>around the world to quickly take a look at the

0:13:01.720 --> 0:13:07.400
<v Speaker 1>models and begin using them, allowing them to be adopted

0:13:07.480 --> 0:13:13.199
<v Speaker 1>much faster in the business and AI ecosystem, thereby outdoing

0:13:13.400 --> 0:13:16.839
<v Speaker 1>the likes of open Ai. Now that's huge. It's not

0:13:16.880 --> 0:13:21.960
<v Speaker 1>only about democratizing models, it's strategically about making sure that

0:13:22.000 --> 0:13:25.000
<v Speaker 1>you cut out your competitor by making things so cheap

0:13:25.240 --> 0:13:33.199
<v Speaker 1>that the world adopts it quickly and then it becomes mainstream.

0:13:33.360 --> 0:13:36.720
<v Speaker 2>As a result, Microsoft and Amazon both offer deep seek

0:13:36.760 --> 0:13:39.760
<v Speaker 2>on their cloud services, and deep seeks models have been

0:13:39.760 --> 0:13:43.600
<v Speaker 2>incorporated into Perplexity, an AI powered search engine that also

0:13:43.640 --> 0:13:46.160
<v Speaker 2>offers models from open Ai and Anthropic.

0:13:47.559 --> 0:13:50.920
<v Speaker 1>There is definitely a question about, you know, how fast

0:13:51.000 --> 0:13:54.400
<v Speaker 1>AI is advancing, and there's a fear around the world

0:13:54.640 --> 0:13:58.920
<v Speaker 1>about having all of the controls rest only with one

0:13:59.000 --> 0:14:01.120
<v Speaker 1>or two companies in the world. I think that was

0:14:01.160 --> 0:14:04.240
<v Speaker 1>what chep Seek and others were trying to put out

0:14:04.240 --> 0:14:06.800
<v Speaker 1>a message in the world, saying that all of the

0:14:06.800 --> 0:14:09.880
<v Speaker 1>controls cannot be left to one or two companies and

0:14:09.960 --> 0:14:13.400
<v Speaker 1>the proprietary models that they are building, it should be

0:14:13.480 --> 0:14:17.240
<v Speaker 1>much more democratic. Therefore, I think the open source philosophy

0:14:17.480 --> 0:14:23.480
<v Speaker 1>is about de risking concentration and allowing more people to

0:14:23.640 --> 0:14:27.960
<v Speaker 1>build with technologies that are much more available.

0:14:29.120 --> 0:14:31.880
<v Speaker 2>Is there potentially a kind of clash of cultures or

0:14:31.920 --> 0:14:34.040
<v Speaker 2>a clash of values as well when it comes to

0:14:34.080 --> 0:14:38.560
<v Speaker 2>building AI between the Wesk's approach and the Chinese approach.

0:14:38.800 --> 0:14:42.840
<v Speaker 1>Very clearly, because if you look at early models of

0:14:42.840 --> 0:14:47.360
<v Speaker 1>deep Seek or even the you know, not tweaked or

0:14:47.400 --> 0:14:50.280
<v Speaker 1>fine tune models of deep Seek, they are very much

0:14:50.440 --> 0:14:54.560
<v Speaker 1>working within the boundaries of China's censorship rules. For instance,

0:14:54.640 --> 0:14:59.240
<v Speaker 1>you cannot ask questions about Taiwan or Sheishiingping without it

0:14:59.320 --> 0:15:04.280
<v Speaker 1>giving a very bland official answer. Whereas if you take

0:15:04.320 --> 0:15:07.960
<v Speaker 1>that same model and you can train it with other

0:15:08.040 --> 0:15:14.120
<v Speaker 1>data and make it culturally suitable to different geographies. That's

0:15:14.160 --> 0:15:17.080
<v Speaker 1>one of the things that deep Seek learned early on

0:15:17.400 --> 0:15:23.520
<v Speaker 1>that by open sourcing the model and by giving developers

0:15:23.560 --> 0:15:28.400
<v Speaker 1>and users a chance to customize to their own cultural context,

0:15:29.080 --> 0:15:33.400
<v Speaker 1>deep Seak could find much more quicker and faster adoption

0:15:33.520 --> 0:15:36.280
<v Speaker 1>around the world than by controlling a lot of it

0:15:36.640 --> 0:15:39.840
<v Speaker 1>and controlling it in a way that it could only

0:15:39.920 --> 0:15:42.920
<v Speaker 1>give China friendly answers around the world.

0:15:45.320 --> 0:15:49.240
<v Speaker 2>And while some applied China's innovations in AI, many of

0:15:49.280 --> 0:15:53.600
<v Speaker 2>the US suspect darker reasons for the success. An April

0:15:53.680 --> 0:15:57.840
<v Speaker 2>report from a US House of Representatives committee alleged significant

0:15:57.880 --> 0:16:01.640
<v Speaker 2>ties between Deep Seek and the Chinese got It concluded

0:16:01.680 --> 0:16:05.680
<v Speaker 2>that the company unlawfully stole data from Open Ai. The

0:16:05.800 --> 0:16:10.520
<v Speaker 2>Chinese embassy rejects those claims as groundless. Meanwhile, Deep Seek

0:16:10.560 --> 0:16:15.000
<v Speaker 2>and Liang haven't commented on the House report. Srita, it

0:16:15.080 --> 0:16:17.360
<v Speaker 2>seems like there is very much of an arms race

0:16:17.360 --> 0:16:19.800
<v Speaker 2>of sorts when it comes to AI, certainly between the

0:16:19.920 --> 0:16:20.600
<v Speaker 2>US and China.

0:16:20.600 --> 0:16:25.360
<v Speaker 1>At the moment, it's very much a race, and I

0:16:25.400 --> 0:16:29.000
<v Speaker 1>think it would be too early to call a winner.

0:16:29.880 --> 0:16:33.040
<v Speaker 1>All I can say is that a year ago I

0:16:33.080 --> 0:16:36.760
<v Speaker 1>would not have called it as a close race. It's

0:16:36.760 --> 0:16:39.120
<v Speaker 1>a marathon, but you have to go at a sprinting pace.

0:16:39.640 --> 0:16:42.000
<v Speaker 1>We are really at the very beginning of it, and

0:16:42.360 --> 0:16:47.120
<v Speaker 1>for whichever country cracks the race, there is a lot

0:16:47.200 --> 0:16:52.320
<v Speaker 1>of economic gains to be had. So every country, particularly

0:16:52.320 --> 0:16:54.560
<v Speaker 1>the US and China, do not want to let up

0:16:54.600 --> 0:16:55.080
<v Speaker 1>in AI.

0:16:55.800 --> 0:16:58.240
<v Speaker 2>And what are the challenges ahead right now for Deep

0:16:58.280 --> 0:16:59.400
<v Speaker 2>Seek that you see.

0:16:59.680 --> 0:17:03.200
<v Speaker 1>I think one of the main challenges is what next,

0:17:03.520 --> 0:17:06.879
<v Speaker 1>What can they do that outdes what they've already done.

0:17:07.040 --> 0:17:10.600
<v Speaker 1>But there's also, I think for Deep Seak competition within

0:17:10.640 --> 0:17:15.200
<v Speaker 1>its own home ground, a bunch of China companies such

0:17:15.240 --> 0:17:19.479
<v Speaker 1>as Ali Baba and Paita and Tencent building models that

0:17:19.600 --> 0:17:25.080
<v Speaker 1>are outdoing Deep Seek's last flagship model. So there's this

0:17:25.720 --> 0:17:30.119
<v Speaker 1>pressure on Deep Seek to do better. But also I

0:17:30.160 --> 0:17:35.240
<v Speaker 1>think there is also the question around commercializing these models.

0:17:35.240 --> 0:17:38.520
<v Speaker 1>How are companies like deep Seak going to make money?

0:17:39.119 --> 0:17:42.359
<v Speaker 1>There is no clear answer yet whether Deep Seak wants

0:17:42.359 --> 0:17:44.399
<v Speaker 1>to make money and if it does, how will it

0:17:44.480 --> 0:17:47.720
<v Speaker 1>make money.

0:17:49.320 --> 0:17:51.960
<v Speaker 2>This is the Big Take Asia from Bloomberg News. I'm

0:17:51.960 --> 0:17:55.680
<v Speaker 2>wan Ha. This episode was produced by Young Young and Naomi.

0:17:56.480 --> 0:17:59.280
<v Speaker 2>It was edited by Patty Hirsh and Joshua Brucein. It

0:17:59.320 --> 0:18:02.159
<v Speaker 2>was fact check by Bloomberg's editorial team and mixed and

0:18:02.200 --> 0:18:06.320
<v Speaker 2>sound designed by Taka Yasuzawa. Our senior producer is named Lushaven.

0:18:06.720 --> 0:18:10.800
<v Speaker 2>Our senior editor is Elizabeth Ponso. Our deputy executive producer

0:18:10.920 --> 0:18:14.359
<v Speaker 2>is Julia Weaver. Our executive producer is Nicole Beemster Bower.

0:18:14.520 --> 0:18:17.520
<v Speaker 2>Sage Bauman is Bloomberg's head of podcasts. If you like

0:18:17.600 --> 0:18:20.280
<v Speaker 2>this episode, make sure to subscribe and review The Big

0:18:20.359 --> 0:18:23.480
<v Speaker 2>Take Asia wherever you listen to podcasts. It really helps

0:18:23.480 --> 0:18:25.760
<v Speaker 2>people find the show. Thanks for listening, to see you

0:18:25.800 --> 0:18:26.160
<v Speaker 2>next time.