(Preview) Happy Lunar New Year and A Few Thoughts on DeepSeek
10:25 Share

Hello and welcome to Sharp China. I'm Andrew Sharp and you are listening to a free preview of today's episode. Hello and welcome back to another episode of Sharp China. I'm Andrew Sharp and on the other line, Bill Bishop. Bill, happy year of the snake. How you doing?

I'm good, thanks. Hey, Andrew, happy New Year to you and all of our listeners. Today is officially Lunar New Year's Day. Indeed. Have you run into any snakes in your travels the past couple days? No, so far, not seen any. I'm hoping, you know, I think it's auspicious to see one. A couple years ago, we had a little baby one slithering through our house, not on New Year's Day, but from the park with a lot of snakes in the park. So maybe we'll get lucky and see one.

I'm still scarred by the story at the end of the last episode. You discovering the snake in the hotel room. That was in Bali, right? Yeah, that was in Bali, yes. Yeah. My favorite part is you having to put on a brave face in front of your daughters and say, oh, this is really auspicious and just play it cool the whole time. Well, to be fair, I think it is auspicious. However, it was still a little...

You'd rather have a bit more of a controlled environment, put it that way.

But we would be remiss if we didn't have a quick conversation about DeepSeek, which has been a trending topic around the world this week. Your line in Monday's newsletter basically sums it up. Who had a PRC AI firm possibly popping the US AI stock bubble on their black swan slash gray rhino list? So quite a couple of days. How are you feeling about the whole thing?

I feel like there's still a lot we don't know. On the one hand, it shouldn't be surprising that there are really talented engineers in China who are very focused on adapting to the constraints they're under and building better products. On the other hand, the reaction in the stock market was quite something to witness. And it really felt like there was some pretty concerted efforts to talk down China.

to pop this bubble and specifically talk down NVIDIA and some of the other chip stocks. Twitter was full of all sorts of... People talk about Think Boys and there are also a bunch of Shill Boys talking about this shows the US tech controls have failed. It was all full of triumphalism, which again, I think is quite premature. But it was really interesting to watch how

sort of information really flowed and how this narrative was created that then led to this, you know, bit of a stock market panic in certain sectors on Monday. Yeah. I mean, as someone who started a business in China and is very familiar with the tech scene over there, are there any elements of the DeepSeek backstory in particular that you found interesting? Because there was...

basically like a creation myth that accompanied the V3 technology and then the R1 model. And then people have been poking holes in that myth over the last week or so. Well, what's interesting is this, you know, this is a quant fund. Right.

Mm-hmm.

They put in this new guy, a bunch of rules have come out, regulations, reforms for the stock market, including cracking down on quant funds. And so I think in his previous quant fund, the high flyer quant fund, I think they were in a bit of a doghouse. But he had bought a lot of Nvidia chips.

to build his quant models before any export controls. We don't know exactly how many he has. There's talk of big numbers. He has some, for sure. But then as the quant fund came under pressure, he clearly saw an opportunity or a need to diversify into doing more

specifically AI large language model focused work. And if you look at the landscape of AI firms in the PRC,

His deep seek was known as having talented engineers, but they were not the ones sucking up vast amounts of capital. They weren't riding the leading edge of the hype machine, both to get government support as well as to attract more investment because they don't have any outside VCs. And so, I mean, good for them. Of course, we learned today or last night that

OpenAI and Microsoft are investigating whether or not they exfiltrated a whole bunch of data more than they were supposed to against the OpenAI terms of use to distill their model they had. And so I think we still, you know, there are lots of people drawing all sorts of conclusions about what DeepSeek's new models mean. I actually don't think we have yet really have enough information yet

To understand that, I think it is so great. I think, frankly, it's great news because it's going to help the companies in Silicon Valley. The information reported that Meta set up four war rooms to figure out how they did all this stuff because they laid out. They were very actually helpful to companies by open sourcing it and detailed papers about what they did. So actually, I think it's going to help any company that really wants to dig in. It's probably going to learn how to make their systems much more efficient.

Yeah.

will be useful for American tech companies. And by the same token, a lot of people saying, well, this is proof that nothing that the U.S. has done on the export control side has been effective and Chinese AI is great and just as good as the U.S. companies is.

I think it's fair to assume that the deep-seek model wouldn't exist. It wouldn't be as effective as it is, if not for some of the leading-edge AI models in the U.S. because of the distillation that allegedly happened. I'm not sure. Well, and there's distillation, and then there's exfiltrating data. I mean, again, there's a spectrum of...

acceptable distillation and unacceptable distillation. And at least Bloomberg, now multiple media outlets have reported that open AI and Microsoft think it was more on the unacceptable side of the spectrum. Indeed. And that could be, I could be asked covering more will be revealed. Exactly. I mean, that's the other thing is every time somebody says that, um,

then there's another crowd that says, oh, this is just cope, blah, blah, blah. Yeah, exactly. My one request is that everyone on the internet stop using the word cope because it's just been very frustrating the last couple of days. I feel like we're running that particular term into the ground. The move by DeepSeek to open source this is really fascinating because it, again, I think helps competitors in the rest of the world

by showing them ways to get more out of the massive investments they've made. It also, again, it's now a non-US-driven, US-controlled model that can be used all over the world. And what's interesting, right, is you go back to the AI chips export diffusion rules that came out the last gasping days of the Biden administration, where...

divided the world into three tiers. Tier 1, close US allies, they get whatever they want. Tier 2, got to apply for things and maybe their quotas. Tier 3, no.

is basically, you know, if you're like a deep seek, so tier two countries, tier three countries, they're all your customers. Totally. Yeah. It's like a, it's like a customer, but potential customer lists. If, if you're selling, I mean, again, this is free, but my point is more that for a lot of places around the world, this is a deep seek model is a gift. Indeed. Indeed.

And for the PRC, a model like DeepSeek is a real soft power tool that could be pretty valuable in the future, particularly if the U.S. is going to seed markets like this by remaining closed source and trying to sort of tier its technology transfers around the world. I think the ultimate takeaway is multiple things can be true at once. The DeepSeek model

We might not all have enough information to make conclusions to. Exactly. And there are a lot of people who are speaking in declarative sentences and more will be revealed on several of these fronts. So a couple of things like on the sort of, you know, mother or what is it, necessities of mother invention, the fact that Deep Seek, you know, had some

pool of media chips, but probably not as many as they wanted. And that forced them to find ways to be much more creative and efficient about how they design the ways to get the model to work. I think the company, the CEO, other employees have gone on the record as saying their biggest constraint is compute.

All right. And that is the end of the free preview. If you'd like to hear the rest of today's conversation and get access to full episodes of Sharp China each week, you can go to your show notes and subscribe to either Bill's newsletter, Cynicism, or the Stratechery bundle, which includes several other podcasts from me and daily writing from my friend Ben Thompson. I'm an incredibly biased news consumer, so I think both are indispensable resources for

But either way, Bill and I are going to be here every week talking all things China, and we would love to have you on board. So check out your show notes, subscribe, and we will talk to you soon.