WEBVTT - How DeepSeek is shaking up AI

0:00:04.880 --> 0:00:08.200
<v Speaker 1>From The Australian. Here's what's on the front. I'm Christinamiot.

0:00:08.440 --> 0:00:16.720
<v Speaker 1>It's Wednesday, January twenty nine. Athletics Australia has lashed at

0:00:16.720 --> 0:00:20.000
<v Speaker 1>the Western Australian Institute of Sport for its treatment of

0:00:20.040 --> 0:00:22.840
<v Speaker 1>a pole volt coach who blew the whistle on sexual

0:00:22.840 --> 0:00:27.760
<v Speaker 1>assault allegations. Chair Jane Fleming and CEO Simon Hollingsworth have

0:00:27.880 --> 0:00:30.680
<v Speaker 1>called for the WA body to better support whistle blow

0:00:30.760 --> 0:00:38.839
<v Speaker 1>up Paul Burgess and reinstate his training privileges. New Zealand's

0:00:38.840 --> 0:00:42.440
<v Speaker 1>Resources minister has launched an attack on banks that refuse

0:00:42.600 --> 0:00:45.640
<v Speaker 1>to finance minds on the eve of the release of

0:00:45.680 --> 0:00:49.479
<v Speaker 1>his government's mineral strategy. Shane Jones also backed at the

0:00:49.479 --> 0:00:54.840
<v Speaker 1>Fijian PM in saying Australia's not to blame for climate change.

0:00:54.960 --> 0:00:57.640
<v Speaker 1>Those exclusive stories are live right now at the Australian

0:00:57.680 --> 0:01:03.800
<v Speaker 1>dot Com dot Au. Science Minister Ed Husick says a

0:01:03.880 --> 0:01:08.720
<v Speaker 1>new Chinese AI platform comes with national security concerns, and

0:01:08.800 --> 0:01:12.000
<v Speaker 1>he says he'll take advice from intelligence experts on any

0:01:12.000 --> 0:01:16.720
<v Speaker 1>potential threats. Deep Seek sent shockwaves through the tech industry

0:01:16.760 --> 0:01:20.600
<v Speaker 1>and financial markets this week. So what is it and

0:01:20.800 --> 0:01:35.400
<v Speaker 1>where's it sending the AI race. That's today's episode. Last week,

0:01:35.480 --> 0:01:38.759
<v Speaker 1>on the frenzied first day of his new presidential term,

0:01:38.920 --> 0:01:42.839
<v Speaker 1>Donald Trump announced a multi billion dollar commitment to artificial

0:01:42.880 --> 0:01:44.320
<v Speaker 1>intelligence infrastructure.

0:01:45.400 --> 0:01:50.800
<v Speaker 2>These world leading technology giants are announcing the formation of Stargate,

0:01:51.080 --> 0:01:54.400
<v Speaker 2>a new American company that will invest five hundred billion

0:01:54.440 --> 0:01:58.000
<v Speaker 2>dollars at least in AI infrastructure in the United States,

0:01:58.040 --> 0:02:00.680
<v Speaker 2>creating over one hundred thousand are jobs.

0:02:00.680 --> 0:02:04.640
<v Speaker 1>Almost immediately about the same time, on the other side

0:02:04.640 --> 0:02:08.640
<v Speaker 1>of the planet, a Chinese company called deep Seek unveiled

0:02:08.639 --> 0:02:12.080
<v Speaker 1>an AI model it says can compete with market leaders

0:02:12.120 --> 0:02:17.120
<v Speaker 1>like Chat GPT for a fraction of the cost. The

0:02:17.160 --> 0:02:21.680
<v Speaker 1>world's financial markets reacted with shock and awe with Navidia,

0:02:21.720 --> 0:02:24.920
<v Speaker 1>the world's most valuable public company and the maker of

0:02:24.960 --> 0:02:29.080
<v Speaker 1>the computer chips most American AI models rely on bearing

0:02:29.200 --> 0:02:32.239
<v Speaker 1>the brunt of it Well, shairs of Santa Clara's in

0:02:32.320 --> 0:02:35.360
<v Speaker 1>video fell hard this morning on the popularity of a

0:02:35.400 --> 0:02:39.960
<v Speaker 1>new Chinese artificial intelligence advancement. Deepseek is impressive enough, but

0:02:40.040 --> 0:02:42.600
<v Speaker 1>it made people wonder, do you really need those Nvidia

0:02:42.760 --> 0:02:43.480
<v Speaker 1>AI chaps?

0:02:43.560 --> 0:02:46.160
<v Speaker 2>That's the whole question, right, I mean ignoring the fact

0:02:46.160 --> 0:02:48.280
<v Speaker 2>that a number of reports.

0:02:48.080 --> 0:02:58.920
<v Speaker 3>That I think it caught everyone off guard, even the

0:02:58.919 --> 0:03:01.480
<v Speaker 3>best and brightest in Wall Street and in Silicon Valley.

0:03:02.200 --> 0:03:04.680
<v Speaker 1>Jared Lynch is The Australian's Technology editor.

0:03:05.480 --> 0:03:09.680
<v Speaker 3>They thought the US was miles ahead, not even miles ahead,

0:03:09.720 --> 0:03:12.440
<v Speaker 3>thousands of miles ahead than China. They had a trade

0:03:12.440 --> 0:03:16.280
<v Speaker 3>embargo on Navidia chips, the world's most advanced chips. And

0:03:16.320 --> 0:03:19.359
<v Speaker 3>then the fact that China could produce at a far

0:03:19.440 --> 0:03:23.440
<v Speaker 3>lower costs, mind you, an AI model that is almost on.

0:03:23.360 --> 0:03:26.600
<v Speaker 4>Par with chat GPT was extraordinary.

0:03:30.400 --> 0:03:33.880
<v Speaker 1>The AI model released by deep Seek is called R one.

0:03:34.440 --> 0:03:38.120
<v Speaker 1>It's a generative AI platform similar to Chat GPT and

0:03:38.320 --> 0:03:42.200
<v Speaker 1>Meta AI. Deep Seek is backed by Leoon Wen Fung.

0:03:42.600 --> 0:03:45.440
<v Speaker 1>He's the manager of the eight billion dollar hedge fund

0:03:45.640 --> 0:03:49.440
<v Speaker 1>high Flyer. The company says it forked out just under

0:03:49.560 --> 0:03:53.080
<v Speaker 1>nine million Australian dollars to train R one, and says

0:03:53.120 --> 0:03:56.560
<v Speaker 1>it did it without Navidia's top of the line computer chips.

0:03:57.400 --> 0:04:01.240
<v Speaker 1>By comparison, the most recent iteration of chat GPT cost

0:04:01.280 --> 0:04:04.560
<v Speaker 1>around one hundred million dollars to train before it was

0:04:04.640 --> 0:04:09.120
<v Speaker 1>released in November, and deep seeks model is open source,

0:04:09.520 --> 0:04:12.880
<v Speaker 1>meaning other developers can adopt and adapt it. For their

0:04:12.920 --> 0:04:20.120
<v Speaker 1>own purposes, hence the financial panic. Jared was that market

0:04:20.160 --> 0:04:21.440
<v Speaker 1>reaction proportionate?

0:04:21.560 --> 0:04:23.719
<v Speaker 4>Do you think time will tell?

0:04:23.880 --> 0:04:26.880
<v Speaker 3>There have been some analysts that have called bs to

0:04:26.960 --> 0:04:30.960
<v Speaker 3>put it politely on deep seaks claims that is that

0:04:31.000 --> 0:04:34.240
<v Speaker 3>it only cost five point six million US to train

0:04:34.320 --> 0:04:38.000
<v Speaker 3>this model, whereas some US models have cost billions to

0:04:38.080 --> 0:04:41.640
<v Speaker 3>train because there's just not enough data to show. Okay,

0:04:42.240 --> 0:04:44.920
<v Speaker 3>how much does this cost to train? And how much

0:04:45.120 --> 0:04:48.080
<v Speaker 3>do the data centers cost to run deep seats? So

0:04:48.120 --> 0:04:51.039
<v Speaker 3>that's all unknown at this stage. At the moment, the

0:04:51.080 --> 0:04:55.400
<v Speaker 3>market's caught up in Wow, China's actually delivered this AI model.

0:04:55.880 --> 0:04:59.000
<v Speaker 3>What risk does that pose to the big American companies

0:04:59.000 --> 0:05:01.560
<v Speaker 3>that have invested billions in creating this tech.

0:05:02.320 --> 0:05:03.560
<v Speaker 4>But the thing is we've got to.

0:05:03.560 --> 0:05:06.000
<v Speaker 3>Rewind the clock back a little bit to the fifties,

0:05:06.040 --> 0:05:08.880
<v Speaker 3>to the sixties, even the early seventies, off the space race,

0:05:09.240 --> 0:05:13.320
<v Speaker 3>the US against Russia. That didn't result in a pullback

0:05:13.360 --> 0:05:17.840
<v Speaker 3>off investment. I mean, competition rarely does it becomes a battle.

0:05:18.600 --> 0:05:22.120
<v Speaker 3>Billions of dollars, trillions of dollars get poured into gaining

0:05:22.200 --> 0:05:24.240
<v Speaker 3>the upper hand, and that is what we're going.

0:05:24.160 --> 0:05:25.320
<v Speaker 4>To see in the AI race.

0:05:26.240 --> 0:05:28.760
<v Speaker 3>What the response has been, more broadly, is this is

0:05:28.880 --> 0:05:31.760
<v Speaker 3>very much now game on. China is saying to the

0:05:31.800 --> 0:05:34.760
<v Speaker 3>world that it can compete and at a far lower cost.

0:05:34.960 --> 0:05:39.040
<v Speaker 3>In fact, that's almost embarrassing to the American companies which

0:05:39.160 --> 0:05:44.039
<v Speaker 3>have spent billions and billions of dollars of investor funds

0:05:44.480 --> 0:05:48.159
<v Speaker 3>on developing and training these sophisticated AI models.

0:05:50.680 --> 0:05:53.880
<v Speaker 1>Deep six founders say they've tested are one against models

0:05:53.920 --> 0:05:57.600
<v Speaker 1>from open Ai, which owns chat GPT, and they say

0:05:57.760 --> 0:06:02.400
<v Speaker 1>it holds up. Jared, You've spent some time in the

0:06:02.480 --> 0:06:06.480
<v Speaker 1>last week playing around with deep siks are one AI model.

0:06:06.920 --> 0:06:09.159
<v Speaker 1>What did you think? Is it actually good? Yeah?

0:06:09.200 --> 0:06:11.560
<v Speaker 3>And the most part it is it works just like

0:06:11.680 --> 0:06:15.640
<v Speaker 3>chat GBT, in some cases even better. It provides lengthy,

0:06:15.839 --> 0:06:20.080
<v Speaker 3>balanced explanations to queries. For example, if you say if

0:06:20.120 --> 0:06:23.159
<v Speaker 3>Trump is a good leader, it'll say whether he's a

0:06:23.160 --> 0:06:26.560
<v Speaker 3>good leader is subjective. Here's some arguments, or here's some

0:06:26.680 --> 0:06:30.360
<v Speaker 3>arguments against. Here some neutral arguments, and this is a conclusion.

0:06:30.440 --> 0:06:32.680
<v Speaker 3>So it sort of directs the user to make up

0:06:32.680 --> 0:06:33.440
<v Speaker 3>their own mind.

0:06:34.480 --> 0:06:35.440
<v Speaker 4>The problem is.

0:06:35.839 --> 0:06:40.400
<v Speaker 3>It's AI veers into what I call autocratic intelligence, and

0:06:40.440 --> 0:06:42.479
<v Speaker 3>by that I mean it acts as an extension of

0:06:42.600 --> 0:06:46.800
<v Speaker 3>China's Communist Party, because if you ask any questions about China,

0:06:47.960 --> 0:06:50.640
<v Speaker 3>it adheres to China's censorship laws. So if you say,

0:06:50.680 --> 0:06:53.800
<v Speaker 3>what is the Tianaman Square massacre, it'll say, sorry, that's

0:06:53.839 --> 0:06:54.839
<v Speaker 3>beyond my current scope.

0:06:54.960 --> 0:06:56.240
<v Speaker 4>Let's talk about something else.

0:06:56.839 --> 0:07:01.960
<v Speaker 3>So it's very much a pro China model, or as

0:07:01.960 --> 0:07:04.560
<v Speaker 3>I like to call it, autocratic intelligence.

0:07:05.279 --> 0:07:08.640
<v Speaker 1>Do you think that, in adhering to the Communist government's

0:07:09.160 --> 0:07:14.440
<v Speaker 1>rules around censorship, Deep Seek could actually count itself out

0:07:14.440 --> 0:07:16.920
<v Speaker 1>of this race before it's really even begun.

0:07:17.640 --> 0:07:19.880
<v Speaker 3>Well, I don't think many people will realize because on

0:07:19.960 --> 0:07:23.680
<v Speaker 3>the whole it will act like a chap GBT. Why

0:07:23.760 --> 0:07:27.040
<v Speaker 3>it matters when people ask questions about China is that

0:07:27.160 --> 0:07:29.200
<v Speaker 3>is where people are getting their information. I mean, we

0:07:29.320 --> 0:07:33.160
<v Speaker 3>used to have libraries, and then Google came along and

0:07:33.200 --> 0:07:36.080
<v Speaker 3>people got their information from Google, and now increasingly people

0:07:36.080 --> 0:07:42.080
<v Speaker 3>want their information via AI summaries. And the fact that

0:07:42.120 --> 0:07:45.400
<v Speaker 3>it is pro China, I think that's almost going to

0:07:45.400 --> 0:07:48.440
<v Speaker 3>become irrelevant because if it gets what they call the

0:07:48.520 --> 0:07:51.840
<v Speaker 3>multiplier effect, where it gets a huge user base, people

0:07:51.840 --> 0:07:54.800
<v Speaker 3>won't realize that they're actually being fed a pro China message,

0:07:54.880 --> 0:07:57.280
<v Speaker 3>and that is a really, really big concern.

0:08:02.000 --> 0:08:23.760
<v Speaker 1>Coming up? What next for the AI race? Back in

0:08:23.800 --> 0:08:27.000
<v Speaker 1>the White House, an advisor to President Donald Trump reportedly

0:08:27.040 --> 0:08:30.440
<v Speaker 1>said the cheap generative AI model released last week by

0:08:30.480 --> 0:08:33.840
<v Speaker 1>deep Seek was one of the most impressive breakthroughs he'd

0:08:33.920 --> 0:08:37.560
<v Speaker 1>ever seen. In many ways, it's an escalation in the

0:08:37.600 --> 0:08:40.839
<v Speaker 1>competition between the United States and China to win the

0:08:40.840 --> 0:08:44.600
<v Speaker 1>AI race, a fact Trump was upfront about at the

0:08:44.640 --> 0:08:45.800
<v Speaker 1>Stargate announcement.

0:08:47.040 --> 0:08:51.360
<v Speaker 2>This monumental undertaking is a resounding declaration of confidence in

0:08:51.400 --> 0:08:54.199
<v Speaker 2>America's potential. What we want to do is we want

0:08:54.240 --> 0:08:56.760
<v Speaker 2>to keep it in this country. China as a competitor

0:08:56.840 --> 0:08:59.800
<v Speaker 2>and others are competitors. We wanted to be in this country.

0:09:01.480 --> 0:09:04.040
<v Speaker 1>So as Deepzick races to the top of app stores

0:09:04.080 --> 0:09:06.320
<v Speaker 1>around the globe, what happens next?

0:09:08.040 --> 0:09:12.280
<v Speaker 3>Here's Jared Lynch, Well, it shows now that the US

0:09:12.400 --> 0:09:16.280
<v Speaker 3>no longer has the monopoly on AI development. In China,

0:09:16.440 --> 0:09:20.839
<v Speaker 3>we've seen strong entrepreneurial culture which has created some of

0:09:20.880 --> 0:09:25.120
<v Speaker 3>the world's biggest companies. Just look at Ali Barber for example,

0:09:25.760 --> 0:09:28.400
<v Speaker 3>And now people are saying, wow, we can do the

0:09:28.440 --> 0:09:30.480
<v Speaker 3>same with a bit of an ounce there is another

0:09:30.520 --> 0:09:33.280
<v Speaker 3>way to be a disruptor, and we no longer have

0:09:33.360 --> 0:09:36.600
<v Speaker 3>to have the deep pockets of Silicon Valley. Anyone can

0:09:36.640 --> 0:09:40.920
<v Speaker 3>now take advantage of this technology and be a David

0:09:41.120 --> 0:09:42.239
<v Speaker 3>taken on the goliaths.

0:09:43.320 --> 0:09:46.800
<v Speaker 1>Elsewhere, and especially in parts of Europe, deepzek has been

0:09:46.840 --> 0:09:50.160
<v Speaker 1>welcomed with open arms by developers who say it's put

0:09:50.240 --> 0:09:54.600
<v Speaker 1>them in a position to compete. And just lastly, Jared

0:09:54.840 --> 0:09:59.400
<v Speaker 1>on the flip side, does that risk creating a marketplace

0:09:59.440 --> 0:10:04.000
<v Speaker 1>where small are already established operations actually can't compete or

0:10:04.080 --> 0:10:06.160
<v Speaker 1>is that just natural selection?

0:10:07.440 --> 0:10:09.160
<v Speaker 3>I think it'll be a bit of both. We're going

0:10:09.240 --> 0:10:12.280
<v Speaker 3>to see some smaller operators be weeded out for a

0:10:12.360 --> 0:10:16.280
<v Speaker 3>variety of reasons. Call that natural selection. And then we're

0:10:16.280 --> 0:10:18.720
<v Speaker 3>going to see a wholely of smaller operators come up

0:10:18.760 --> 0:10:22.559
<v Speaker 3>in areas that otherwise we had never thought of before,

0:10:22.760 --> 0:10:26.439
<v Speaker 3>as we've seen with deep Seek. So we are entering

0:10:26.640 --> 0:10:30.240
<v Speaker 3>uncharted territory, almost like the wild West. And it's been

0:10:30.280 --> 0:10:33.720
<v Speaker 3>even made more so by Donald Trump's decision to rescind

0:10:33.880 --> 0:10:39.040
<v Speaker 3>an executive order of Joe Biden's which gave the government

0:10:39.160 --> 0:10:43.240
<v Speaker 3>too much control of AI. So the message from Trump

0:10:43.320 --> 0:10:46.400
<v Speaker 3>is telling us companies to go forth and innovate and

0:10:46.440 --> 0:10:47.120
<v Speaker 3>do it quickly.

0:10:47.320 --> 0:10:49.439
<v Speaker 4>I've got your back because I.

0:10:49.440 --> 0:10:52.679
<v Speaker 3>Want the US to be the AI leaders, and that

0:10:52.800 --> 0:10:54.160
<v Speaker 3>means we've all got to buckle up.

0:11:02.040 --> 0:11:06.040
<v Speaker 1>Jared Lynch is The Australian's Technology editor. You can read

0:11:06.080 --> 0:11:09.560
<v Speaker 1>all our reporting on artificial intelligence and technology right now

0:11:09.559 --> 0:11:11.440
<v Speaker 1>at the Australian dot com dot au