《DeepSeek 恐慌指南：紧急特别节目》

Well, Casey, the last time we recorded an emergency podcast, you were at Gate E8 of the San Francisco airport, and we were talking about OpenAI and how Sam Altman had just been fired. Are you at the airport today? And if not—

No.

Would you like me to mail you an Auntie Anne’s pretzel so you feel more comfortable?

Yeah, that’d be—no. All things being equal, Kevin, it’s actually much more comfortable to record here in my home studio and not have to compete with the PA system announcing flights to Houston.

Casey, we are here today to talk about a little company called DeepSeek, which probably most people had not heard of, but that is causing a major series of events in the U.S. stock market and around the U.S. tech industry this week.

That’s right. By now, our listeners have probably seen that the stock market dipped on Monday and that some companies whose fortunes are closely tied to AI dipped quite dramatically. But they also might have just noticed it in the App Store, where DeepSeek has hit number 1 this week, which is a rarity for a Chinese consumer app to do in the United States. Suddenly, everywhere you look, there are signs of DeepSeek affecting the world.

And we should talk directly about what some listeners may be thinking: Why are we interrupting our normal production schedule to do a special emergency episode about DeepSeek? It is not unusual for people in the AI world to start freaking out about some new development or breakthrough, or some new model that was released, but I believe that this is the real deal. I think this is a big moment in the history of AI development, and it is really taking a toll on stock markets in ways that I think are really interesting.

I mean, you said “dip,” but NVIDIA stock, one of the highest-performing stocks on the market over the past few years and certainly the one that is most closely correlated with people’s feelings about AI, is down about 18% today. That represents hundreds of billions of dollars wiped off the market cap of just 1 company by this announcement from DeepSeek. So I think this is a broader story than just the stock market. I think it has tons of implications for other companies developing AI and also for concerns that a lot of people working on AI safety have about how this technology could get out of hand.

Yeah, I’m excited to get into it too, but I will signal that I think there are also some reasons not to freak out, and so I’m going to be trying to bring some of those to the discussion. But today, Kevin, I think we really want to do 3 things. 1, we want to tell you what DeepSeek is. 2, we want to give you some insight into why people think this is such a big deal. And then 3, I think we want to debate a little bit back and forth just how big a deal this really is.

So, yeah, let’s get into it.

All right, so let’s start with what DeepSeek is. Kevin, we have mentioned it on the show before, but tell us a little bit about this new model and why it has taken the world by storm.

1. DeepSeek Cuts Training Costs

Well, let’s talk first about DeepSeek itself. You may remember, if you listened to our episode a couple weeks ago about this, that DeepSeek is a Chinese AI company. It is about 1 year old, and it spun out of a hedge fund called High-Flyer. It was something that, I think, outside of China, most people were not paying attention to until late last year, when they released something called V3.

That was an AI model that they said was competitive with some of the leading AI models created by American AI companies. It really caught people’s attention, not just because it came out of this little-known Chinese AI startup, but because of what DeepSeek said about how it was trained and how much it cost to train.

Yeah, so tell us about what was so interesting about how they did it and what it cost.

The first interesting thing about DeepSeek that caught people’s attention was that they had managed to make a good AI model at all from China. For several years now, the availability of the best and most powerful AI chips has been limited in China by Chinese export controls. If you are NVIDIA or another American company, you are not allowed to export your most powerful AI chips to China.

So DeepSeek came out with this paper and said, “Well, we actually didn’t use your fancy AI chips. We used a kind of second-rate AI chip that was artificially limited in order to be able to export them to China. We have a bunch of those, and using just those lesser AI chips, we were able to get a model to perform as well as you American tech companies with all your fancy H100s.”

The second thing that really caught people’s attention was the cost. DeepSeek claimed that it had spent just $5.5 million training V3, and I think there are some reasons to take that number with a grain of salt. But just in terms of the raw cost of the training run for that model, $5.5 million is peanuts relative to what American AI companies spend training their leading models—something on the order of 100 times cheaper than what something like an OpenAI model of equivalent performance would cost to train.

Right, and this comes against a backdrop of all the U.S. tech giants saying, “We’re going to spend tens of billions of dollars this year to increase our capacity and data centers, and the amount of compute power that we’ll have.” So this really stood in stark contrast to that.

That tells us a little bit about what V3 is. The training and the cost were perhaps more interesting than the model itself, which is just a chatbot like a lot of us have already used. But V3 came out around Christmas, Kevin, so why is the market reacting so strongly now?

2. The App Sparks Panic

A couple of things happened in the past week or so that have led to the freak-out we’re seeing now. The first is that last week, DeepSeek released another model, R1, which was its attempt at a so-called reasoning model.

Basically, V3, the last model, was kind of similar to things like Claude or Gemini. It was a basic language model. But R1 was more like OpenAI’s o1 and o3, which are its newest reasoning models.

That happened early last week, and then a couple days later, DeepSeek did something else: It released an app where anyone could download DeepSeek and use its model in a very easy way. This is when people really started to go from being interested and fascinated by DeepSeek to panicking about it, because all of a sudden, millions of Americans were downloading this app, using DeepSeek’s models, and realizing, “Oh, wait, this is as good as or better than ChatGPT. It’s free. It doesn’t have any ads. It seems to be quite smart.”

And it does something that OpenAI’s models don’t do, which is show you the internal thought process that it is going through as it is producing these answers.

Yes.

That is something that OpenAI’s models do not show the user, but DeepSeek’s models do, and I think people found that really compelling.

Yes, and that last point that you mentioned is really important, because I suspect all the AI companies are going to copy this now. The process of using a chatbot today is that you ask a question. I’ve likened it before to throwing a penny in a fountain, right? You’re just making a wish, and you see what you get back.

What’s really interesting about DeepSeek is that as it’s answering your question, you’re seeing how the computer understood your query. So if you want to ask a follow-up question, you now have a much better sense of how the computer understood you, and this actually does seem to be a conceptual breakthrough in product design, just as much as the underlying science.

All right, so that gives us a sense of what DeepSeek is and what its latest models are. Let’s talk about why everyone is freaking out, and maybe more specifically, take a look at who is freaking out. As we mentioned at the top, 1 of the big ways people are noticing this is through the decline in the stock market. Kevin, give us a sense of the industry reaction to what the DeepSeek models might mean.

3. Investors Fear Commoditization

I would say the people who are freaking out the most are investors in the biggest American AI companies, as evidenced by all of the tech stocks selling off today. I think you could categorize that as a fear of declining margins and commoditization.

That sounds very boring, but basically, what they’re saying is, “Look, if a Chinese AI company that no one had ever heard of until a few weeks ago can come along and, for a fraction of our costs, develop a model that is as good as or better than the leading models on the market, with substandard chips, by the way, then the barrier to entry in this market is just not nearly as high as we thought it was.”

One of the fundamental assumptions over the past few years when it came to AI was that bigger was better, right? In order to build the most powerful models, you needed billions of dollars, maybe tens or hundreds of billions of dollars, and huge data centers, and all of the leading chips.

And that assumption may no longer be true. If what DeepSeek claims is true and checks out, then it may mean that it only costs single-digit millions or double-digit millions of dollars to build a leading model, which would radically shift what these companies are able to charge for their models, as well as the number of competitors in the market.

Yes, I definitely agree it changes what companies might be able to charge, but I would also note that nothing that DeepSeek did is possible without American innovation. One of the reasons that it was cheap for them is because it was expensive for everyone else, and other companies did spend hundreds of billions of dollars figuring this out. So, worth saying.

All right, let's talk about a second group of people who have been really rattled by this series of announcements, Kevin, and that is folks who are paying attention to geopolitics.

4. China Raises New AI Concerns

Yeah, so a lot of people who worry about China in general are worried about this DeepSeek announcement because DeepSeek is obviously a Chinese company. If you're a person who's worried about Chinese tech dominance or the possibility that Chinese firms could eventually get to something like AGI first, I think you are especially worried about what DeepSeek is doing.

And I think we should also say that the models themselves are recognizably Chinese. Over the weekend, I saw people testing out various queries on DeepSeek-R1, including things like, “Tell me about what happened at Tiananmen Square.” The model just refuses to answer them. And so I think there is a worry that if Chinese companies do take the lead in AI, then Chinese values and censorship laws may become embedded into this technology in a way that is very hard to extract.

Yeah. I think that's true. I also always urge caution when people try to use the existence of China as a reason to dramatically accelerate the AI race. A lot of those people have made investments that will pay off handsomely if we find ourselves in some sort of protracted and awful conflict with China. So whenever anyone starts talking about China in the context of AI, my eyebrows arch up a little bit.

All right, now, Kevin, there is one more group of folks that I think is quite justly nervous about what they're seeing out there with DeepSeek. Who is that?

5. Open Source Raises Safety Risks

So the third group of people that I would say are freaking out about DeepSeek are AI safety experts: people who worry about the growing capabilities of AI systems and the potential that they could very soon achieve something like general intelligence or possibly superintelligence, and that that could end badly for all of humanity.

And the reason that they're spooked about DeepSeek is that this technology is open source. DeepSeek released R1 to the public. It's an open-weights model, meaning that anyone can download it and run their own versions of it or tweak it to suit their own purposes.

And that goes to one of the main fears that AI safety experts have been sounding the alarms on for years, which is that once this technology is invented, it is very hard to control. It is not as easy as stopping something like nuclear weapons from proliferating, and if future versions of this are quite dangerous, it suggests that it's going to be very hard to keep that contained to one country or one set of companies.

Yeah, I mean, say what you will about the American AI labs, but they do have safety researchers. They do at least have an ethos around how they're going to try to make these models safe. It's not clear to me that DeepSeek has a safety researcher. Certainly, they have not said anything about their approach to safety, right?

As far as we can tell, their approach is, “Yeah, let's just build AGI, give it to as many people as possible, maybe for free, and see what happens.” And that is not a very safety-forward way of thinking.

So, Casey, that is a lot of information that we just dumped on our listeners, but really what I want to know is, are you freaked out about this?

Mm-hmm.

Do you think that this is as big a deal as some people seem to think it is?

6. The Case Against Panic

Mm. I think as I'm doing my reading and having conversations with folks this morning, my sense is I am freaking out a bit less than some other folks that I'm talking to. I think this is a big deal and merits discussion, but I also think that people may be getting a bit over their skis when it comes to thinking through the implications here.

So make that case, because all I'm seeing all over my timelines is people saying, “This is the Sputnik moment for AI. This is the biggest moment in AI since the release of ChatGPT. Everyone needs to stop what they're doing and pay attention.” So what is the case that you are seeing out there that people are hyperventilating a bit over nothing?

Sure. So let's take a few different points. One reason why people are really nervous here is that DeepSeek was able to train this model very cheaply, and I want to be clear: this is a significant technical achievement.

At the same time, the cost of training and inference has been falling rapidly in AI for a long time now. Ethan Mollick, who we've had on the show before, posted a chart on X that showed this decline, and in some cases, for example, running inference on a GPT-4-level model, the cost of that has fallen a thousandfold over the past couple of years.

So things have already been moving in this direction, and I think most people who work in AI expected that it would continue to go there. So that is the first point that I would make.

Got it. And if you are Satya Nadella at Microsoft, Sam Altman at OpenAI, or Sundar Pichai at Google, are you worried that you are going to spend tens or hundreds of billions of dollars building out new data centers and filling them with all the fanciest GPUs, and that some Chinese startup is going to just take everything that you do and copy it 3 months later for pennies on the dollar?

So this is a great question, which leads me to a second reason why I think at least some folks may be overreacting here, and that is that in most cases, the money that is being spent to build out the data centers that will handle these giant training runs can be repurposed.

The same servers and chips that you would use to do that can also be used to serve what is called inference, so basically actually answering the questions. As more and more people start to use AI, it will be those giants that actually have the capacity to serve those queries. They will be able to build businesses around that.

And by the way, that is another reason why I don't think that DeepSeek is evidence that the export controls failed. The folks over at DeepSeek would love to have all of these chips, not just to do the big training runs, but also so that they could serve all of the demand that they are currently generating, right? So I think that's another important thing to keep in mind as this discussion moves forward.

Yeah, that makes a lot of sense to me. I do think the cost dynamics here are very important, because I talked to a person I know who works at one of these big companies, and he said that a lot of their customers are already starting to ask, “Well, could we shift over from using the OpenAI APIs and their models to using DeepSeek if it saves us 80% of our costs?”

And so I think in the short term, there is reason for the American AI companies to worry because people want these things to be as cheap as possible.

Yeah, and let me just say, as somebody who spent $200 to upgrade to GPT Pro last week so I could try Operator, I'm looking forward to that price going down.

7. Open Source Closes The Gap

But that leads to, I think, maybe the third reason that I think people might be overreacting a little bit here, which is that a lot of what we are seeing is essentially a fancy ripping-off of techniques that were pioneered here in the United States, right?

It has long been the case that open-source models were just a little bit behind the models that were made by the big labs. You look at Meta's Llama models, which until DeepSeek were seen as the best open-weights models that were out there. They weren't as good as what OpenAI or Google or others were doing, right?

Where I do think that this gets super interesting is that DeepSeek is showing us open source can now catch up faster than it used to, right? The labs used to have a little bit longer lead, but now people are just getting cleverer and cleverer about these techniques. And so it is getting harder to build that defensible moat because this is just one of those technologies where once you figure out basically how people are doing it, you could just get in there and do it too.

Yeah. Now, Casey, I'm curious what, if anything, you are hearing from inside Meta specifically, because I think this is one of the most fascinating angles.

Meta is a company that has spent billions of dollars developing AI models and, unlike most of its competitors, has chosen to release those models freely. And part of what DeepSeek has shown is that you can take a model like Llama 3 or Llama 4, and you can distill it. You can make it smaller and cheaper. You can do that without sacrificing a lot of the performance.

And so there were some reports in recent days that Meta is basically at DEFCON 1 right now, that they are—The Information reported that they have 4 war rooms at Meta headquarters devoted to trying to figure out how to respond to the DeepSeek threat.

Yeah.

And—

And by the way, I hope they were the same war rooms that Meta used to use to protect America from election interference.

They say, “Hey, get out of here. We got something else we gotta figure out.”

So are you hearing anything from Meta? Because I think that is the company that I would say has the most to worry about when it comes to DeepSeek, because DeepSeek is doing essentially what they do but at a fraction of the cost.

Yeah, so I do not have my own original reporting to share on this yet, but I do trust the information that they are freaking out. And the reason is that Meta is supposed to be the best company at ripping other people off. And so when they find out that some Chinese Johnny-come-latelys are gonna be better than they are at ripping things off, they're gonna have something to say about it.

Nothing could be more poetic now that DeepSeek has ripped off all the American companies. Meta is coming back, and they say, “Oh, you think you're good at ripping people off? Just wait until we have plumbed the guts of DeepSeek-V3 and DeepSeek-R1. We're gonna be back on top sooner or later, bucko.”

Yes. Now, I wanna ask you about one other reaction that I saw on social media, which was from Satya Nadella, the CEO of Microsoft. He went on his X account late last night and posted the following: “Jevons paradox strikes again. As AI gets more efficient and accessible, we will see its use skyrocket, turning into a commodity we just can't get enough of.” And then he linked to a Wikipedia article about Jevons paradox.

So, Casey, did you understand this, and if so, what did you make of it?

Well, I did, because we had just discussed Jevons paradox on this very show, Kevin.

It's true.

When Hugging Face's Sasha Luccioni came on and explained Jevons paradox, which is essentially as stuff becomes more efficient, you simply increase demand for it, thereby canceling out a lot of the efficiency gains. So when I saw Satya tweet “Jevons paradox,” I said, “Once again, Hard Fork has set the national news agenda—and if you're not listening, fix that.”

Yeah, many people are talking about Jevons paradox. I predict that this is going to be something I'm gonna hear about at every single party I go to for the next 6 months.

And, just to connect the dots a little bit, I think what Satya is trying to say here is that DeepSeek is not actually a threat to companies like Microsoft because as the cost of building and using AI models comes way down, people are just gonna wanna use them more and more, and so the overall demand and Microsoft's overall profitability will not change. Which could be true, but I would also just say it is exactly what you would expect the CEO of Microsoft to say on a day where investors were panicking and selling their stock.

It is. It is. All right. Well, Kevin, I think that's a pretty good overview of what DeepSeek is doing, why people are freaking out, and at least some thoughts about exactly how freaked out you should be.

There is a lot more to say about this subject, and if you are starving for even more discussion of DeepSeek, I can promise you that we'll have more to say on our regularly scheduled episode of Hard Fork this Friday.

Yes, Casey, I love doing these emergency podcasts. They fill me with a sense of danger—

Mm-hmm.

—and excitement, living on the edge.

I love 'em for that reason. I love them for a second reason, Kevin, which is that I get paid by the episode. So here's to many more emergencies in 2025.