This is The Guardian. Today, how a Chinese hedge fund boss created a chatbot that could change the world forever, but at what cost?
Right, don't tell the Guardian's technology security team, but I have just downloaded a new app. It is free on the App Store. Number one for productivity apparently. There are a lot of Chinese characters on the description. It's got an avatar of a very chewy looking blue whale, and it is promising me seamless interaction with an AI assistant. Excellent, I've always fancied an assistant. Right, I'm gonna ask it something.
High, deep, seek. Does the Chinese government read everything I ask you to do? Do question mark? It's thinking, no, apparently, no. The Chinese government does not have access to the specific questions or interactions you have with me. Your conversation with AI models, I mean, are private.
This is DeepSeek, an innocuous-looking chat book that this week sent shockwaves far beyond Silicon Valley. Down sharply this morning amid fears of an AI stock bubble burst. It is mind-blowing and it is shaking this entire industry to its core. A major league sell-off in AI stocks and big tech. It's happening now. The Dow is off 112. S&P is down 103 points.
Within just a few days, it had shot to number one in the App Store, knocking rivals like chat GPT and Gemini off their perch, leaving some rather red faces among the tech Broligarchy.
Deepseek claims to be far cheaper and far greener than its US competitors, promising to enhance our lives and make us vastly more efficient. But should we be worried about the Chinese Communist Party, gaining a copy of everything we type into Deepseek and potentially using it for nefarious means? From The Guardian, I'm Helen Pitt. Today in focus. How Deepseek rocked Big Tech?
Robert Booth, welcome back to today in focus. You're a familiar voice to many of our listeners from your previous role as Social Affairs correspondent, but this is your debut as the Guardian's UK technology editor. So congratulations on the new job. We're talking today about deep-seak and it's a product that I confess I had not heard of at all until Monday. What is it?
I'd only heard of it a few days earlier when I was meeting an expert in Chinese tech here in London, and they'd mentioned Deep Seat making a lot of progress, and then suddenly it bubbled up. So it's another AI chatbot. You know, listeners will have heard about chat GPT.
Googles versions of these chat bots, meta versions of chat bots, perplexity, there's lots of them around, most of them are based in California, this one and usually based in China. It seems to be doing pretty much the same thing as all those other chat bots.
It's not any more sophisticated or quick or breaking any kind of records, but what it is doing is the same thing, but using clever engineering to deploy far fewer resources. In other words, it's much, much cheaper to run.
That means it's sort of the biggest news in this space of AI chat box since November 2022 when chat GPT came out and everyone started playing with it and so the same thing's happening again and it's been seen as quite a big game changer. And who is behind Deepseek?
So, yeah, the youngish-looking man, who goes by the name of Liang Wen Feng, he's 40 now. He's a hedge fund entrepreneur from South China's Guangdong Province. He set up a startup in Hangzhou, which is just outside Shanghai.
that's deep seek and that was founded in May 2023 so it's relatively new company and it really has been focused as a sort of research organization rather than a kind of commercial tech company and that may be a key reason why it's been so successful is that it you know he's a very wealthy guy he's got his hedge fund running and this seems to be something of a passion project and so he's been just very focused on developing it and suddenly it's burst out into the world.
And when we think of these kind of tech entrepreneurs, tech geniuses, we've got Elon Musk at one end in terms of publicity, hungry, and then you've got the much more sort of shy, reclusive geniuses on the other end. Where would you say that this chap falls? Well, we don't know too much about him at the moment. And obviously, this news has broken over the Chinese New Year holiday. And he, I think, probably among
hundreds of millions of other Chinese are away. He hasn't been speaking a lot about it. He is clearly a high flyer. His company is called High Flyer Capital. He had previously been using AI to spot patterns in stock prices.
making a lot of money for himself and for his company. But now he's kind of working with a very relatively small team, a few hundred people it seems, although very highly talented data scientists and engineers. And he's done something pretty special. The artificial intelligence arms race as a new competitor out of China, a new AI app made by Chinese tech startup DeepSeek, appears to be better and significantly cheaper.
then it's American counterparts. Right now it sits at the top of the charts on Apple's App Store, despite close ties to the Chinese Communist Party and vast collection of users' personal data.
And have I been very foolish downloading this onto my phone? Can I now assume that the Chinese government, anything I type into there, I guess, they will know? I think that's the key thing that people seem to be warning about is whatever you put into DeepSeek is going to be going on to Chinese servers that held in China. And so people often use chat GPT and these kind of chatbots for work. You know, they might put into them
work documents or emails or reports and then ask them to do some sort of heavy lifting on them. And the danger there is that you're passing potentially sensitive documents over or if you might want to talk about your family or you might want to talk about those sorts of details. That's all going on to Chinese service. And under Chinese law, their own national security law, anybody in China, individuals, companies, enterprises,
have to, in their own words, shall support, assist and cooperate with national intelligence efforts. So it's incredibly broad law and it means essentially that the Chinese state almost certainly can obtain any information that you put into it. And what do we know about how much it costs to build and how they built it?
Well, a lot of the focus over the last week has been on this. And the key figures that are floating around is that it costs less than $6 million to train the model, which is many, many times less than its rivals. But that might be a slightly misleading figure because that doesn't necessarily account for all of the costs. Some estimates suggest that the wider company of the exceed must have spent hundreds of millions of dollars on buying processing power previous to that.
I don't think I would particularly be disingenuous about it. There are six million relates to a sort of final training runners, I understand it. Nevertheless, it does seem to be a lot cheaper, many times cheaper, to have developed this model than the approach that has been taken in Silicon Valley over the last couple of years, where
only kind of last week we heard about a planned investment of half a trillion dollars in computing power and you compare that to the sort of six million figure and you can see there's a complete difference in scale.
So there may be a certain amount of smoke and mirrors here when it comes to the cost claims, and I've definitely heard a lot of skepticism that something so powerful could be done on the cheap like this. And anyway, why have we only heard about DeepSeek now? Given the acres of coverage there have been about AIs like chat GPT or Gemini, which is Google's version. A lot of people can now pay $20 a month for these, don't they?
I think the thing is that both of those that you mentioned have been, you know, got huge marketing budgets behind them because they're pushing really, really hard to try and monetize them to one degree or another. And Deepseek hasn't been like that. Deepseek has been known about by some people in the tech world for a little while.
It wasn't set up as a money-making enterprise, so it hasn't been pushing its product. And then last Wednesday, 22nd of January, it drops this paper, 22 pages, which is explaining that it's launching this thing called DeepSeek R1, which is the thing that we're talking about. It's really not a piece of marketing material. The title of this document is not snappy. It's called incentivizing reasoning capability in LLMs via reinforcement learning.
I started looking out of the window halfway through that sentence. Yeah, exactly. They are not trying to sell it. But it was boasting of how powerful and intriguing this thing seemed to be and that it was comparable to open AIs, the $20 one that we were talking about. And then he had a really big tech-investical Mark Andreson who took to Twitter. He's a Trump supporter as well. And he talked about it as the Sputnik moment, which was a wonderful two-word phrase that captured everyone's imagination. So likening it to the space race.
Yeah, you know, to the Russians getting the satellite up, I think, before America. Deepseak then hits number one in the Apple Store in the US and then in the UK. And then the sort of markets wake up to it and think, goodness me, what's going on? And in terms of the actual technology, what is Deepseak doing differently? What's so cool about it?
So the way to think about this is that it's essentially a piece of engineering. So if you think about it as a sort of physical machine, they've worked out how to make the machine much more efficient. I often think about this moment in AI, a little bit like kind of the early industrial revolution, you're in Manchester in the 1830s and someone has come forward with a much more efficient way of making a mill or a loom work. And it's that kind of a- I'm smiling behind my bonnet.
Exactly, but it's that kind of a shift. It's like goodness, we can do it differently. What they seem to have done is that they've worked out a way of doing more with less and they've got less chips. What they're doing then is activating only the relevant chips when you put your search in rather than all of them. So it's answering more efficiently. It's reducing its requirement for memory.
The other thing it seems to have done is it's trained it in a slightly different way to the ones that we've been talking about. It's managed to turn the kind of large language model that the chat GPTs and so on have mostly previously used into this reasoning model. And that means that it's sort of fact checking itself as it goes along. It takes a little longer to come back to you with the information.
than a non-reasoning model. But the upside is that this needs to be more reliable, particularly in domains like physics and maths and science. You can see the sort of chain of thought that it's taking, and that kind of makes it feel a bit more human as well. So you can sort of engage in a slightly different way. I mean, essentially it's doing pretty much the same thing, but the engineering means it's far, far more efficient.
and cheaper, and why does it matter that it's so much cheaper to run than its rivals? One of the big things here is that you have a kind of oligarchy of AI companies in the world at the moment. A small number who are able to do what they're doing, develop their AI because they've got billions and billions and billions of dollars behind them.
enabling them to buy the huge number of microchips that it was thought to when necessary to create the very best models. Nvidia is the big American company that until this week was, I think, the largest company in the world. And it has been producing huge numbers of these chips for the big American companies. But I think the fact that they can do it much more cheaply, deep-seeker doing it much more cheaply, just reduces the barriers to entry into this market.
So it means that you need far less capital to be able to try and join in. So likely we're going to see lots more competition. And that's one of the reasons why the British government, I Peter Kyle, the tech secretary, has said he's really excited about it, putting to one side the fact that there are security concerns about how it's used. But he's excited about the fact that this has happened because it enables the smaller AI players like Britain, which is kind of normally considered probably the biggest in Europe, but like third in the world.
behind American China. It could enable us to kind of punch harder because we won't need as much investment. Who then are the biggest losers with the advent of deep-seek?
Well, I mean, immediately it's the people who own stocks in several of the big tech companies. So the markets responded to the efficiency of DeepSeek by wiping nearly a trillion dollars from the NASDAQ composite, which is the big US tech index. Nvidia, which is the chip maker we were talking about,
It shares fell by 17% won seven, which is the biggest fall in a US stock market history. That was $600 billion just wiped off. It's come back a bit. Google's parent company, Alphabet, lost $100 billion on that day and then Microsoft.
less, lost 7 billion, but it wasn't just deep-seek that precipitated this. There was a sense among some investors that they were already spending too much capital on chips, and it might not have been absolutely necessary. Those companies are losers. I think there is also something to think about in terms of
the competition between America and China geopolitically because Trump coming into the White House would have felt that US supremacy in AI would have been a useful part of his attempts to do better deals with China. And that calculation probably has to change. And what has he had to say about it?
He said that it's good because you don't have to spend as much money, which is obviously something that any sort of businessman would see. So he said, I view that as a positive asset. But he also said that it should be a wake up call for our industries. The release of DeepSeek AI from a Chinese company should be a wake up call for our industries that we need to be laser focused on competing to win
By not needing as many US chips to create deep-seek, at least that's what it seems on the face of it, it probably reduces the amount of leverage that the US has to use over China in terms of banning the export of chips from America's China. Yeah, because Biden was quite keen on that. He really made it hard to export this. So there have been export bans on chips to China and deep-seek seems to have kind of
just worked with what they had, and also worked with what they could import. So they'd stockpiled something? Yeah, it seems that prior to the export ban, in 2022, 23, they were able to get chips in there that were powerful, and they've also used these less powerful chips, and they've been savvy about it and produced a kind of design breakthroughs.
It's really like kind of, you know, if you imagine a really heavy weight bridge between two places, you know, using tons of concrete and steel and cement. And then a sort of whizzing new architect comes in and suggests, actually, we can do this in a much more lightweight way. And everyone goes, of course, you know, and then it becomes the new normal. And that's maybe what's happened here. And what have Nvidia and OpenAI expressed concerns about?
Well, Nvidia obviously is kind of reeling from the shock of its stock fall, but now the dust is settling slightly. It seems to think that more of its chips will probably be needed in the future to meet the demand that has been created for deep-seek services. OpenAI for its part, it says that
It thinks that there's evidence that Deepsea has used its AI model to train its chatbot to one degree or another. And that's certainly something that the White House has echoed as well. So, I mean, open AI, the clues in the name, or is it not cricket to do that?
Well, I mean, it seems to me that all these technological developments are building on each other, right? And, you know, OpenAI has had to train its own chatbots on large amounts of, you know, material from all around the world. That's been fed every word you and I have ever written for the Guardian, I'm sure.
And there are lots of legal suits in the States where copyright holders are trying to push back against that. But I suspect that it is what it is. It's been created and developers are always building on the shoulders of each other. And so, you know, Sam Altman, who's the founder of OpenAI, is basically saying, look, we've got a new competitor. We're going to step up.
coming up. Deepseek has assured me that it will always tell me the truth. But what if I ask it something about Tiananmen Square?
Rob The Guardian has been reporting this week that deep-seek won't answer difficult questions on matters that are sensitive to the Chinese Communist Party, whether that relates to Tibet or Taiwan or Tiananmen Square. What does that tell us?
But it's really interesting because clearly there are guardrails that Deepseek is putting on what it will actually tell users about very sensitive subjects in China. And one of these cropped up this week when a reader from Mexico got in touch, Guy called Salvador, and he'd asked whether
free speech was a legitimate right in China. And then what he shared was his screen. So he showed us Deepsea kind of giving its preamble about the things that it might consider when it was going to answer this question. And that was things like, and I'm just going to read a persecution of human rights lawyers, crackdowns on protests in Hong Kong, censorship of discussions of reeducation camps, China's social credit system punishing the centers, all these sorts of things.
And then it started to give its answer proper. It was explaining how China's governance model dismisses or rejects the framework of free speech to a large degree and instead prioritizes the state authority. And then it starts to get into kind of much more blunt stuff and saying that in China, the primary threat to this framework of free speech is actually the state itself, which actively suppresses dissent is what it said. And it's at that point that his Salvador screen goes blank.
It's all wiped away and the different messages put in place. And that says, sorry, I'm not sure how to approach this type of question yet, although clearly it was. Let's chat about maths, coding and logic problems instead, exclamation mark. So there it is. That's the censorship and action.
There have been long-running tensions between the US and China over TikTok, another Chinese-owned app because of the security concerns. Are we seeing the same kind of worries about deep-seek and what it's going to do with everybody's personal data?
Well, I mean, it's perhaps less an issue for the government that's talked about in the States because they seem much more worried about kind of competition. But for users, they would want to pay attention to the fact that the Deep Seats privacy policy does state that the personal information that it collects from you is held on secure servers located in the People's Republic of China. So it's clear that
If you're depositing your information, your data on Chinese servers, the Chinese intelligence agencies can potentially get access to it. Right, so there are some very real concerns then, but overall, given that it's cheaper and potentially a lot greener to run if it's more energy efficient, could the rise of deep-seek be good news?
I think it can, from the perspective of the pace of technological development, and if people can work out what it's done in terms of being so much more efficient and it's proved that it is based on lower cost, then that will allow lots more competitors into the market and you could see a second flourishing, if you like, of this kind of technology.
The flip side of that, I suppose, is that it's clear that China and perhaps other states as well are going to have the ability to release these kind of AIs at similar pace to the West, if not faster. And one of these AIs at some point is going to become kind of dominant and influential. And if it's an AI that's operated by China or overseen by the Chinese Communist Party,
or another state that's potentially hostile to the West. It could be quite a powerful agent in the geopolitical tensions that are already existing in the world. So that is another side of it that we need to be very aware of. But I think maybe the final word should go to Sam Altman, the open AI boss, who said he thought it was legit invigorating.
to have a new competitor. And so perhaps the prognosis is that this is going to be good news for consumers, at least in the short run. Well, all this for an app that most of us only heard of on Monday. Rob, thank you very much. Thank you.
That was Rob Booth, the Guardian's new UK technology correspondent. You can read his coverage of Deepsea at TheGuardian.com, along with much more from our whole technology team about the rapid developments in the world of AI. The Guardian is committed to following all of these tech developments as they change our world for better or worse. And keeping tabs on these tech brawler garks takes time and tenacity. If you want to support our work in this field, then please consider supporting TheGuardian.
You can find out how at TheGuardian.com forward slash today in focus pod. And that's all for today. This episode was produced by Lucy Hoff and presented by Mii Helen Pitt. Some design was by Hannah Varal and the executive producer was Elizabeth Cassin. We'll be back on Monday. This is TheGuardian.