WEBVTT - Here's Why DeepSeek is a Wake-Up Call for AI Titans

0:00:02.560 --> 0:00:07.040
<v Speaker 1>Bloomberg Audio Studios, podcasts, radio news.

0:00:09.240 --> 0:00:11.840
<v Speaker 2>I'm Stephen Carol and this is here is Why, where

0:00:11.840 --> 0:00:14.200
<v Speaker 2>we take one news story and explain it in just

0:00:14.240 --> 0:00:16.760
<v Speaker 2>a few minutes with our experts here at Bloomberg.

0:00:19.200 --> 0:00:23.000
<v Speaker 1>It was done very cheaply. So why the crazy amounts

0:00:23.040 --> 0:00:25.600
<v Speaker 1>of spending that are happening from the US big tech companies.

0:00:25.720 --> 0:00:30.000
<v Speaker 2>What's happening here is the collapse in the cost of innovation.

0:00:30.400 --> 0:00:34.200
<v Speaker 1>What this development has shown is the hardware that these

0:00:34.200 --> 0:00:37.959
<v Speaker 1>companies have procured, they weren't using it efficiently. A lot

0:00:38.000 --> 0:00:41.120
<v Speaker 1>of question marks around deep Seek, in particular, how much

0:00:41.320 --> 0:00:43.920
<v Speaker 1>did it really cost for them to develop this new

0:00:44.080 --> 0:00:46.400
<v Speaker 1>quote unquote cheap model that they've self reported.

0:00:46.560 --> 0:00:49.159
<v Speaker 2>The implication is, of course that things got cheap, that

0:00:49.200 --> 0:00:51.360
<v Speaker 2>things got more simple, but we'll have to see that

0:00:51.440 --> 0:00:54.840
<v Speaker 2>manifest itself in the real world. Shocking, a game changer,

0:00:54.960 --> 0:00:57.480
<v Speaker 2>a Spotnik moment, just a few of the things that

0:00:57.480 --> 0:01:00.680
<v Speaker 2>have been said about the Chinese AI startup deep Seek.

0:01:00.920 --> 0:01:04.600
<v Speaker 2>The company shot to global fame over a weekend after

0:01:04.640 --> 0:01:07.679
<v Speaker 2>the launch of it's R one chatbart, seen as arrival

0:01:07.720 --> 0:01:10.840
<v Speaker 2>to the likes of Chat GPT, but crucially, the company

0:01:10.840 --> 0:01:15.080
<v Speaker 2>says it used less expensive chips. So here's why Deep

0:01:15.080 --> 0:01:17.880
<v Speaker 2>Seek is a wake up call for the AI Titans.

0:01:20.240 --> 0:01:22.840
<v Speaker 2>Joining me Iwo to discuss Boomberg's TV anchor. Tom mackenzie,

0:01:22.880 --> 0:01:25.880
<v Speaker 2>a former China correspondent and our resident tech watcher. Thanks

0:01:25.880 --> 0:01:27.920
<v Speaker 2>for being with us. Tom, First of all, tell me

0:01:27.920 --> 0:01:30.680
<v Speaker 2>about R one, the product that Deep Seek makes. Is

0:01:30.720 --> 0:01:32.440
<v Speaker 2>it something that's really comparable to.

0:01:32.600 --> 0:01:35.960
<v Speaker 1>Chat GPT or to Gemini so on a number of

0:01:36.120 --> 0:01:40.280
<v Speaker 1>metrics measured by experts, and there are platforms online where

0:01:40.280 --> 0:01:43.039
<v Speaker 1>you can see how these measurements are displayed and how

0:01:43.040 --> 0:01:47.240
<v Speaker 1>they categorize these models. Then yes, R one competes with

0:01:47.400 --> 0:01:50.880
<v Speaker 1>chat gbt's most sophisticated model, the latest model to zero one,

0:01:50.880 --> 0:01:53.600
<v Speaker 1>but also the models from Gemini and Anthropic. It is

0:01:53.800 --> 0:01:57.240
<v Speaker 1>up there top of the list alongside those kind of players.

0:01:57.520 --> 0:02:01.160
<v Speaker 1>It is sophisticated. It is a text based model and

0:02:01.280 --> 0:02:03.760
<v Speaker 1>chat bop you can download it onto your phone. It

0:02:03.840 --> 0:02:06.360
<v Speaker 1>is not multimodal, so it doesn't produce video, it doesn't

0:02:06.400 --> 0:02:10.680
<v Speaker 1>produce pictures, but what it does do is reasoning. Like

0:02:11.080 --> 0:02:14.320
<v Speaker 1>one from chat GBT, it goes through how it comes

0:02:14.320 --> 0:02:17.400
<v Speaker 1>to its responses, and unlike O one, it actually displays

0:02:17.440 --> 0:02:19.640
<v Speaker 1>those for you, so you can see the chat bot

0:02:19.760 --> 0:02:23.120
<v Speaker 1>processing your question, going through how it's going to get

0:02:23.160 --> 0:02:26.679
<v Speaker 1>to the answer, and that level of transparency has led

0:02:26.720 --> 0:02:29.960
<v Speaker 1>to a lot of tech enthusiasts out there who are

0:02:29.960 --> 0:02:32.640
<v Speaker 1>getting their hands on this thing responding very favorably, and

0:02:32.639 --> 0:02:35.079
<v Speaker 1>that's led to a lot of optimism and some very

0:02:35.080 --> 0:02:38.280
<v Speaker 1>positive feedback about what this chatbot based on deep seeks

0:02:38.320 --> 0:02:40.440
<v Speaker 1>model can actually do. Now, one thing it can't do

0:02:40.960 --> 0:02:44.080
<v Speaker 1>is answer an honest question about Hi Jinping and his

0:02:44.200 --> 0:02:47.200
<v Speaker 1>leadership or what happened in Chaneman Square in nineteen eighty nine,

0:02:47.240 --> 0:02:50.520
<v Speaker 1>because it is not allowed to full foul of Chinese

0:02:50.520 --> 0:02:53.400
<v Speaker 1>government censorship. So there is that important caveat.

0:02:53.160 --> 0:02:55.639
<v Speaker 2>Okay, all very interesting. What do we know about how

0:02:55.680 --> 0:02:56.399
<v Speaker 2>it was developed?

0:02:56.800 --> 0:02:59.359
<v Speaker 1>What we know is what is claimed by deep Seek.

0:02:59.680 --> 0:03:02.240
<v Speaker 1>So they claim that it was developed at a fraction

0:03:02.480 --> 0:03:05.160
<v Speaker 1>of the cost of some of their competitors, around six

0:03:05.440 --> 0:03:09.040
<v Speaker 1>million US dollars. They claim the model was trained on

0:03:09.560 --> 0:03:13.080
<v Speaker 1>much older chips, not the most cutting edge in video chips,

0:03:13.120 --> 0:03:17.160
<v Speaker 1>because those are restricted from the Chinese markets. We know

0:03:17.360 --> 0:03:21.280
<v Speaker 1>that they have very innovative and sophisticated engineers. They've been

0:03:21.280 --> 0:03:25.440
<v Speaker 1>recruiting talent for the top universities domestically in China systems engineers.

0:03:25.480 --> 0:03:28.239
<v Speaker 1>We also know that they put in place an infrastructure

0:03:28.320 --> 0:03:30.360
<v Speaker 1>and a method of building models called a mixture of

0:03:30.360 --> 0:03:33.160
<v Speaker 1>experts procedure, which basically has lots of mini models much

0:03:33.160 --> 0:03:35.560
<v Speaker 1>easier to put together, and if you align them, you

0:03:35.560 --> 0:03:39.120
<v Speaker 1>can create these efficiencies. So the engineers, the mixture of

0:03:39.200 --> 0:03:42.640
<v Speaker 1>experts method that they've used, they say has led them

0:03:42.640 --> 0:03:45.800
<v Speaker 1>to creating these efficiencies, building a model much more cheaply,

0:03:45.920 --> 0:03:48.640
<v Speaker 1>using less efistigated chips, and much more quickly. They say

0:03:48.640 --> 0:03:52.080
<v Speaker 1>they built this model, design this model within about two months.

0:03:52.120 --> 0:03:55.040
<v Speaker 1>Now there are question marks. Microsoft and open ai are

0:03:55.040 --> 0:03:58.200
<v Speaker 1>scrutinizing whether or not deep Seak actually lent on open

0:03:58.240 --> 0:04:00.760
<v Speaker 1>AI's own model to learn from the outputs from that

0:04:00.760 --> 0:04:02.560
<v Speaker 1>model that they then fed into the training of the

0:04:02.560 --> 0:04:06.480
<v Speaker 1>deep Seat model that potentially went over and above what

0:04:06.760 --> 0:04:10.080
<v Speaker 1>was allowed. That is being scrutinized, And we have to

0:04:10.080 --> 0:04:11.600
<v Speaker 1>take them on face value when they talk about the

0:04:11.680 --> 0:04:15.000
<v Speaker 1>chips they're using, because we don't know exactly how they

0:04:15.000 --> 0:04:16.520
<v Speaker 1>built this. But if we take them on face value,

0:04:16.560 --> 0:04:19.880
<v Speaker 1>the more cheaply, more cost effectively, in a shorter time

0:04:20.320 --> 0:04:23.360
<v Speaker 1>and with a slightly different method, that created a very

0:04:23.400 --> 0:04:25.640
<v Speaker 1>efficient and very capable model.

0:04:25.880 --> 0:04:29.360
<v Speaker 2>Okay, got questions being asked, as you know, more broadly,

0:04:29.800 --> 0:04:32.119
<v Speaker 2>when we talk about AI open on now we've mainly

0:04:32.120 --> 0:04:35.160
<v Speaker 2>been talking about the models developed by American companies. What

0:04:36.040 --> 0:04:38.599
<v Speaker 2>is this revealing something that we didn't already know about

0:04:38.680 --> 0:04:41.960
<v Speaker 2>China's AI industry? How does it compare to what we

0:04:42.000 --> 0:04:42.839
<v Speaker 2>know out of the US.

0:04:43.160 --> 0:04:46.040
<v Speaker 1>There is an argument that we've had a blind spot

0:04:46.080 --> 0:04:48.520
<v Speaker 1>when it comes to the innovation that's coming out of China.

0:04:48.760 --> 0:04:50.680
<v Speaker 1>There's been a lot of talk about the slowdown in

0:04:50.720 --> 0:04:52.920
<v Speaker 1>the economy rightly, so, there's been a lot of concern

0:04:52.960 --> 0:04:54.680
<v Speaker 1>about the real estate market. There's been a lot of

0:04:54.680 --> 0:04:57.640
<v Speaker 1>concern about a crackdown on technology in recent years out

0:04:57.680 --> 0:05:01.040
<v Speaker 1>of China. That has allowed some to overlook the real

0:05:01.080 --> 0:05:03.919
<v Speaker 1>innovation that is happening in that Chinese market. One of

0:05:03.920 --> 0:05:07.240
<v Speaker 1>the most competitive places to build and test technology is

0:05:07.279 --> 0:05:09.440
<v Speaker 1>in the Chinese markets because you have so many people

0:05:09.440 --> 0:05:11.880
<v Speaker 1>who pile in, they test their products, they fight to

0:05:11.960 --> 0:05:14.000
<v Speaker 1>the death, and then the survivors come out on top.

0:05:14.000 --> 0:05:16.080
<v Speaker 1>And if they can compete in the Chinese market, my goodness,

0:05:16.080 --> 0:05:17.920
<v Speaker 1>they can compete globally. And we've seen that, whether it

0:05:18.000 --> 0:05:20.839
<v Speaker 1>is with drone makers like DGI, whether it's like TikTok,

0:05:20.920 --> 0:05:23.839
<v Speaker 1>social media companies like TikTok, or whether it's indeed the

0:05:23.880 --> 0:05:26.520
<v Speaker 1>solar panel makers of China. Across all those different areas,

0:05:26.560 --> 0:05:28.480
<v Speaker 1>they compete in the domestic market, they win, and then

0:05:28.480 --> 0:05:30.480
<v Speaker 1>they go to compete internationally. I haven't even mentioned the

0:05:30.520 --> 0:05:32.919
<v Speaker 1>electric vehicle makers that are posing a challenge to Europe's

0:05:32.960 --> 0:05:35.920
<v Speaker 1>EV model now as well. They have the engineers, They

0:05:35.920 --> 0:05:38.200
<v Speaker 1>have almost double the number of engineers of the US

0:05:38.279 --> 0:05:41.560
<v Speaker 1>in terms of AI engineers. They have the data in

0:05:41.839 --> 0:05:44.880
<v Speaker 1>vast quantities what they don't have, and the most sophisticated chips.

0:05:45.200 --> 0:05:47.799
<v Speaker 1>If Deep Sea really is an example of how to

0:05:47.839 --> 0:05:52.320
<v Speaker 1>circumvent that by using older chips, then China has all

0:05:52.360 --> 0:05:56.400
<v Speaker 1>the ingredients it needs to compete on the global level. Interestingly,

0:05:56.680 --> 0:05:59.720
<v Speaker 1>we also saw another Chinese company, Ali Baba, coming out

0:05:59.720 --> 0:06:03.000
<v Speaker 1>with model this week that is also as capable as

0:06:03.000 --> 0:06:06.080
<v Speaker 1>the most sophisticated models coming out of the US. So

0:06:06.400 --> 0:06:09.520
<v Speaker 1>every indication suggests that China is a serious player, and

0:06:09.560 --> 0:06:11.920
<v Speaker 1>if it hasn't caught up with the US yet, then

0:06:11.960 --> 0:06:13.560
<v Speaker 1>it's very very close to doing that.

0:06:13.720 --> 0:06:15.960
<v Speaker 2>It does seem like all of a sudden everyone has

0:06:16.000 --> 0:06:19.839
<v Speaker 2>an opinion on deep Seek and its breakthrough. Even the

0:06:19.960 --> 0:06:21.239
<v Speaker 2>US President.

0:06:21.160 --> 0:06:25.120
<v Speaker 1>The release of deep Seek AI from a Chinese company

0:06:25.160 --> 0:06:27.760
<v Speaker 1>should be a wake up call for our industries that

0:06:27.800 --> 0:06:30.760
<v Speaker 1>we need to be laser focused on competing to win.

0:06:31.160 --> 0:06:34.400
<v Speaker 2>So, Tom, what do the AIA companies in the US

0:06:34.600 --> 0:06:37.960
<v Speaker 2>need to be woken up to with this deep Seek story?

0:06:38.080 --> 0:06:40.800
<v Speaker 1>Well, interestingly, it's earnings week as well, so we've been

0:06:40.800 --> 0:06:42.800
<v Speaker 1>hearing from the CEOs of some of those major players.

0:06:42.839 --> 0:06:44.960
<v Speaker 1>We've been hearing from Mark Zuckerberg from Meta, We've been

0:06:45.000 --> 0:06:47.680
<v Speaker 1>hearing from sati In adell At, CEO of Microsoft. What

0:06:47.760 --> 0:06:50.400
<v Speaker 1>you're not hearing from them is any walk back on

0:06:50.440 --> 0:06:53.320
<v Speaker 1>the spend around CAPEX, the spend on the chips, the

0:06:53.360 --> 0:06:55.320
<v Speaker 1>spend on the data centers, the spend on the servers,

0:06:55.600 --> 0:06:57.960
<v Speaker 1>and the spend on the energy that's needed to power

0:06:58.000 --> 0:07:00.599
<v Speaker 1>this AI revolution. Because one of the major questions that

0:07:00.600 --> 0:07:02.960
<v Speaker 1>Deep Seek has posed this week is is all of

0:07:03.000 --> 0:07:05.839
<v Speaker 1>that spending worth it? Do we need to spend tens

0:07:05.839 --> 0:07:08.360
<v Speaker 1>of billions of dollars and all of that AI infrastructure?

0:07:08.800 --> 0:07:12.160
<v Speaker 1>If a bootstrap company in China with six million dollars

0:07:12.240 --> 0:07:14.800
<v Speaker 1>can produce a model that competes with Chat GBT at

0:07:14.840 --> 0:07:18.000
<v Speaker 1>a fraction of the price. But the CEOs are not

0:07:18.160 --> 0:07:22.520
<v Speaker 1>walking back from these investment commitments. In fact, Mark Zuckerberg,

0:07:22.520 --> 0:07:25.040
<v Speaker 1>who praised deep Seat and said there was real innovation there.

0:07:25.240 --> 0:07:28.280
<v Speaker 1>He's committed to sixty five billion dollars of spending this year.

0:07:28.680 --> 0:07:31.520
<v Speaker 1>He says they're going to be building their own AI agents,

0:07:31.520 --> 0:07:33.320
<v Speaker 1>and he hopes the Meta is going to be getting

0:07:33.360 --> 0:07:35.360
<v Speaker 1>those agents to a billion people by the end of

0:07:35.360 --> 0:07:37.200
<v Speaker 1>this year, and that they'll be leading in that space.

0:07:37.280 --> 0:07:40.320
<v Speaker 1>Microsoft Satiy and Nadella also praising deep Seek, but his

0:07:40.480 --> 0:07:42.800
<v Speaker 1>take was this is going to lead to greater adoption

0:07:42.840 --> 0:07:46.120
<v Speaker 1>of artificial intelligence. It will drive down prices, and longer

0:07:46.200 --> 0:07:48.960
<v Speaker 1>term that will be good news for Microsoft. Microsoft committing

0:07:48.960 --> 0:07:52.680
<v Speaker 1>to spend eighty billion dollars this fiscal year on AI infrastructure.

0:07:53.080 --> 0:07:55.560
<v Speaker 2>Okay, a rude awakening for some, perhaps not for everyone.

0:07:55.600 --> 0:07:58.040
<v Speaker 2>Tom McKenzie, our Blimberg TV anchor, thank you very much

0:07:58.080 --> 0:08:00.880
<v Speaker 2>for joining us for more xp like this from our

0:08:00.920 --> 0:08:03.679
<v Speaker 2>team of twenty nine hundred journalists and analysts around the world.

0:08:03.720 --> 0:08:06.560
<v Speaker 2>Search for Quick Take on the Bloomberg website or Bloomberg

0:08:06.600 --> 0:08:11.320
<v Speaker 2>Business app. I'm Stephen Carroll. This is Here's why I'll

0:08:11.360 --> 0:08:13.600
<v Speaker 2>be back next week with more thanks for listening