Chinese AI Is Spooking Silicon Valley

[MUSIC] >> Live from New York, it's the journal. >> That's right, Jess and I are back with another live show in New York City. This time on Wednesday, September 9th. >> We're going to be at City Winery for an evening of sharp reporting inside our conversation and special guests. >> And probably wine? >> I hope so. Either way, you won't want to miss it. >> Tickets are on sale now. Check out the link in our show notes. See you there. [MUSIC] >> China has made some big advancements in AI recently. >> You break throughs in powerful low cost Chinese AI models are sending shockwaves through Washington and Silicon Valley. >> The latest Chinese offering, seen as a credible challenge. >> It is no longer a given that the US Frontier Labs have a safe lead in AI. >> Big USAI companies are alarmed by this. >> Our colleague, Omrith Romkumar, covers tech policy in AI. >> The view was essentially that the US has a decent size lead. You could say six to nine months, a lot of people argue over China and other countries. >> These models are made by companies in China like Alibaba, Z.ai and Deepseek. >> The most recent model that has everyone's attention is called Kimi K3. It was made by a company called Moonshot AI and it's basically been shown in some benchmarks to get me getting close to advanced US capabilities and that's what turned a lot of heads. [MUSIC] >> It's not just that these models are good. They're also way cheaper than American AI models from companies like OpenAI and Anthropic. >> These things cost a fraction of what companies are currently paying to use in Anthropics, top products and OpenAI's top products and especially at big companies with a large number of employees, they're telling them to use AI as much as possible. But the bills are just racking up and they're quite insane. >> And these new models are really attractive to a lot of US companies. >> We've written a lot about how executives are struggling with that and finance departments are sort of fretting. It's millions and millions or tens of millions of dollars spent on AI compute. If you can cut that bill in half or 75% and get something close to the same performance. I mean, that's the no-brainer for a lot of executives. >> But when it comes to the US companies that make AI products like OpenAI and Anthropic, they have a lot to be worried about. Some AI researchers and analysts have questioned whether this could be the beginning of the end of America's dominance in AI race. You have Anthropic and OpenAI barreling toward IPOs with valuations around a trillion dollars or other really high figures and the notion that China could come out with more products to undercut those businesses is really scary for a lot of people. >> Welcome to the journal, our show about money, business, and power. I'm Ryan Knudsen, it's Tuesday, July 20th. Coming up on the show, who's afraid of Chinese AI? >> This episode of the journal is presented to you by Accenture. Behind this podcast and the music you listen to, Spotify has a sophisticated advertising business. Accenture is deeply embedded in that business, working with Spotify to reinvent the rhythm of ad sales, building, maintaining, and optimizing. Accenture finds the friction and builds the systems that make things run smoother and faster. The result, less time spent on operations, more time connecting brands with the moments and fandoms that matter most, learn more at Accenture.com/Splotify. >> This episode is brought to you by Indeed. Running a business means dealing with a dreaded hiring process, and the stakes are high when the right hire could make or break your business strategy. But good news, Indeed sponsor jobs is here to make things less painful and nerve-wracking. It gives your job post a boost to reach quality talent faster. Plus, it can help put your job in front of exactly who you're looking for based on criteria you set. Like if you need certain skills or certifications. And it works. Sponsored jobs posted directly on Indeed are 95% more likely to report a hire than non-sponsored jobs. So spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a $75 sponsor job credit at Indeed.com/podcast. That's Indeed.com/podcast. Terms and conditions apply. Need a hiring hero? This is a job for Indeed sponsor jobs. >> All right, so these new Chinese AI models are much cheaper and have apparently gotten a lot better. Talk about how they did that starting with how have they gotten so much better? >> China has been thought to be several months to a year behind and one way they've been trying to catch up is this technique called distillation where you essentially train off another company's AI models. >> China does an enormous amount of its own innovation and research on AI. But some Chinese companies have also been accused of distilling US models, basically studying how they respond to prompts and then working backward. Kind of like trying to reverse engineer all the secret ingredients in a can of coke. >> With AI and how sophisticated these techniques can get, it's like just imagine millions and millions of cans of coke and pouring them out, seeing what's in there and trying to mimic those. And a lot of people have said if you go and use these Chinese models, the language they use when they respond will be similar to what is used in Anthropics latest models, for example. So that is definitely part of the equation and the stunning thing to a lot of people is that China is doing this, even while the US has restrictions on what advanced chips it can get. >> US companies make the world's most powerful chips, which are essential to developing advanced AI, and the US has made it hard for Chinese companies to get them. >> How is it that China is able to catch up if they're still being restricted on what US chips they can use? >> I mean that's really the billion dollar question that people continue to ask. One answer, maybe a tech loophole called remote access. >> And that is when user basically can access data center space and chips in another country. So Chinese companies are essentially able to use remote access to train their models on chips that might not be located in China. They might be in Singapore, Thailand, et cetera. Frankly, there's not a ton known. This is sort of like the secret sauce for why these Chinese companies are so successful. The combination of distillation and finding new workarounds to get more essentially out of less powerful chips has allowed the Chinese companies to not quite match the US capabilities, but they're pretty close. Chinese authorities have denied that its companies have used unauthorized distillation techniques to copy US models, and have said the allegations are politically motivated. Now let's talk about why China's AI models are cheaper. First, Chinese AI companies are often subsidized by the Chinese government and use different training techniques to cut costs. The other thing is that most of their models are what's known as open weight. Open weight means that the parameters that make up the AI model are public and can be shared and adapted. You might be more familiar with the phrase open source, which is a pretty similar idea. >> Open weight tools are very significant because the open source technology generally to zoom out is the backbone of our modern internet. It's what's allowed Silicon Valley to thrive over the past several decades. It's how many companies have been built essentially. >> Open weight AI models give the companies that use them a lot more control. >> It really spurs a lot of adoption among businesses because they want to take these AI tools and plug them into their own products, use their own proprietary data, and that's much easier to do with open weight products than closed products like those offered by Anthropic and open AI. One of the most popular models from U.S. companies, like the ones that power chat U.P.T. or cloud, are not open weight. Customers can't see or control the underlying code. >> It's sort of like Android versus iOS on Apple that Android can run on any smartphone and smartphone makers can put their own spin on it and customize and change it in any number of ways where if you want iOS, it's just on Apple devices. >> That's a really good way to think about it. It's very easy to just take the model, take how it was built, and integrate that in your own organization in ways you can't really do with Anthropic and OpenAI. You're basically captive to sort of their tools, their interface, and it's a much tougher process and much more expensive. >> Their business model is also much different from companies like Anthropic or OpenAI. For Chinese Openweight AI companies, the hope is that mass adoption will bring in revenue down the road through upsells or additional services. For a lot of businesses, Openweight systems are much more appealing. >> There's a whole other swath of Silicon Valley and smaller startups that are saying this is great. They're a cheaper AI tools on the market. I can go in. I can download the parameters. I can modify them. This will help my business. It's cheaper. It's more accessible than Anthropic's cloud, OpenAI's models, et cetera. >> For all these reasons, some American businesses, like Airbnb and Coinbase, have already ditched USAI models for their Chinese competitors. >> If this Chinese open source model succeeds, how bad could it be for companies like OpenAI and Anthropic? There was a worst case scenario. >> The worst case scenario for these companies is that this is just the beginning, that these companies have to really cut prices, that cuts into their profit margins. They're still spending a ton of money. Investors. and then you're in sort of a potential, that's spiral. So few people expect that right now, but that's what the fear that keeps people up at night and what they're thinking in the back of their mind every time, China comes out with one of these great models. (upbeat music) - How USAI companies are responding, that's next. (upbeat music) - People disagree about a lot in this country. Taxes, healthcare, education. But wherever you land politically, I think most of us agree on this. Better information leads to better decisions. That's the mission of USAFacts to make government data clear and accessible. Because public data should be easy to access, trust, and understand so we can use it to make informed decisions. Read and sign our open letter to support public data at USAFacts.org. A.I. is moving fast across the enterprise. But without visibility, it's just chaos, different tools, different models, different teams using A.I. in completely different ways. ServiceNow turns that chaos into control. With the A.I. control tower, you see all your A.I. across the business in one place. What it's doing, what it's done, and what it's about to do. So you stay in control. To put A.I. to work for people, visit serviceNow.com. (classical music) - Earlier this month, an executive at Open A.I. named Dean Ball started sharing his thoughts about Chinese open-weight models. - Dean Ball is a former Trump administration official who was in the administration for a very short time last year. And he recently started at Open A.I. in a position focused on essentially policy strategy. - In this position, Ball said that when he comments on A.I. policy, he's speaking for himself, not Open A.I. - But he said on X After a Kimi K3 came out that he was worried that Chinese A.I. models were essentially going to take over and that there would essentially be this sort of dystopian hellscape where communism ran the A.I. world. - In Ball's view, if China wins the A.I. race and open-weight models come to dominate, A.I. could become something provided by governments rather than sold by companies. - The idea is that if China basically gains control of the A.I. race and that it's open-weight models are the ones that proliferate widely, essentially the companies lose a lot of control over that. And because the Chinese government has a role in how Chinese companies are developing these things, that essentially it becomes a utilitarian product that's available to everyone as opposed to premium product, that's reserved for, in the most powerful cases, the U.S. government and powerful companies. - The worry is that this won't be an industry anymore, it won't be a business, it'll just be like this service that the Chinese government provides to the world. - Or other governments, even the U.S., if the U.S. goes that way and it's all Open models. - David Sachs, a top white house official on A.I. policy, criticized Ball's take and said it sounded like he was just trying to protect Open A.I. business. - Writing quote, the weaponization of regulatory uncertainty as a competitive tool should be completely unacceptable. More broadly, there are other concerns about these Chinese A.I. companies, like political censorship. - So there's been testing by cyber security researchers and others that's basically shown. The Chinese models do have a bit of a bias if you ask them about something like Tiananmen Square. They're still going to to the party Y in there. - They're like, "I'm not sure what you're talking about." (laughing) - Exactly. - Interesting. - So there is a bit of bias there. - Chinese open weight models also have fewer systems in place to prevent bad actors from using A.I. to cause harm. - They also don't have the same safeguards and so that's why people are so alarmed by this. They don't have to go through the same safety testing and government oversight that the U.S. has recently started to do and that other advanced economies are doing. - Where there's limits on, "Hey, can you teach me how to make a bomb?" or whatever, the U.S. models will be like, "No." - Exactly. - Or they should say no. - Exactly. The U.S. companies say they're increasing their safeguards and the government has more of a role there. And then there's national security. - As with any Chinese technology, there are huge security concerns here, especially for a lot of lawmakers and administration officials others in Washington. These models are open weight so that does give you more insight into how they are built, how they work, et cetera. But there are still huge data privacy concerns and a fear that the Chinese government could somehow have access or influence into how this all works. - Some A.I. executives have asked the White House to consider stepping in to take a clearer stand against Chinese A.I. companies or even ban Chinese A.I. models entirely. The Trump administration has said that it's considering what to do about it. In recent days, we've seen several administration officials come out with statements like this is very troubling. We're studying what's going on. - But then late last week, an open letter about these open weight models shook things up even more. It was written by Jensen Huang, the CEO of NVIDIA, which designs many of the semiconductor chips used to power A.I. The letter took a strong stance defending open weight models. - Saying that they're good for Silicon Valley and good for America and the rest of the world. And they're urging the government not to take any action essentially restricting these models from China or anywhere else because they think they benefit innovation. - Just a few days ago, Jensen Huang talked about his position with great A.I. open models. It's great for the whole industry. Obviously, when there's great A.I., even if it's open, wherever it comes from, there will be more use. Huang wrote in the letter that open models strengthen safety and cyber security and give companies that use open weight A.I. tools more freedom. - These Chinese models are excellent. Open source models that are excellent should be used. - In the following hours and days, other tech company leaders signed on. - We've basically had this onslaught where top CEOs, including NVIDIA's Jensen Huang, Metta's Mark Zuckerberg, and many other tech luminaries have planted a flag essentially defending open weight models. NVIDIA's role in all of this is fascinating. They're the biggest trip designer in the world. They supply all these customers. Their interest is to make the ecosystem as diverse and robust as possible. So they sell as many chips as possible. - So what kind of pressure did that put on open A.I. in Anthropic, who are kind of with his letters about? - The letter created enormous pressure on others in the tech ecosystem. You had Metta, you had NVIDIA, and you also had Microsoft right out of the gate three of the biggest players in this space. And then that sort of shot the spotlight on Anthropic, Open A.I. in Google. - By the end of the weekend, Open A.I. in Google had endorsed the letter. - Both of those companies also develop open models. And again, the letter is framed in a way that's basically like open weight models are great. Let's sing Kumbaya. It's like easy to sort of get behind that, even if you have some security concerns and privately you're talking to the government about what can be done. - Anthropic is the one major A.I. company that's yet to endorse the letter. Yesterday, CEO Dario Amade published a response. He wrote that he agrees with many of the points in the open letter, but disagrees that broad access to open weights makes the ecosystem safer. - They've made clear they see risks with open source and from a competitive standpoint, they don't have open models and they're pretty much against them. But now they are standing against essentially big swats of the White House, big swats of Silicon Valley, pretty much the rest of the A.I. ecosystem in the US. - I mean, it seems like, you know, one of the concerns about the race over A.I. between the US and China is, the China is like really leading away on these open weight models. That creates a concern. But with this letter, it feels like almost all of Silicon Valley, all of the tech world, other than just one giant, are also endorsing that open weight path. So what is that, what do you think that's going to mean for the future of the way this technology develops? - The race is definitely entering a new chapter and China's ability to have government support behind all of this and long-term policy clarity is really coming into focus for a lot of people. I think they'll continue to be a roller coaster here where when the open models have breakthroughs, it stirs up this debate again about whether there need to be restrictions on the value of open weight and how much China is catching up. And then in a matter of weeks or months, you'll see probably more advances from open A.I. andthropic and others when the pendulum might swing back a bit. So I think it's just hard for people to feel like they're on stable ground in the A.I. race. (upbeat music) (upbeat music) - That's all for today, Tuesday, July 28th. The journal is a co-production of Spotify and the Wall Street Journal. Additional reporting in the episode by Jonathan Chang, Rafael Huang, Tina Lee and Katrina Northrop. Thanks for listening, see you tomorrow. - Hank joined B.J.'s wholesale club the day he became a father of 30. - I coach football. now. Dr. Chang saves up to 25% off grocery store prices, 30 pounds of pasta, three cases of protein bars, 75 sports drinks, and that's just pregame. He knows teamwork, and BJs knows savings. This is your home coach, home of the save. Join for just $20 at bj's.com/maskete and save 10 cents per gallon for six months, open soon limited time offer new members only. BJs, home of the save. When you shop Kroger Delivery, you can expect the savings you love and fresh groceries delivered right to your door. Our refrigerated trucks keep your items fresh while our associates are committed to getting every detail right, carefully selecting and packing your order with care, because bringing you fresh, quality groceries is what we do best. And right now, enjoy $20 off your first online order of $75 or more. Restriction supply, seasite for details. Kroger, fresh for everyone.