The AI Model That Tanked the Stock Market

SPEAKER_2 (0:02)

Bloomberg Audio Studios, podcasts, radio, news.

Joe Weisenthal (0:18)

Hello, and welcome to another episode of the Odd Lots Podcast. I'm Joe Weisenthal.

Tracy Alloway (0:23)

And I'm Tracy Alloway.

Joe Weisenthal (0:24)

Tracy, the DeepSeek sell-off.

Tracy Alloway (0:27)

That's right, it's pretty deep. Has anyone made that joke yet?

Zvi Mowshowitz (0:30)

We're in DeepSeek, yeah.

Joe Weisenthal (0:31)

I don't think anyone has made that joke yet.

Tracy Alloway (0:33)

I will say, like, you know it's bad in markets when all the headlines are about standard deviations. Yeah, right. And then you know it's really bad when you see people start to say, it's not a crash, it's a healthy correction. And that's the real cope.

Joe Weisenthal (0:49)

But just for like real scene setting, you know, we've done some very timely interviews about tech concentration in the market lately and how so much of the market is this big concentrated bet on AI, et cetera. Anyway, on Monday, I think people will be listening to this on Tuesday, markets got clobbered. Nvidia, one of the big winners as of the time I'm talking about this, 3:30 PM on Monday, down 17%. So we're talking major losses really across the tech complex.

Basically, it seems to be catalyzed by the introduction of this high-performance, open-source Chinese AI model called DeepSeek. It was born from what we know, out of a hedge fund. Apparently, it was very cheap to train, very cheap to build. You know, the tech constraints at this point didn't seem to be much of a problem. They may be a problem going forward. But yes, here is something the entire market betting on a lot of companies making AI and are now concerns about, of course, a cheap Chinese competitor.

Tracy Alloway (1:45)

I just realized, Joe, this is actually your fault, isn't it? Yeah. Because last week you wrote that you were a DeepSeek AI bro. And look what you've done. You've wiped $560 billion off of Nvidia's market cap.

Joe Weisenthal (1:58)

Might be.

Tracy Alloway (1:59)

That's you.

Joe Weisenthal (2:00)

Anyway, one of the interesting questions, though, is that this was sort of announced in a white paper in December. Why did it take till January 27 for it really to freak people out? Big questions. Anyway, let's jump right into it. We really do have the perfect guest. Someone who was here for our election eve special, a guy who knows all about numbers and AI and quant stuff, and he writes a substack that has become for me a daily absolute must read, where he writes an extraordinary amount. I don't even know how he writes so much on a given day. We're going to be speaking with Zvi Mowshowitz. He is the author of the Don't Worry About the Vase blog or substack. Zvi, you're also a DeepSeek AI bro. You've switched to using that?

Zvi Mowshowitz (2:43)

I use a wide variety of different AIs. I will use Claude from Anthropic. I will use O-One from Chatsheep BT from OpenAI. I'll use Gemini sometimes, and I'll use Perplexity for web searches, but I'll use R1, the new DeepSeek model, for certain types of queries where I want to see how it thinks and see the logic laid out, and then I can judge, did that make sense? Do I agree with that?

Tracy Alloway (3:08)

So, one of the things that seems to be freaking people out as well as the market is that purportedly this was trained on a very low cost. Something like $5.5 million for DeepSeek V3. Although I've seen people erroneously say that the $5.5 million was for all of its R1 models, and that's not what it says in the technical paper. It was just for V3. But anyway, oh, I should mention it also seems like a big chunk of it was built on Lama, so they're sort of piggybacking off of others' investment. But anyway, $5.5 million to train. Is that A, realistic, and then B, do we have any sense of how they were able to do that?

Zvi Mowshowitz (3:53)

So we have a very good sense of exactly what they did because they are unusually open, and they gave us technical papers that tell us what they did. They still hid some parts of the process, especially with getting from V3, which was trained for the $5.5 million, to R1, which is the reasoning model for additional millions of dollars, where they tried to make it a little bit harder for us to duplicate it by not sharing their reinforcement learning techniques. But we shouldn't get over-anchored or carried away with the $5.5 million number. It's not that it's not real, it's very real. But in order to get that ability to spend $5.5 million and get the model to pop out, they had to acquire the data, they had to hire the engineers, they had to build their own cluster, they had to over-optimize to the bone their cluster because they're having problems of chip access thanks to our export controls and they're training on H800s. And the way they did this was they did all these sorts of little optimizations, including just exactly integrating the hardware, the software, everything they were doing in order to train as cheaply as possible on 15 trillion tokens and get the same level of performance or, you know, close to the same level of performance as other companies have gotten with much, much more compute. But it doesn't mean that you can get your own model for $5.5 million, even though they told you a lot of the information. In total, they're spending hundreds of millions of dollars to get this result.