Transcription
Deepseek just dropped their latest flagship model, V4. It's massive, powerful, open-source, and a fraction of the cost. And it might be the model that ends America's lead in artificial intelligence. Not because China caught up, but because of what happens next.
So, usually at this point, I would do a model overview. I would tell you about the model. I would show you the benchmarks. I would test it and show you what I think. But as I looked at it, I realized there was actually a much bigger story here.
America has the best chips. It has the most money flowing into AI labs. Yet, China was able to release a frontier-level model that matches the best of them. Completely open-source, completely open weights, and at a fraction of the cost and resources. They are literally working with nerfed Nvidia GPUs. That's not supposed to be possible, and the fallout will be bigger than people realize.
And so today, I'm going to tell you what's special about DeepSeek V4, why it matters, and what it means for the world. But first, if you love seeing videos about AI models, go ahead and hit the like and subscribe button. I want to reach as many people as possible and teach them about artificial intelligence, get them excited about it. And hitting the like and subscribe button helps the channel more than you realize. So, thank you for doing that in advance.
Okay, so what is so special about DeepSeek V4? First, let me tell you about who DeepSEK actually is. If you remember, about 18 months ago, they dropped a model that literally changed the world. It was called Deepseek R1, and it was an open-source, open-weights model that could think. Remember back then, models that could think were only developed by the closed-source AI labs in the United States. They dropped R1, showed the world that other countries and open-source labs could develop models that were at the frontier, and the stock market dropped 20% pretty much overnight.
And what was really special about Deep Seek Rar1 was how efficient they were able to train it, a fraction of the price than the hundreds of billions of dollars paid by the frontier US AI labs. And so people thought, "Wow, if they can train it at a fraction of the price, then maybe Nvidia GPUs are not actually worth that much." But it turns out they were very wrong about that. When things get cheaper in price, we actually use a lot more of it. That's called Jevons' Paradox.
Okay. And now, fast forward to today. Deepseek is back with V4. And they wrote an incredibly thorough white paper explaining how they did all of it, including being super honest about some of their failures, much more honest than any of the closed-source AI labs in the United States.
All right, so here's the post. It came out late last night. Let me tell you about it. Deepseek V4 preview is here, and it comes in two flavors: Pro and Flash. First, it has a million-token context length. That is amazing because that is the frontier. So immediately check that box. They are at the frontier of context limits.
Next, it is a 1.66 trillion total parameter model with 49 billion active parameters. This is called Mixture of Experts. It basically allows you to have a massive model but run only parts, small parts of the model that are specific to the question or the prompt that you're giving it. They also have V4 Flash, which is their workhorse model. It's going to be smaller, it's going to be faster, and it's going to be much cheaper. This is 284 billion total parameters with 13 billion active. And if we look at this screenshot, we can see both of them were trained with about 33 trillion tokens of training data.
So, some of the characteristics of these models: they have enhanced agentic capabilities. It is comparable to the state-of-the-art agentic coding models like Opus 47 and GPT 5.5. Literally, the models that were just released in the last week from Anthropic and OpenAI. It has rich world knowledge and world-class reasoning, beats all current open models in math, STEM, coding, rivaling top closed-source models.
All right. So let me show you some of the major benchmarks. Here we have MMLU Pro, which is knowledge and reasoning. We can see here in the dark green bar, this is DeepSk. With orange being Opus 46, purple being GPT54, and then in the stripes, those are the new models, Opus 47, and GPT55. But what we're seeing is, although it is slightly behind here, it's right up there. Okay? And remember, it is a fraction of the price here. GPQA Diamond, same thing. Sweetbench Verified. And basically, what you're seeing across the board is, yes, it is behind, but just a little bit. And that's the real story here. Most use cases, the vast majority of use cases, do not require the absolute frontier-level intelligence. And the fact that DeepSeek is so much more efficient and so much cheaper is actually the problem for the United States.
And so let's talk about cost, because that is really what we need to be scared of. And if you're not sure why, I'm going to explain. Let's look at the cost first. This is AI model price versus performance. On the Y-axis, we have intelligence. Just think, the higher the smarter. On the X-axis, it is the price. The more to the left it is, the cheaper it is. Cheaper is better. And so what you want is to be up here, in this top left. You want to be as cheap as possible and as intelligent as possible.
And so what do we see? We see GBT 5.5, which was just released. At the very top, we have Opus 4.7 right next to it. And I'm just measuring intelligence right now. GBT 5.4 extra high right over here. And then we have Deepseek V4 Pro. A little bit behind, a little bit lower on intelligence, but much, much cheaper. And then look at Flash down here. Certainly a big drop in intelligence. Still really good, but this is an absolute workhorse model price right here. This is pennies per million tokens.
Now, I want to show you how the rivalry between the US frontier labs and Chinese frontier labs has gone over the last few years. So we have GPT4 that came out in May 2023, and we had this massive gap. This is the ELO score in Arena, and then Quen, then GLM4, and at this point, right after 01 preview came out. Remember, 01 was the first thinking model. Right after that, just a few months, Deepseek R1 changed the world and closed the gap almost completely. The US labs did shoot ahead, and there's been this back and forth ebb and flow between them. Every time the US shoots ahead, Chinese open-source catches up. They have always been behind, but that might not always be the case.
And so that brings us to a geopolitical question: Are export controls actually working? Export controls basically means the US, specifically Nvidia, is not allowed to sell its top chips, its best GB300, and a few others, to China directly. Now, there are a lot of rumors that China is going around those export controls and importing them into other countries and smuggling them. And there's an entire story there. We're not going to get into that today, though.
But are export controls working? Assuming that they are actually enforced. Well, the answer is kind of yes and kind of no. Export controls are working because China doesn't have the same compute resources that the United States has. This is just a fact. Even if they're able to smuggle in chips, it is difficult, and they certainly don't have as much compute as we have in the United States. But if they did, imagine what they'd be able to do.
Because on the flip side, the export controls kind of aren't working because they are innovating on the algorithm side. They are coming up with incredible algorithmic unlocks that make training and running inference of these models, including DeepSeek, incredibly efficient. And so even using nerfed GPUs, even using Chinese native GPUs, they're still able to create a frontier-level model.
And in fact, Nvidia, specifically Jensen, has made arguments for selling our top GPUs to them. "China is going to be developing and producing their own AI chips. They should be built on American technology." And that argument is actually why Deepseek V4 is actually such a big deal and such a big threat to the US economy. But just the flip side to it, they're going to make their own chips. They're going to make their own incredible models, and they are going to be very attractive to US companies and our allies. But more on that in a minute.
All right, I want to talk about distillation hacking because it's all related. Just a few weeks ago, Anthropic put out a report basically saying they have proof that the top Chinese AI labs have been distillation attacking them for their Claude model. And what does that actually mean? The simplest way to explain what a distillation attack is: the Chinese AI labs are essentially trying to steal the data from Claude and from ChatGPT. They're asking it questions, getting the answers, and then using those question-and-answer pairs to train their own models. Those question-and-answer pairs are everything. That's the IP of companies like Anthropic and OpenAI.
And just yesterday, the US government put out a statement on distillation attacks. This is Director Michael Kratzios: "The US has evidence that foreign entities, primarily in China, are running industrial-scale distillation campaigns to steal American AI. We will be taking action to protect American innovation." Now, this was already reported by Anthropic a few weeks ago. So, this is not really new news, but the US government actually saying yes, it's happening, is the new part of it. And I'm going to explain why this ties into this overall story in a moment.
These foreign entities are using tens of thousands of proxies and jailbreaking techniques and coordinated campaigns to systematically extract American breakthroughs. But here's the thing. If you look at Anthropic's report, the Chinese labs, and specifically DeepSeek, didn't really steal all that much data. And there is actually an argument that they weren't stealing at all. Maybe it's against the terms of service, but a lot of it can be explained by simple benchmark comparisons. If you're a frontier lab and you want to know how well does my model do against my competitor model, well, the only way to know is to run benchmarks against both. And those benchmarks look exactly the same as a distillation attack.
All right, so this is the report from Anthropic. I just want to very briefly show one thing. The scale of Deep Seek's distillation attack is just 150,000 exchanges. That is not much. Now, Moonshot, the company behind Kimmy, had 3.4 million, and Minimax has 13 million. So certainly, Deepseek, of the Chinese labs, have been doing this "quote unquote" distillation attack far less than the other labs. And 150,000 exchanges is not really enough to explain the level of quality that DeepS has been able to achieve. And then you pair that with the fact that they've open-sourced the whole thing. They have an incredibly detailed and thorough white paper that explains exactly how they were able to achieve it. It just doesn't mesh.
And so, back to our export controls actually working well. Twitter user Jukcon pointed out something very interesting in the report because, of course, like I said, DeepSeek put out a very thorough report. It says, "Due to constraints in high-end compute capacity, the current service capacity for Pro is very limited. After the 950 super nodes are launched at scale in the second half of this year, the price of Pro is expected to be reduced significantly." So they are very compute-constrained. They were able to bake and produce this model, but they can't even serve it in the most optimized way. And they're also charging more than they would have otherwise. So the price is going to continue to drop, and the price is what I want to focus on now.
So why is the price and efficiency of Deep Seek V4 such a big deal? Yes, it is nearly state-of-the-art. Nearly, not quite. It is almost as good as the top models, Opus 47 and GPT 5.5. But here's the thing, it doesn't need to be as good. And just being nearly as good is good enough for almost everybody, including enterprise companies in the United States. And that is what matters.
Imagine you are a CEO of a company in the United States or one of our allied countries, and you're looking at Opus 47. You're looking at GPT 5.5, and you're looking at the costs. And you see GPT 5.5 is $30 per million output tokens. You see, Opus 47 is similarly priced. And then you look at DeepSeek, and it can accomplish all of your use cases because you're not doing frontier scientific research. You're not trying to crack some of the hardest coding problems in the world. You have a business, and you're trying to run your business. And you look at the price, and it is literally a fraction. And you get to control it more precisely. It's open-source. You can fine-tune it all you like. You can make it exactly how you like, host it how you like, and your bill will be a fraction of the size it would be otherwise.
The calculus that these CEOs are making becomes very obvious. Why would you pay so much more for a US frontier lab to serve you their model over an open-source Chinese model? And that's where the problem comes in because more and more US and our allied countries' enterprise companies are going to think about this and make the decision to build on top of Chinese open-source technology. And that's the big argument. Remember, Jensen just had the argument that, hey, China is going to be building their own chips. They're going to be building their own models. They might as well be built on US chips. Well, the same argument is on the flip side with US companies building on top of Chinese open-source models. That is a big security risk for the United States because if Chinese companies decide to change their architecture or cut us off suddenly, we're in a really bad spot.
And so, let's think about this. We have trillions of dollars pouring into the AI industry in the United States. We have infrastructure buildout happening more quickly than any infrastructure buildout in history. So if you have all of this investment that requires a return, and all of a sudden we're not getting that return, there is the potential for the US economy to collapse, especially because we are betting so heavily on artificial intelligence being the future of our economy.
And then think about culturally. Think about how social media changed the world, and social media came from the United States. We were able to control the narrative in a lot of places. Now, flip that on its head. Imagine we're all built on Chinese models, and they're dictating what the models are able to say and what they're not able to say. These are big questions that we're going to have to grapple with if US companies decide to build their AI strategy on top of Chinese open-source models, and they are looking very attractive right now.
All right, so where do we go from here? Well, I think there needs to be two big initiatives in the United States. Number one, we need to go much harder on open source. The frontier labs in the US are not open-source friendly for the most part, maybe with the exception of Google, but Google is building small open-source models, not the same level and not the same capability as a DeepSk V4.
And then we also need to work on efficiency. Even if we are to maintain closed-source and they're being served by OpenAI and Anthropic, they need to get much cheaper, much more quickly, because US enterprise companies need to look at these different models, and it needs to make sense cost-wise. That's going to enable the entire world to build on top of US artificial intelligence.
So, if DeepSeek is doing everything right, Anthropic might be doing everything wrong, at least lately. I made a video about it, so check out the video on the screen right now. People say I went a little bit too hard on them, but I've been really frustrated.