Odd Lots - The AI Model That Tanked the Stock Market

Episode Date: January 28, 2025

On Monday, the stock market tanked, seemingly in reaction to the emergence of DeepSeek, an open source AI model developed in China. Nvidia, the semiconductor giant that has been the largest winner of the AI boom, erased $589 billion in market cap, for the biggest one-day wipeout in US stock-market history. Other chipmakers and big tech giants also swooned. So how did DeepSeek do it? Is it a big threat to the American AI giants like OpenAI and Anthropic? What does this say about export restrictions on US chips? On this special emergency session of the podcast, we spoke with Zvi Mowshowitz, an AI expert who authors the excellent Substack, Don’t Worry About the Vase. He answered all our questions and more to help understand what it means. Read more: AI-Fueled Stock Rally Dealt $1 Trillion Blow by Chinese UpstartWorld’s Richest People Lose $108 Billion After DeepSeek Selloff Only Bloomberg.com subscribers can get the Odd Lots newsletter in their inbox — now delivered every weekday — plus unlimited access to the site and app. Subscribe at bloomberg.com/subscriptions/oddlots See omnystudio.com/listener for privacy information.

Bloomberg Audio Studios, Podcasts Radio News. Hello and welcome to another episode of the Odd Lots podcast. I'm Joe Wisenthall. And I'm Tracy Allaway. Tracy, the deep seek sell-off. That's right. It's pretty deep. Has anyone made that joke yet?

We're in deep seek, yeah. I don't think anyone has made that joke yet. 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. Yes. That's the real cope. 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, etc. Anyway, on Monday, I think people will be listening to this on Tuesday. Markets got clobbered in video. One of the big winners as of the time I'm talking about this, 3.30 p.m. on Monday, down to 17%. since we were 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. I 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. I just realized, Joe, this is actually your fault, isn't it? Yeah, yeah. Because last week you wrote that you were a deep seek AI, bro, and look what you've done.

You've wiped $560 billion off of NVIDIA's market. Yeah, my being. That's you. Anyway, one of the interesting questions, though, is this was sort of announced in a white paper in December. Why did it take until January 27 for 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 Moshevitz. He is the author of the Don't worry about the Vaz blog or substack.

Zvi, you're also a deep seek AI. bro. You've switched to using that? So I use a wide variety of different AIs. I will use Claude from Anthropic. I will use O1 from Chad CTPT from OpenAI. I'll use Gemini sometimes and I'll use perplexity for

web searches. But yeah, I'll use R1, the new deep seat model for certain types of queries where I want to see how it thinks and like see the logic laid out and then I can judge like, did that make sense? Do I agree with that? So one of the things that seems to be

freaking people out as well as the market is that purportedly this was trained on like 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? So we have a very good sense of exactly what they did because they're 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-optimized to the bone their cluster because they're having problems of chip access thanks to our export controls.

and training on 800s. And the way that they did this was they did all these sorts of mini-optimist, 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. Wait, explain that further. Why does it still take hundreds of millions? And does this mean if it takes hundreds of millions of dollars that the gap between what they're able to do versus the, say, American labs is perhaps not as wide as maybe people think? Well, what Deepseek is doing is they have less access to chips. They can't just buy Navidia chips the same way that, you know, Open AI or Microsoft or Anthropic can buy Navidia chips. So instead, they had to make good use, very, very efficient killer use of the chips that they did have. So they focused on all these optimizations and all these ways that they could save on compute.

But in order to get there, they had to spend a lot of money to figure out how to do that and to build the infrastructure to do that. And once they knew what to do, it cost them $5.5 million to do it. And they've shared a lot of that information. And this has dramatically reduced the cost of somebody who wants to follow in their footsts. steps and train a new model because they've shown the way of many of their optimizations that people didn't realize they could do or didn't realize how to do them that can now very easily be copied. But it does not mean that you are $5.5 million away from your own V3. So the other thing

that is freaking people out is the fact that this is open source, right? We all remember the days when open AI was more open and now it's moved to closed source. Why do you think they did that And like how big a deal is that? So this is one of those things where they have a story and you can believe their story or not believe their story. But their story is that they are essentially ideologically in favor of the idea that everyone should have access to the same AI. That AI should be shared with the world, especially that China should help pump out its own ecosystem and they should help grow all of the AI for the betterment of humanity.

And they're going to get artificial general intelligence and they are going to open source that as well. And this is their the main point of Deepseek. This is why Deepseek exists. They're disclaiming even having a business model really. And, you know, they're an outgrowth of a hedge fund. And the hedge fund makes money. And maybe they can just do this if they choose to do that. Or maybe they will end up with a different business model.

But it was obviously very concerning from a lot of angles if you open source increasingly capable models because, you know, artificial general intelligence means something that's, you know, as smart and capable as you and I, as a human, and perhaps more so. And if you just hand that over in open form to anybody in the world who wants to do anything with it, then we don't know how dangerous that is, but it's existentially risky at some limit to unleash things that are smarter, more capable, more competitive than us that are then going to be free and loose to, you know, engage in whatever any human directs them to do. I have a really dumb question, but I hear people say artificial general intelligence all the time, AGI.

What does that actually mean? There is a lot of dispute over exactly what that means. The words are not used consistently, but it stands for artificial general intelligence. Generally, it is understood to mean you can do any task that can be done on a computer that can be done cognitively only, as well as a human. I mean, most of these things do things much better than me. I don't know how to code. But I get that there are still some things. Maybe they wouldn't be as good as proving some of the are you human tests.

Everyone's talking about Jevin's paradox. And so we see InVITA and Broadcom shares, these chip companies, they're getting crumbled today. And one of the theories like, oh, no, with all these optimizations and so forth, researchers will just use those and they'll still have max demand for compute. And so it won't actually change the ultimate end for compute. How are you thinking about this question? So I'm definitely a Jevon's paradox, bro, right now from the perspective of this debate. So that you don't think it'll have a negative impact and just the amount of compute demanded.

The tweet I sent this morning was Navidia down 11% pre-market on news that his chips are highly useful. And I believe that what we've shown is that, yes, you can get a lot more, in some sense, out of each Navidia chip than you expected. you can get more AI. And if there was a limited amount of stuff to do with AI, and once you did that stuff you were done, then that would be a different story. But that's very much not the case. As we get further along towards AGI,

as these AIs get more capable, we're going to want to use them for more and more things more and more often. And most importantly, the entire revolution of R1 and also open AIs O1 is inference time compute. What that means is every time you ask a question, it's going to use more compute, more cycles of GPUs, to think for longer, to basically use more tokens or words, to figure out what the best possible answer is. And this scales, not necessarily without limit, but it scales very, very far. So Open AI's new 03 is capable of thinking for, you know,

many minutes, it's capable of potentially spending, you know, hundreds or even in theory thousands of dollars or more on individual query. And if you knock that down by an order of magnitude, that almost certainly get you to use it more for a given result, not use it less, because that is in fact starting to get prohibitive. And over time, if you have the ability to spend remarkably little money and then get things like virtual employees

and abilities to answer any question under the sun, yeah, there's basically unlimited demand to do that or to scale up the quality of the answers as the price drops. So I basically expect that as fast as NVIDIA can manufacture chips and we can put them into data centers and give them electrical power, people will be happy to pie those chips.

At the risk of angering the Jevins Paradox bros, just to push on the InVideo point a little bit more. So my understanding of Deepseek is that one of the reasons it's special is because it doesn't rely on specialized components, custom operators. And so it can work on a variety of GPUs. Is there a scenario where, you know, AI becomes so free and plentiful, which could in theory be good for Invidia, but at the same time, because it's easy to run on a bunch of other GPUs, people start using, you know, more like ASIC chips,

like customized chips for a specific purpose. I mean, in the long run, we will almost certainly see specialized inference chips, whether they're from David or they're from someone else. And we will almost certainly see various different advancements. Today's chips are going to be obsolete in a few years. That's how AI works, right? there's all these rapid advancements. But, you know, I think Nvidia's in a very, very good position to take advantage of all this.

I certainly don't think that, like, you'll just use your laptop to run the best AGIs, and therefore, we don't have to worry about buying GPUs is a poor position. It's certainly possible that rivals would come up with superior chips. That's always possible.

Navidia does not have a monopoly. But Navidia certainly seems to be an dominant position right now. It seems to me, I mean, I know there's others, but it seems to be in the U.S., there's like three main AI producers of models that people know about. There's open AI, there's Claude, and then there's meta with Lama. And it's worth knowing that meta is green today, that the stock is actually up as of the time I'm talking about this, 1.1%.

Just go through each one real quickly how the sort of deep seek shock affects them and their viability and where they stand today. I think the most amazing thing about your question is that you forgot about Google. Oh yeah, right. Yeah, that's very telling, isn't it? But everyone else has forgotten about Google as well. It wasn't that surprising. Gemini Flash thinking, their version of 01 and R1 got updated a few days ago. And there are many reports that it's actually very good now and potentially competitive.

And effectively it's free to use for a lot of people on AI Studio. But nobody I know has taken the time to check and find out how good it is because we've all been too obsessed with being deep seek pros. And Google's had its rhetorical lunch eaten over and over and over again. December, like Open AI would come out with advance after advance after advance. Then Google would have advance after advance. And Google's would be seemingly actually, if anything, more impressive. And yet everyone would always just talk about Open AI. So this is not even new.

Something is going on there. So in terms of Open AI, open AI should be very nervous in some sense, of course, because they have the reasoning models. And now the reasoning model has been copied much more effectively than previously. And the competition is a hell of a lot cheaper than when OpenA. open AI is charging. So it's a direct threat to their business model for obvious reasons. And it looks like their lead in reasoning models is smaller and faster to undo than you would expect. Because if Dipsy can do it, of course, Anthropic and Google can do it and everyone else can do it as well. Anthropic, which produces Claude, has not yet produced their own reasoning model. They clearly are

operating under a shortage of compute in some sense. So it's entirely possible that they have chosen not to launch a reasoning model even though they could or not focused on. on training one as quickly as possible until they've addressed this problem. They're continuously taking investment. We should expect them to solve their problems over time, but they seem like they should be directly concerned because they're less of a directly competitive product in some sense. But also they tend to market to effectively much more aware people, so their people will also know about Deepseek and they will have a choice to make. If I was meta, I would be far more worried, especially if I was on their Gen AI team and wanting to keep my job, because

meta's lunch has been eaten massively here, right? Meta with Lama had the best open models, and all the best open models were effectively fine tunes of Lama. And now, Deep Seek comes out, and this is absolutely not in any way a fine tune of Lama. This is their own product. And V3 was already blowing everything that META had out of the water. R1, there are reports that it's better than their new version that they're training now. better than Lava4, which I would expect to be true. And so there's no point in releasing an inferior open model of everyone on the open model community

just being like, why don't I just use deep seek? Tracy, it's interesting that Zvi said the people who should be nervous are the employees of meta, not meta itself, because meta is up. And so you've got to wonder, it's like, well, maybe they don't, I don't know. Maybe they don't need to invest as much in their own open source AI. if there's a better one out there, and now the stock is up. Anyway, the market has been very strange, from my perspective, on how it reacts to different things that meta does.

For a while, meta would announce, we're spending more in AI, we're investing in all these data centers, we're training all of these models, and the market would go, what are you doing? This is another metaverse or something, and we're going to hammer your stock, and we're going to drag you down. And then with the most recent $65 billion announced spend, then meta was up, presumably they're going to use it mostly for inference, effectively, in a lot of scenarios. because they have these massive inference costs to want to put AI over Facebook and Instagram.

So if anything like, you know, I think the market might be speculating that this means that they will know how to train better llamas that are cheaper to operate and their costs will go down and then they'll be in a better position. And that theory isn't crazy. Since we all just collectively remembered Google,

I have a question that's sort of been on the back, in the back of my mind. I think Joe has brought this up before as well. But like when Google debuted, it took years and years and years for people to sort of catch up to the search function. And actually, no one ever really caught up. Right. So Google has like dominated for years. Why is it when it comes to these chatbots, there aren't like higher, wider boats around these businesses?

So one reason is that everyone's training on roughly the same data, meeting the same data, meeting the. entire internet and all of human knowledge. So it's very hard to get that much of a permanent data edge there unless you're creating synthetic data off of your own models, which is what OpenAI is plausibly doing now. Another reason is because everybody is scaling as fast as possible and adding zeros to everything on a periodic basis. In calendar time, it doesn't take that long before your rival is going to have access to more compute than you had. And they're copying your techniques more aggressively. There's just a lot less secret sauce. There's only so many algorithms. Fundamentally, everyone is relying on the scaling law. It's called the bitter lesson. It's the idea that you know, you just

scale more. You just use more compute. You just use more data. You just use more parameters. And deep seek is saying maybe you don't, you can do more optimizations. You can get around this problem and still get a superior model. But mostly, yeah, there's been a lot of just I can catch up to you by copying what you did also because I can see the outputs, right? I can query your model. and I can use your model's outputs to actively train my model. And you see this in things like most models that get trained, you ask them who trains you? And they will often say, oh, I am from Open AI.

The internet's gotten so weird. The internet is so weird. Zvi Moshavits, thank you so much for running over to the odd lots and helping us record this emergency pod on the Deep Seek sell-off. That was fantastic. All right, thank you. Tracy, I love talking to Zvi. We got to just sort of make him our AI guy.

I mean, to be honest, we could probably have him back on again this week because there's going to be stuff happening. Maybe we will. And obviously, we could go a lot longer. This is a really exciting story. This is a really exciting story. And things are just getting really weird these days. It is kind of crazy how fast all of this is happening. And then the other thing I would say is just the bitter lesson, great name for a band.

Oh, totally, totally great. Maybe when we do our AI-themed Prague rock band, Tracy. That can be our name. Yes. Let's do that. Okay, shall we leave it there? Let's leave it there. This has been another episode of the Odd Lots podcast.

I'm Tracy Alloway. You can follow me at Tracy Alloway. And I'm Joe Wisenthal. You can follow me at the stalwart. Follow our guests, Zvimoshavits. He's at the Zvi. Also, definitely check out his free substack. It's a must read for me.

Don't worry about the Vaz. Really great stuff. every single day. Follow our producers, Carmen Rodriguez at Kermann, Dashel Bennett at Dashbot, and Kill Brooks at Kail Brooks. For more Oddlots content, go to Bloomberg.com slash oddlots. We have transcripts, a blog, and a newsletter. And you can chat about all of these topics 24-7 in our Discord. Discord.g.g. slash oddlots. Maybe we'll get Zvi to do a Q&A in there with people. Oh, yeah. That'd be great. And if you enjoy Oddlots, if you like it when we roll out these emergency episodes, then please leave us a positive review on your favorite platform. Thanks for listening.