Speaker 1: The single most important thing that DeepSeek show, you have to understand this, this is the single biggest thing. So the last couple of days, DeepSeek has really blown up in the mainstream media. I think a lot of people in tech already kind of knew about it, but it's really exploded. Like I think everywhere you look, everyone is talking about DeepSeek. They're talking about US versus China. They're talking about how all these big tech companies are going to crash now. And I think there's a lot of sort of misinformation going on. And maybe even more importantly, I think a lot of people are taking this news and misinterpreting it. And it's not entirely a people's fault. I think the news is not really helping either. So I kind of wanted to clarify certain things and also offer my opinion for whatever it's worth. So first, I just want to add a little bit of context. DeepSeek R1 released on January 20. And so this is the DeepSeek API docs. And if you look at the benchmarks, it's pretty much on par or slightly exceeding most of the state of the art models that are pretty much offered by US companies. I think the vast majority of the state of the art models before DeepSeek were all US, whether it's OpenAI, Claude, Llama, Gemini, et cetera. And so now that there's a Chinese model, people think that this changes everything. And actually, it does change everything, but not for reasons that people think. Let's just go back in time just a little bit, just one month ago, and take a look at this. One month ago, they released DeepSeek V3. So first of all, R1 is more of that like slow reasoning model that OpenAI released, the O1 model, the one that in some cases can be really, really good at leak code style problems, but sometimes it can be really, really bad at general purpose things. Well, DeepSeek, prior to R1, released V3. And to me, this was actually the more interesting model, the one that's probably more useful in most cases. And this one was kind of the one that's comparable to Claude 3.5 SUNET or GPT 4.0. And it was practically on par or slightly exceeding most of the models in most categories. You can see maybe there's a couple that it is slightly worse Now, this happened on December 26 of last year, just a month ago. And so you might be wondering, why wasn't anybody talking about this back then? I can't really say for sure. I can tell you, everybody in the tech community, or at least a lot of people in tech were talking about it on tech Twitter. And even people who weren't directly mentioning it, take a look at this. This was December 26. Take a look at this subtweet from Sam Altman, who I was hating him before it was cool. So I just want people to know that. But take a look at this subtweet. What do you think he was talking about here? December 27, a day after he said, it is relatively easy to copy something that you know works. It is extremely hard to do something new, risky, and difficult when you don't know if it'll work. Individual researchers rightly get a lot of glory, blah, blah, blah. It's the coolest thing in the world. He's not a very genuine person, in my opinion. But what do you think he was talking about? He was talking about DeepSeek, in my opinion. I don't think that's a conspiracy theory to say he's subtweeting about DeepSeek. Why didn't he mention it by name? Because he didn't want it to blow up. He didn't want it to get attention. It took a little while for the media to catch up and realize that DeepSeek existed. But eventually, they did. And you can see that some people did actually know. If you scroll through the replies, there were at least one or two people saying DeepSeek has left the chat. And some people knew what he was talking about. Most people didn't. And just yesterday, I think, if you look at his Twitter, now he's mentioning it by name because everybody already knows about it. Now he's saying DeepSeek R1 is an impressive model. Why didn't he say anything about V3 last month? Because that's just how this works. It's kind of a game of politics and a game of marketing and a game of hype. So it's very hard to kind of clearly know what's going on. While I mentioned that DeepSeek is on par with the state-of-the-art models, in some ways, it actually is significantly better. It's actually more transparent. And I'm not just talking about the fact that it is open source, that you do not have to actually use the DeepSeek API directly. You can download the model weights and run the system locally, provided you have a good enough computer. And so that's one of the big benefits of DeepSeek. It is true. It is open source. I believe that's a good thing because OpenAI is obviously not open sourced. I think they're trying to extract as much profit as they possibly can. They believe that building a really, really good model gives them some kind of competitive advantage. And what DeepSeek showed is that is absolutely not true. And that is a very good thing. Let me tell you why. First of all, take a look at this. DeepSeek will actually give you its entire chain of thought reasoning when you ask it something. That is something that O1 does not do. Like OpenAI is not as transparent as this Chinese model, this Chinese model that everybody should be scared of now. China is going to take over. Well, they're doing it a little bit better. And the criticism, of course, is that since this is coming out of China, there's a lot of censorship when it comes to Chinese laws. There are certain things you're not really allowed to talk about in China. And so DeepSeek does have to adhere to those laws. So if you ask it certain things about China, it's not going to get involved in Chinese politics. I think that is a very significant thing. I'm not kind of making light of that. The censorship aspect is very real. But I don't think DeepSeek is necessarily doing it maliciously. I think that's just the laws that they have to follow. And so in some cases, they're probably going to be aligned with the Chinese government. But the single most important thing that DeepSeek showed, you have to understand this, this is the single biggest thing. And most people aren't really talking about it. How did DeepSeek get so good? How is it that it's on par with GPT and the other models? Well, some people asked DeepSeek, which model are you? And it thinks it's ChatGPT. Why is that? Well, it's not a coincidence, because the reason part of the reason that they were able to create such a good model so quickly and at a relatively low cost was that they practically copied GPT, ChatGPT's models, the one that OpenAI has created, because they violated ChatGPT's terms of service, or they probably use like the OpenAI APIs, but they violated the terms of service because they used the inputs and outputs of the GPT models to actually train their own models. And that's against terms of service. But you want to know what's also against terms of service, the things that OpenAI has been doing by scraping everybody's data, everybody's videos, everybody's images and stuff, and then using those to train their models. That's also against terms of service. So it's kind of a little bit of karma for OpenAI to have the same thing happen to them. They stole everybody else's data to create these GPT models. And now somebody just kind of stole their models and took them and then made them open source. And the reason that this is so significant is that it shows that if this were to happen with the GPT 4.0 model, if it was to happen with the 0.1 model, well, couldn't it happen with every single model? OpenAI is investing billions, hundreds of billions of dollars to build these models. And if it turns out that anybody can just copy the models and then make them open source, which OpenAI does not want to do, well, then why would OpenAI invest so much money? Why would they be the first mover when there isn't a first mover advantage? In fact, it's the opposite to do that. And it's not just going to be deep seek now. This is very, very good for the world, in my opinion, because now my prediction, and I don't think it's a crazy one, a lot of the US companies, including Meta, who has also been doing a lot of open source work, they're going to do the exact same thing. Now they're going to have state of the art models. And so it pretty much guarantees that the best models are pretty much going to be open source eventually. And it's probably going to happen pretty quickly, just like it has with a deep seek. Now, what exactly does this mean for like the big tech companies that are investing billions and billions into this? Well, I think some of them are probably going to think twice about it, even though, you know, in Sam Altman's position, he's not going to tell them that he's going to say, well, we still now, this is just a reason for us to compete even harder against China. Let's pour more money into this, because then we'll have an advantage over China. China's catching up. No, China did not show that they're catching up. They showed that everybody is already on an equal footing. That's what they showed. And like, do not miss that. Sam Altman is not being entirely truthful with what he's saying. And of course, like there's plenty of other misinformation going on where people think that the CEO of Deep Seek is a memer. I think that's pretty unlikely. And yeah, that's that's misinformation. I would agree with this kind of conclusion. Deep Seek is a really, really good model, but it's not generally better than all the models that are available right now. It has kind of some pros and cons. That's why I think it's like roughly on par with those models. Most people wouldn't really be able to notice the difference. And so those are kind of my thoughts. Maybe some of you guys disagree. I'd be happy to hear what you have to say in the comments. But I think I am roughly correct with most of the things that I've said here. I think it's a very, very interesting thing that's going on. It's fun to follow. I can't really say what it means for how the tech market is going to go. Is Nvidia as a company that sells GPUs now going to go up or is it going to go down? Who knows? What does this mean for tech hiring? I really don't know. Maybe somebody on the Internet can tell you what it means and they can tell you it very, very confidently. But I think this shows that in many cases, even experts have no idea what they're talking about. And that's what I've been kind of saying, I think, for the last year or two, that you can't really trust the experts because either they could be lying or they could just be flat out wrong. I don't think most people would have predicted that deep secretive came and that kind of just annihilated the advantage of creating a state of the art models. Nobody would have predicted this. None of the experts. I didn't see any of the experts predicting this, but it happened and it happened, I think, sooner than people would have expected.