The Daily - China Challenges Silicon Valley for A.I. Dominance

Episode Date: February 3, 2025

Financial markets went into a panic last week over an obscure Chinese tech start-up called DeepSeek. The company now threatens to upend the world of artificial intelligence and the race for who will dominate it.Kevin Roose, a tech columnist at The Times, discusses how DeepSeek caught us all off guard.Guests: Kevin Roose, a technology columnist for The New York Times and co-host of the Times tech podcast, “Hard Fork.”Background reading: DeepSeek’s model has rocked Silicon Valley and upended several fundamental assumptions about A.I. progress.Listen to “Hard Fork”: Your guide to the DeepSeek freakout.For more information on today’s episode, visit nytimes.com/thedaily. Transcripts of each episode will be made available by the next workday. Photo: Greg Baker/Agence France-Presse Unlock full access to New York Times podcasts and explore everything from politics to pop culture. Subscribe today at nytimes.com/podcasts or on Apple Podcasts and Spotify.

From the New York Times, I'm Natalie Kittroweth. This is The Daily. Last week, financial markets went into a panic over an obscure Chinese tech startup called DeepSeek. That company now threatens to upend the world of artificial intelligence and the race for who will dominate it. company now threatens to upend the world of artificial intelligence and the race for who will dominate it. Today, my colleague Kevin Roos, a tech columnist and the co-host of the podcast Hard Fork,

on How Deep Seek caught us all off guard. It's Monday, February 3rd. Hi, Kevin. Hello. So great to be here. So let's jump in. Kevin, how did this giant AI tech freakout begin? Tell us that story. So the freakout really started in earnest

with a Chinese AI company called DeepSeek. And DeepSeek had released a new AI model. Models are released all the time. Generally, they don't make international news. But this model was different in a few ways. DeepSeek released its new chatbot app, which is said to perform as well as chat GPT. One of them was that it just appeared to be a really good model,

like better than the leading Chinese models at the time and on par or close to on par with the leading American models. A new AI king was crowned today, well at least for now. And so the DeepSeek app. Unseating open AI's chat GPT. Goes to number one on the app store charts. It faults ahead of chat GPT in all these better known apps. Now the most stunning thing here isn't necessarily

that China has developed a pretty good AI app. It's how cheap it is. And the more notable thing, the thing that really caused the American AI industry to start to panic, was how cheaply this model appeared to have been built. How cheap are we talking? So we think that this model costs DeepSeek about $5.5 million to train.

Now that might sound like a lot of money, but it's really not compared to what many of the American AI companies are spending. Meta said that it was spending $65 billion. Microsoft says it plans to spend about $80 billion. Microsoft said it was spending $65 billion. Microsoft says it plans to spend about $80 billion. Microsoft said it was spending $80 billion. And some of the world's most prominent technology leaders pledging to invest an initial $100 billion. And OpenAI had just announced this giant partnership where they were planning to spend up to half a trillion over the next four years.

As much as half a trillion dollars to build the infrastructure for AI. Whoa. Yeah, it's wild. And on top of that, DeepSeek says that they built their model without access to the latest and greatest American AI chips, which up until now were thought to be necessary to build the most powerful models.

A lot of people put a lot of money into AI, and now they're wondering if that money is needed the way that some of these American companies have said it is. And so investors start saying, wait a minute, if it only costs $5.5 million to train a leading edge AI model, then what the heck are all these American companies doing spending hundreds of millions of dollars or even billions of dollars to train roughly equivalent models? Tech stocks plummeted Monday as investors raised concerns about advancements in Chinese artificial intelligence.

And so the stocks of many of the American tech companies start to fall. Right. And so after all of this, people in the American tech industry start asking questions. Like, who is DeepSeq? And how are they getting these incredible models with so little money spent on them? little money spent on them. Okay, we're going to get to those questions of who this company is and how they did this. But I just first want to dig into the anatomy of the market panic.

What are the real fears driving this? So it depends who you ask, because there are a couple kind of overlapping panics that are starting to happen around this time Of course again, there's the investor panic I mean imagine if you had your whole portfolio invested in American AI companies it would be like if you just bought like a very high-end sports car like a Lamborghini and You had been driving it around and were so proud of how fast it could accelerate and how well it handled. And then like some random guy shows up with like a soapbox car made of balsa wood

and it can go just as fast as your car. You'd be like, what the heck? Why did I just spend all this money on this Lamborghini? Yeah, and should I maybe be investing in balsa wood cars? Yes. And then, of course, there's the geopolitical freakout because DeepSeek is a Chinese AI company. And there has been this race happening between primarily the US and China for years about

AI and AI supremacy. Who was going to be able to build the most powerful AI models before the other one. And that is a very important question for things like assessing the future of military conflict. If one country's AI is way better than another country's AI, they might have an advantage. In fact, the US has banned the export of the most powerful AI chips to China for exactly this reason, to try to hobble the Chinese AI companies to keep them from catching up when it comes to building

the bleeding edge models that could become very important. So instead, DeepSeek had to make do with these Kirkland signature chips that are pretty good, but they're not the best. And so that combined with the amount of money spent really made people say, how do they pull this thing off? Kevin, it certainly seems that at least based on what DeepSeek is saying, it has managed to pull off a pretty impressive feat here. But I'm wondering, can we trust what the company says? Can we trust their claims about how they pulled this off? Yeah. So there are a lot of people who are skeptical of what DeepSeek has claimed. In

particular, the cost of the model, $5.5 million, might not be the real figure. It doesn't include all of the research and the engineer salaries and things that went into that so that the real cost is probably Significantly higher than that but there are questions about you know Did they smuggle in very powerful chips that would have actually allowed them to build a model this good? You know, is there something going on? Is the Chinese government funneling money to them and not telling us about it? So there are lots of theories. But then as time wears on and people who are experts in this stuff start digging through the details, they're coming

to the conclusion that, well, yeah, maybe the cost is a little higher than Deepsea claims. Maybe they have a few more chips than they're telling us about. But in general, this seems like they actually just did build a really good model using some very clever engineering techniques. Okay, so let's talk about those engineering techniques. I mean, how actually did Deep Seek do this? Make a chatbot on a shoestring budget potentially with second-rate chips? So because Deep Seek did not have access, we don't think to the most powerful chips that American companies are using,

they had to get clever about becoming more efficient with their model. I won't bore you with the technical details, it includes terms like mixture of experts architecture. But basically, they were able to use some clever tricks to squeeze the most power out of the chips that they did have. And it occurs to me, Kevin, that this company was operating under a lot of constraints.

And it sounds like that may have forced the engineers to think about how to tackle this problem differently. As in, it seems possible that not having these critical ingredients actually bred innovation. Yeah, I mean, there's this saying in the tech industry that constraints inspire creativity. And that is definitely true here. DeepSeek did not have access to the best American AI chips. They did not have the largest budget or the most sophisticated team. But they were really scrappy and smart. They had a lot of really good young engineers, and

they were able to pull this off. So Kevin, how do the big American tech companies contend with that? I mean, what do they say to investors who are wondering about whether maybe these companies have been throwing money away when some of this work on these AI models could have been done much more cheaply? So what the AI companies in America are saying in response to this market panic is, look, we've still got to build these big, expensive supercomputers to stay at the forefront of AI, to have the best models.

And if we take the techniques that DeepSeek has now shown are possible, these efficiency gains, we could have them too. Think about how powerful our models would be if we put a billion dollars into the same kind of model that DeepSeek was able to make for much less. So that is what the American AI companies are saying. But I think there are real questions among investors about whether the scale of investment

that they have been planning is really necessary. And for you, Kevin, I mean, obviously you've been covering this world for a very long time. Does that show you that more money doesn't necessarily mean more innovation in the world of AI where more money has been kind of a given, you know, as an assumption of what's needed. I mean, does it actually suggest that maybe smaller, as you said, scrappier startups could make huge gains in this world?

Yeah, I think it threw into question this fundamental assumption that only the big dogs could play in AI, right? You had to be Microsoft or Amazon or Google if you wanted a chance to build the state of the art AI models. And I think what the DeepSeek story suggested is that there may be a whole other world of competitors out there trying to stay close to the frontier and that they might not have to have the resources of one of the world's largest corporations to do it.

But there was one other piece of this that I think really suggests that the AI race has entered a new phase, which is that DeepSeek did something that a lot of American companies have been hesitant to do, which is that they released their AI models as open source software, meaning that anyone on the internet can download and use, can make their own versions of, can adapt, can tweak. It is software that can be reused and remixed and improved upon by anyone. And so when DeepSeek released its models this way, they really sent a message to the world that says, we are serious here about competing and we're so serious that we're going to give away our models for free so that anyone who wants to can make them better. And so all of a sudden it just flipped the entire

AI race onto its head and really sent it into a new gear. We'll be right back. Kevin, it sounds like DeepSeek has already or is about to really change the landscape of AI. And my question is, is that good? Like for people like you and me, I mean, maybe you more than me who use chatbots for consumers? So it's a complicated question to answer because there are ways in which it is probably good and ways in which it is probably bad.

The case that this is a good thing is that in general when you make things cheaper they can be accessible to more people. I mean remember DeepSeek is not just free to use in the app or on the website. It was also released as open source software, meaning that anyone with an internet connection can download it and install it on their own computers or maybe tweak it to serve their own purposes. So if you are a person who wants to use AI, maybe you have a small business or maybe you just want to use this to help you write letters or emails.

Maybe you're a student who wants to use this. You can now access a very powerful model for free. Maybe you are a developer or a startup who wants to build your own AI tools. Well, now you have this DeepSeq model that you can kind of take off the shelf as open source software and build your own version of it or run it on your own hardware. And so the people that I talk to in the tech industry who are at startups or smaller companies are very excited about this.

This is a great development for them. And it also means that if you are a person who worries that all the AI power is going to go to a few huge companies, then the democratization of AI through open source models like DeepSeek probably makes you feel optimistic. So that is the positive case for this. But there are also a lot of people who are really worried about what DeepSeek has done. I think the DeepSeek moment has really sparked

a lot of new fears about how quickly this whole field of AI is progressing. I mean, just in the last few years, the leading AI models have gone from maybe being as smart as the average high school student to as smart as a college student to now being able to complete a lot of tasks that would have taken a PhD to complete. So these models are just getting much better very quickly. And a lot of folks in the AI community are just nervous about that. They say things like, well, maybe we're going to get an AI that is as smart as the smartest humans in just a few years.

And we don't really have a playbook for dealing with technology that is more intelligent than us. And so there are people who worry about these sort of runaway AI scenarios where you get super intelligent AIs that can sort of take control or maybe even harm humans. But even if you're not a believer in that superhuman intelligence risk, there are just a lot of questions about whether we as a society are ready for advanced AI.

Are we ready for the possibility that it could eliminate jobs? Are we ready for the possibility that it could really lead to a proliferation of misinformation or propaganda or even automated cyber attacks and things like that. lead to a proliferation of misinformation or propaganda or even automated cyber attacks and things like that. So all of that is swirling around the conversation about DeepSeek because we have just accelerated the AI race again and now it is not just American companies competing with each other over who is in the lead of that race. China has also stepped in and there's a lot of fear and anxiety about what happens if we fall behind.

Kevin, if this really is that important of a moment in the global AI arms race, how should we expect the United States to react? I mean, does the US government just ban this thing? You know, we saw TikTok banned because it was owned by a Chinese company. Is that the move here? Well, it may well be because if you're a person who believes that TikTok is a national security threat, there's nothing about DeepSeek that is less of a threat, right? It is a Chinese company. It is subject to all the same laws and censorship guidelines as other Chinese

software companies are. So for example, if you ask DeepSeek to tell you what happened at Tiananmen Square or to say something mean about Xi Jinping, the leader of China, it won't do it. And I would not be surprised if in the coming weeks and months, we do see lawmakers in the US saying, wait a minute, we passed a law to ban TikTok, why are we not also passing a law to ban DeepSeek? So I think that's one potential outcome here, but there's a key difference, which is that TikTok is not open source software. You cannot download TikTok and create your own version of it. And so already, the DeepSeek models have been downloaded and recreated all over the world by lots and lots

of different people and companies. And so I think what the DeepSeek story suggests is that it is going to be quite challenging to contain the spread of powerful AI without some big moves. Kevin, if we really are past that point of containment here, if we're off to the races, does it matter that this innovation happened in China by a Chinese company? I mean, isn't this bigger than that at this point? So there are people in the American tech scene who are calling this DeepSeek moment the Sputnik moment for the AI race, because just as the launch of Sputnik by the Soviet Union kicked off the 20th century space race

and created profound fear and anxiety among Western nations about whether they were behind their biggest political adversary when it came to technology, a lot of people are looking at this moment with DeepSeek as kicking off a new era in the AI race when it came to technology, a lot of people are looking at this moment with DeepSeek as kicking off a new era in the AI race where we really want to stay ahead of China. And there are people who say that having a lead in AI, if you are the United States, even if it's just a lead of a couple months or a couple years over your political adversaries is very important.

And that may be true, but people who study AI, people who look at this industry closely, who are paying attention to the trends in AI, believed that these models would become cheaper and cheaper over time, as well as becoming more powerful over time. So this really fits neatly with a lot of what people had been predicting for years. Now, they might not have predicted that this moment would happen from a Chinese AI company. They might not have predicted exactly what the breakthroughs would be that allowed for the models to get cheaper.

But this is in keeping with the overall trend in AI that we've seen over the past few years, which is that the models keep getting better and they keep getting more efficient. In a way, it follows the normal progression of any new product. At first, it's expensive and then more and more competition leads to innovation, the thing gets cheaper, everything becomes more democratized. Yeah. So that is what happened here, but just maybe a little faster than people had expected.

But I think the larger point is that these systems are now improving so rapidly and in so many places all at once that I think it is only a matter of time before nearly everyone in the world has access to very powerful AI models. And I just think that world looks a lot different than the one we live in today. Kevin, thank you so much. Thanks for having me. We'll be right back. Here's what else you need to know today. On Saturday, President Trump declared tariffs of 25% on all goods from Canada and Mexico,

with a partial carve-out for Canadian energy and oil exports. He also announced an additional 10% tariff on products coming in from China. The tariffs are set to take effect at 12.01 a.m. Eastern Time on Tuesday and have raised concerns of an escalating trade war with America's largest trading partners. Those countries responded swiftly to Trump's announcement. The Canadian government detailed its own retaliatory 25% tariffs on more than $100 billion worth of U.S. goods, including everything from American-made honey, tomatoes, and whiskey, to refrigerators and toilets. President Claudia Sheinbaum of Mexico said she would unveil her country's response on Monday, while China's Commerce Ministry promised to bring a legal case at the World Trade Organization and quote, corresponding countermeasures.

President Trump acknowledged on social media that his tariffs could cause quote, some pain, but said it would be quote, worth the price that must be paid. And in a new effort to tighten the immigration system, the Trump administration is ending temporary protected status for 300,000 Venezuelan migrants in the United States, according to government documents obtained by the Times. The move could make those migrants vulnerable

to potential deportation in the coming months. The designation of temporary protected status was meant to protect migrants fleeing dangerous situations in their home countries. But Republican critics have said it allowed migrants to stay in the United States for much longer than intended. Today's episode was produced by Shannon Lin, Claire Tenesgetter, Alex Stern, and Kaitlyn O'Keefe with help from Will Reed. It was edited by Mark George and Chris Haxel with help from Paige Cowitt, contains original music by Pat McCusker, Marian Lozano, and Dan Powell, and was engineered

by Alyssa Moxley. Our theme music is by Jim Brunberg and Ben Landsberg of Wonderly. That's it for the daily. I'm Natalie Ketroff. See you tomorrow. 1