How DeepSeek is shaking up the world of AI
13:55 Share

This BBC podcast is supported by ads outside the UK. A few days ago, hardly anyone had heard of the Chinese AI chatbot DeepSeek. Now it's topped the Apple App Store charts and the buzz surrounding it has wiped a trillion dollars off the value of US tech stocks. One of the things everyone is talking about is the price tag. DeepSeek's creators say it costs just $6 million to make compared to the multi-millions spent by giants like Microsoft and Google.

This has got Silicon Valley shuddering. The AI race has been blown wide open and, rightly or wrongly, alarm bells are ringing about things like censorship, spying and cybersecurity. So in this episode, you're going to hear what DeepSeek is and how it's changing the world of AI. I'm Hannah Gelbart, and this is What in the World from the BBC World Service. ♪

Let's find out more now from our cyber correspondent, Joe Tidy. Hi, Joe. Hi. First of all, where has DeepSeat come from so suddenly and who's behind it? Well, it's come from a investment banker, a sort of quantitative hedge fund manager who's doing incredibly well in China. And he's been doing a lot of work on the

And it's spun out of a university project. And it really has come out of the blue that we first started hearing about this incredible development that was taking place around. I think it was Boxing Day just for Boxing Day. We heard about this lab in China that's releasing some really impressive models.

But there was largely kind of ignored. And then, of course, the latest version came out, which was on the 20th of January. And we started seeing some AI researchers say, whoa, there's something really, really impressive going on here, especially because the price tag that they're claiming is attached to this model is so low. Less than $6 million, they claim, is how much it costs to train. I don't actually know how it happened, but somehow we went from like the AI press

to being top of the app charts in America and really high up in the Google Play Store as well, a smash hit around the world. I don't know how he went from AI geek press to world domination in eight days. I still haven't figured that out. Have you played with it yourself? Yeah, I have. Yeah, it's exactly the same as you would get from a kind of chat GPT or a Gemini. The latest model that's really impressed everyone is called a reasoning model.

And it's very similar to chat GPT-4-0, which takes its time and it really thinks and it shows you its thinking. It says, right, the user has asked me how the moon orbits the Earth. I need to let him or her know about centripetal force and gravity, but they might not know about this. So I need to explain that. And it goes on and on and on. And then it gives you a really thorough answer. And it's that reasoning, that kind of taking its time and going through the thinking steps

that's really impressed people. You said it has been made much more cheaply or trained much more cheaply than its competitors. How did they manage to do that? They haven't invented anything new here. What they seem to have done is kind of innovated on techniques that have previously been used. And I think one of the reasons for this is necessity, because what's happening to AI companies in China is they are being forced to make do with less powerful computer chips

because of sanctions from the US stopping the most powerful computer chips getting into China. So we know that the founder of DeepSeek said that they stockpiled thousands of these less powerful GPUs before the ban came in. And they've been using these kind of ones that have been developed by the giant NVIDIA, which is an American company, that are just about safe enough not to be in breach of sanctions. And through the necessity of having to use these less powerful chips and fewer of them,

They've developed these techniques and they've honed these techniques to train the model at a much cheaper rate. And when you compare it, that's $6 million they claim to have spent, which is being disputed, by the way. But when you compare that to OpenAI's latest model, ChatGPT 4.0, that's about $100 million they've spent on developing that. So that's where the real fear has come in in the industry. Hang on a minute. Maybe there's a much cheaper way of doing this. And maybe we don't need to spend as much as we are.

So is everyone pretty freaked out about this because it's come seemingly out of nowhere, cost a fraction of the price and seems to be able to do pretty much the same as the stuff that we already were using?

Yeah, and I think there are two reasons for that. Firstly is the kind of crazy commercial success. I say commercial, but actually they don't charge any money. It's free, which is the thing that's getting the likes of OpenAI scared because they charge £20 a month for this. So it's how on earth this app became such a smash hit. That's the first kind of thing that shocked the markets.

But also, as you say, it's the way they've done it, the cost and the kit they're saying that they used. Again, some people are disputing this already. We only have the company's word for it. They haven't done any interviews because it's Chinese Lunar New Year. And all the requests, certainly from me and my team, are being ignored by the company. The real kind of shock to the system is maybe we don't need to do it in the way that the U.S. has been doing, which is throw money at the problem.

them. NVIDIA, which is some days when you check that the most richest and most valuable company in the world, they charge an enormous amount of money for their latest and greatest chips to the big AI companies like Google and Microsoft and OpenAI. And the worry here and the reason why it's shocked the stock market is maybe we don't need all that kit.

If they have been able to do it as cheaply as they say, and I hear you, we don't fully know that yet. Presumably there are huge benefits because it can be rolled out to different parts of the world that otherwise might not have been able to afford having access to that AI. There's also all this talk about the environmental impact of AI. If the costs come down, could it benefit us?

Well, I think there are two ways of looking at this. First, you think, oh, hang on a minute. Here's a company that says that it can do it with less power, energy, electricity and less money. That's a good thing. That means that it opens up the market to other countries. For example, in the UK, the UK government's been talking about becoming an AI superpower. But the thing that always holds back other countries, if you're not China or America, is investment and it's money.

And it's infrastructure and it's having the right kit and the right expertise. But if you can do all of that, if you can if you can shock the market and become a world leader without any of those benefits that China or the US give you, then maybe anyone can enter this market. But I think the other thing to think about this is that the race to be the best in AI has just stepped up a notch.

So I think there's a pessimistic view, which is actually my view, which is that now the kind of the gloves are off. And I feel like all thoughts of energy saving, money saving, ethics, safety. Personally, I'm a bit worried that's all going to go out the window because now the race is on. You say that on the surface, it looks more or less like chat GPT. What are some of the differences in how it works?

The only difference that I've noticed is that when you do ask it things that are controversial to the Chinese government. So, for example, if you ask it about Tiananmen Square, what's really interesting is that it will start writing an answer. It will say the Tiananmen Square massacre was blah, blah, blah, blah. And then it will stop and it will kind of really strange. It kind of remembers things.

That it's a Chinese model and it deletes that and it says, sorry, I can't talk about that. Let's talk about something else. And you get the same kind of thing when you talk about controversial topics, for example, Taiwan or Winnie the Pooh, the cartoon, which is sometimes linked to President Xi Jinping in a very controversial way, which the government blocks.

So you do realize quite quickly when you're using this app that it is a bit different from the US giants and it doesn't answer things that we are kind of used to. But then again, will that hold back the app? I don't necessarily think it will because how often do you need to have those kind of conversations and that kind of research? Will you put up with a little bit of censorship in exchange for what is a very powerful AI chatbot and it's free?

That's something that people are perhaps overestimating is the impact that that kind of censorship will have on the user. Because if it can answer the coding questions that you ask when you're trying to build a program or if it can help you write emails better, it can help you come up with marketing plans. That's the kind of stuff that people use every day. The TikTok factor looms large here because people,

In the US, for the last few years, there's been this massive concern, which has spread around the world to some degree, that TikTok is a Chinese-owned firm taking all this data from Western users. And what's it doing with that data? Is it ending up back in Beijing? Could it be used as a spying tool, as some sort of espionage weapon against the West or worse?

And right now, there's this really strange limbo that TikTok is in where the US law courts have said, yes, it's a national security risk. But Donald Trump, who wanted the ban originally, but has now come in in the second term as president, he's now a fan of the app because he did quite well on it during the campaign. So will TikTok survive the next couple of months? We don't know. And then suddenly you have this other Chinese app that's come through and is a smash hit and

Everyone is thinking, whoa, hang on a minute. If there's a problem with TikTok, should we have a problem with DeepSeek? We are seeing the first sort of murmurings of dissent. In Australia, one of the government ministers spoke about people should be careful when they download this app because it is from China. We saw overnight that the US Navy has banned DeepSeek on staff phones. I think we're going to see more of that kind of creeping around the world as people come to terms with the fact that this is a Chinese app.

But I think the TikTok factor will have a big bearing on what happens to DeepSeek, because we will know in the next, what, I think six weeks now, what TikTok's fate will be. And they will be very much tied to DeepSeek in the sense of however the US treats TikTok will probably be how it treats DeepSeek. If anything, DeepSeek is more Chinese than TikTok.

because at least TikTok has made efforts to appear to have, for example, an international CEO and the data is processed in different parts of the world. There's none of that with DeepSeek. It's come in overnight and been a smash hit and people are signing up to the terms and conditions without even reading them. When I signed up, it didn't even let you read the terms and conditions or privacy policy because those links were dead. I had to go on and read them separately. And I don't know how many millions of other people didn't even bother to read them.

Jo mentioned there are concerns in the US over the use of Chinese tech. But what about in Taiwan? Taiwan's an island and for all practical purposes, it's been independent since 1950. But that's not how China sees it. China sees it as a breakaway province that will eventually be part of the country. And Beijing hasn't ruled out the use of force to get there. That means that security issues are always a point of consideration for the Taiwanese.

Benny Liu is a reporter with BBC Chinese. He's based in Hong Kong and has been covering the reaction in Taiwan. There are so many concerns about deep-seek safety and privacy in Taiwan. Many Taiwanese are bringing up sensitive topics like the Tiananmen Square or gay marriage, and they feel it's impossible to have those discussions with these Chinese AI robots.

Some worry that it could be a spy tool or that their data might be used or reused by the Chinese government, especially since the data is stored in China, similar to TikTok. So they hope the Taiwanese government will be more cautious about this product, without even suggesting it should be banned in government and military sectors. However, there are many people who think using this product is not a big deal.

for example young people are using TikTok and Raynode to promote their product or making friends and many people use WeChat to work with their Chinese clients so they don't think it has any security problem here. Overall it really highlights how divided opinions are in Taiwan regarding Chinese technology. This isn't the first product to spark this kind of debate and it probably won't be the last.

Joe, all of this comes at a really interesting time. So early on in Donald Trump's second term, he just announced the Stargate project to invest in AI and keep big tech in the US. How is DeepSeek going to work with all of that, this huge rivalry between American and Chinese tech firms?

It's really fascinating. Donald Trump has, you know, I would have thought that we just talked about security. I would have thought he would have said something about that, but he didn't. He said it's a wake up call for US AI tech companies. He sort of welcomed the competition. So this is what I'm talking about with my my slight worry that this is going to be a gloves off race now, because I think what we are going to see is everything is going to be ramped up to meet the deep seek threat.

Because, of course, this isn't just about tech companies doing well against each other. This is a national security priority for the Trump administration to be a leader in AI. We just saw, as you say, $500 billion has been earmarked for infrastructure to help the U.S. tech giants get ahead and stay ahead in the race. But certainly what we've seen here is really put a cat amongst the pigeons. Joe, thank you so much. Thank you.

That brings us to the end of today's episode. Thank you so much for joining us. I'm Hannah Gelbart. This is What In The World from the BBC World Service. And we'll be back with another episode soon. Bye. Yoga is more than just exercise. It's the spiritual practice that millions swear by.

And in 2017, Miranda, a university tutor from London, joins a yoga school that promises profound transformation. It felt a really safe and welcoming space. After the yoga classes, I felt amazing. But soon, that calm, welcoming atmosphere leads to something far darker, a journey that leads to allegations of grooming, trafficking and exploitation across international borders. ♪

I don't have my passport, I don't have my phone, I don't have my bank cards, I have nothing. The passport being taken, the being in a house and not feeling like they can leave.

You just get sucked in so gradually.

And it's done so skillfully that you don't realize. And it's like this, the secret that's there. I wanted to believe that, you know, that whatever they were doing, even if it seemed gross to me,

was for some spiritual reason that I couldn't yet understand. Revealing the hidden secrets of a global yoga network. I feel that I have no other choice. The only thing I can do is to speak about this and to put my reputation and everything else on the line. I want truth and justice.

And for other people to not be hurt, for things to be different in the future. To bring it into the light and almost alchemise some of that evil stuff that went on and take back the power. World of Secrets, Season 6, The Bad Guru. Listen wherever you get your podcasts.