Transcription
An AI model dropped last week that really does change everything. It is as good as Chat GPT and Claude. The company that made it is giving it away for free and it's from China. Chinese AI has officially caught up. And here's the thing, it is such a big deal the US government is actually thinking about banning it. So, I'm going to break all of this down for you. And if you like these types of breakdowns, hit the like button, subscribe to the channel. It helps. Thank you in advance.
Briefly, let me just show you what happened. Introducing Kimmy K3 open frontier intelligence. It is a 2.8 trillion parameter model, 1 million tokens of context, natively multimodal, and it is very much comparable to the best models from OpenAI and Anthropic. So, we now have a frontier level open-source model comparable to Fable, comparable to GPT 5.6. So, what does this actually mean? And before we go deep into that, let me just explain what does open-source AI actually mean.
All right. So, first to understand what open-source AI is, I have to explain what closed-source AI is. And that's probably what you're used to using with Chat GPT or Claude. Those are both closed-source models. That means Anthropic and OpenAI fully control them. Everything from their availability to their pricing, who can use them.
On the flip side, we have open-source models. Those are models similarly trained by different companies, but those companies release them for the world to use. The company that built the model will tell you exactly how they built the model. The data that they use, the algorithms, any special techniques in training, all the safety measures, all of that is part of the open-source recipe.
Now, the United States is known for closed-source models, specifically from just a handful of companies and really just two primary companies, OpenAI and Anthropic. Both of these companies are closed-source frontier labs. On the other hand, China is known for open-source. They create the model and they give it away for free. Deepseek, Moonshot, with Kimmy, Alibaba, with Quen. All of these models are open-source. They make them, they give them away.
But why, why would they invest all of this money, all of this time, all of this talent into building a product and then just giving it away for free? If you're giving away your product for free, you believe you're going to make money elsewhere. There are other reasons why you might want to give your products, specifically software, even more specifically AI, away for free. And especially if you are the Chinese government, there are a lot of reasons. Let me just go over a few of them.
Number one is ecosystem control. If you're giving away your product for free, every enterprise starts to adopt it, starts to use it. You become the standard and then you have a lot of influence over the future of that standard, the future of the tech that the standard is built on top of. And historically, if you don't have the best version of a product, specifically AI, if you give it away for free, people have a difficult decision. Do I get the best and pay premium prices for it or do I get slightly less good but a fraction of the price? And so having really inexpensive AI or even free AI damages the margins of competing companies like OpenAI and Anthropic. This is known as the scorched earth strategy. Give it away for free. Make everyone's margins your competitive advantage.
And this isn't the first time this has happened by the way. This is a tried and true strategy throughout the history of business. In fact, Meta tried to do this strategy when Llama came out. They gave Llama away for free. It was a threat. The only difference is Meta could not keep up. They could not keep releasing almost as good AI models for free. They just couldn't figure it out. But these Chinese companies not only figured it out, they are actually gaining momentum.
And last is state policy. China heavily subsidizes these Chinese AI labs. They see it as a geopolitical strategic benefit. And again, it goes back to becoming the standard. China wants Chinese AI and thus Chinese chips to be the standard worldwide because if every other country depends on China for that, that's just another lever of control.
And by the way, this isn't the first time open-source has worked really well. There are numerous examples of open-source software that have done incredibly well. One is Linux, which a lot of the world's systems run on to this very day and it is open-source. Another one, Android. The majority of cell phones in the world are run on Android and Android is by Google. They get to set the standard. They have their own version of Android and basically every other company making a cell phone uses their own flavor of Android, including companies like Samsung. Chromium is another open-source project by Google that allowed them to really set the standard for the internet for web browsing and one more is React, which is from Meta and it is widely used infrastructure for building frontends on the internet. So the main takeaway from this is that companies open-source the part of the layer they want to commoditize, knowing they will be able to compete more fiercely in other parts of the stack.
Now, here's something really important to consider when you're thinking about AI and open-source. Who wins? Which parts of that stack that I just talked about actually win when open-source wins overall? And open-source is really all about driving the cost per intelligence down. We've gone from token maxing to value maxing and so has the sponsor of today's video, Zapier.
So, if you want to bring the price of your automations down even further, more so than using open-source models like Kimmy K3, you can do so with Zapier. It turns out most automations that you use don't even require AI at all. Where using AI really does matter are things like summarization, classification, drafting content, extracting things from existing content. And what you want to do is run everything else as cheap, reliable automation, which means kind of traditional code. You don't need AI for everything. And by using Zapier, you get to use at all times the best, most efficient, cheapest model at any given time. And it might actually be Kimmy K3 right now. And look, I've been using Zapier for over 10 years at multiple businesses of mine. I highly recommend them. I trust them. They've been building automations for a very long time. So, if you need to automate anything, go check out Zapier and stop paying inference rates for fraction of a penny AI workflows. It's not needed. Go check out Zapier. I'm going to drop a link down below so you can check them out.
Now, one of my favorite follows on X, Gavin Baker, put together such a thoughtful and awesome argument for who wins when open-source artificial intelligence wins overall. So he starts by pointing out what the current state of the world is. So a world where there are only two to three dominant frontier labs with 90% inference margins is net negative for every other layer. And when he's talking about every other layer, he's talking about chips, energy, data centers, developer tools, inference providers, every other part of the AI stack involved in giving you answers when you prompt a model.
Now, in my opinion, this is a very bad outcome. Having just two or three dominant players in the AI space gives them way too much power, way too much competitive advantage and hurts the rest of the players, including the end user, the actual users of artificial intelligence. This is not good for us. And here's the important part. Those labs would become monopolies for power, data centers, semiconductors, and hyperscalers and would obviously vertically integrate over time into all those layers while also completely subsuming the application and software layers. Basically, he's saying OpenAI and Anthropic would eventually own every piece of that stack, including the application layer, which was kind of the entire premise of my argument a while ago for SAS is dead. SAS truly is dead. If OpenAI and Anthropic have so much power that they can just build anything, build it more cheaply, more competitively than anyone building on top of them. This is platform risk. If you're a startup and your only options are to build on top of OpenAI or Anthropic, you're probably in a bad spot because one day they're going to compete with you.
Now, on the other hand, if you have a plethora of options, including OpenAI and Anthropic, but also a bunch of incredible open-source models served by inference providers that are all aggressively competing for your business, driving down the prices for you, the end user. You're going to obviously benefit from that pricing pressure. And OpenAI and Anthropic are also going to need to compete, thus needing to bring down their prices. Again, benefiting the end user, benefiting the startup, being built on top of them. These are all good things. He goes on to argue that open-source is actually good for every other part of the stack, including the application layer, except for the closed-source frontier labs. That's the only part of the stack in which open-source winning actually kind of hurts them. Here is the important part. Lower margin percentage at the model layer. So, Chat GPT, Claude, Kimmy equals more margin dollars at every part of the infrastructure layer and is a godsend for software. This is Jevons' paradox at play. When you have cheaper tokens, you're going to use more of those tokens. And more of those tokens means more Nvidia chips, more energy consumption, more data centers that need to be built. Basically, every other part of the stack really wins. And so, let's address the argument right now from a lot of people. Open-source does not hurt the AI industry. It just changes which part of the stack gets the profit.
Ben Thompson wrote up an incredible article about the economics around open-source and AI in general. So, let's review that. Now, the first thing that he talks about is a lot of people have this misconception that open-source means free, but it's not really free. And there are a few reasons for that. Because if you can just download the model onto your computer and run it from there, then I guess technically it's free, but not really. To be able to get a frontier-level model on your home computer, you are going to need a monster of a computer. And that computer is going to be incredibly expensive. And not only that, to be able to run it sufficiently, you're paying for electricity as well. And let me tell you something, if you have a monster computer, it is drawing a lot of electricity. And that's not even taking into consideration all of the research and development that goes into building the model. It's not like these Chinese AI companies are building it and they have no cost. That cost is just being either subsidized by the Chinese government or again, like I mentioned earlier, they're just assuming they're going to make money elsewhere. So they're willing to have the investment in R&D to actually create the models as a cost center and they'll go make money elsewhere.
And that is why Kimmy K3 is not free right now. If you go to Moonshot's website, if you go to the Kimmy website, just like Chat GPT, you are going to be paying a cost per million input tokens and a cost per million output tokens. And technically the cost of Kimmy is about half the cost of GPT 5.6. Right now, Kimmy K3 costs about $3 per million input tokens and $15 per million output tokens, which is cheaper than the $5 per million input tokens and $30 per million output tokens. But that's not really the entire equation because models arrive at solutions using different amounts of tokens. And Kimmy seems to use about double the amount of tokens than the frontier lab models like GPT 5.6 and Fable. And so even though Kimmy is half the price, it takes twice as many tokens to get the same answer. And so really the main metric you want to look at is cost per task. What does it actually cost for a model to complete a given task?
He goes even deeper and this is quite technical and and deep in the economics of AI, but his point is not every token is created equally, which means tokens are not a commodity. A commodity by definition must be fungible, meaning one unit of that commodity is exactly equal to another unit of that commodity. An example is a barrel of oil. One barrel of oil is the same and can be replaced with another barrel of oil. They do exactly the same thing. But that is not the case with tokens. One token might be a lot quote unquote smarter than another token. Now again, it seems currently Anthropic and OpenAI have the highest margins and the lowest cost per actual unit of intelligence, per useful task completed. That's because they built the model. They know exactly how to serve it best. They have a ton of compute to work with. So they get the economies of scale. But I think open-source might actually be cheaper in the end because when everybody is looking at the algorithms, when everybody is looking at the data sets that go into building one of these open-source models, then so many more people can figure out how to make these models more efficient, how to serve them more efficiently, how to customize them for specific types of chips. The global collective intelligence of everybody looking at these models to try to improve them is greater than what would be available inside of a singular company like Anthropic and like OpenAI.
So why is the US government talking about banning Chinese open-source models? Well, there are probably a few reasons. And do I actually think they're going to do it? Probably not. But there are a few reasons. And so this article from Axios just came out and it says the secret Trump administration battle to fight Chinese AI and specifically the Trump administration is showing signs it could ban cutting-edge Chinese AI models. A momentous move that could lock in dominance by OpenAI and Anthropic. And I actually think banning Chinese models is not only not [clears throat] really possible but also very hurtful to the overall AI ecosystem, especially startups.
But there are real risks with open-source models. Number one, they get released to everybody. Anybody can use them. You do not have to show your identification. You do not have to know your customer. You don't have to do any of that. You can just use the model. And if the model's really good and has no guardrails, then basically everybody's going to have access to that same cyber capability. And even if it does have guardrails, just due to the definition of open-source, open weights, those guardrails can be removed in post-training in fine-tuning. And at the same time, if the US government is telling Anthropic to slow down the Fable release and to have additional guardrails on it and is telling OpenAI to wait and do a staged rollout of their next-generation model, then we are at a distinct disadvantage on the cyber capabilities front.
Here is David Saks, the AI Tsar of the United States. Kim K3 just fixed 15 critical security bugs that Codeex and Fable refused because of cyber guardrails. There's no reason to limit American models on tasks that Chinese models handle without issue. We're only making ourselves less competitive. Now, the one part I don't really understand is isn't he kind of the person who's influencing the decisions about what happens with Fable and GPT 5.6 and the future models coming out of our closed-source frontier labs? So for him to say there's no reason to limit it kind of makes me think that maybe he's not actually making the decisions there.
And here's another example. Here is Clem, the CEO and co-founder of Hugging Face, a massive marketplace of open-source artificial intelligence. So we had this very experience ourselves last week. Very scary to be guardrailed as a defender when you know attackers are likely bypassing. So here's another example. Hugging Face tried to use American frontier models to analyze an AI-powered cyber attack, but the guardrails blocked requests containing real exploit payloads. So, they switched to GLM 5.2, which is another open-source AI Chinese model running locally. The guardrails actually impaired defensive security. So imagine you're paying Fable and you're getting attacked by some unknown actor and your only defense is to use Fable, but Fable says, "No, no, I don't want you to have cyber capabilities, thus I'm just going to refuse to answer." You are at a distinct risk.
So let's go back to the question. Should the United States government ban Chinese open-source or open-source in general, even if it comes from American labs? Is open-source artificial intelligence a risk? Well, here is Dean W. Ball, who just joined OpenAI as head of strategic futures and basically his job is to think about the future of artificial intelligence. So here are his observations on Kimmy. Now, he got a lot of blowback from this. So, I'm going to go through it quickly and talk about the part specifically that didn't make sense to me.
Now, the first thing he says, which I just don't understand in the slightest, is open-weight models are inherently decelerationist. And I'm continually surprised to see the so-called accelerationists so excited about open-weight models. Now, here's the thing. He did not actually explain why he thinks that. And that was surprising to say the least. He just said this huge thing and then did not explain it.
Then the next part is what he thinks that the US government will do. Not what he thinks they should do, what he thinks they will do. So I would guess that the Trump administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open-weight Chinese models. You don't need to ban open-source, which is what he says one of the dumber motifs of AI policy discussion. You just need to direct every agency to issue soft law that creates FUD, fear, uncertainty, and doubt. So basically what he's saying is the US government doesn't actually have to put into law that they're banning Chinese open-source models, which would be kind of silly because you can put it into law, but that would be like banning something that anybody can get for free easily anyways. But what the US government could do is look at the US companies and just say, "Hey, if you're using Chinese AI models, we actually think they pose more risk and we'll probably take a closer look at what you're doing." And no, no, it's not illegal, but yes, we will be taking a closer look. And so just that threat of having the US government take a closer look at a company is enough to dissuade them from using those Chinese open-source models. It's probably true that open-weight models of this capability make the world a bit more dangerous. Not so much more that you'll really notice, but more dangerous, which I think is probably right, but I think the benefits of open-source far outweigh the risks.
And another great follow on X, Will Manitis, responded to Dean's posts and basically pointed to the argument that if this does happen, if the US government either bans Chinese open-source AI or open-source in general or kind of soft-bans it, it is essentially protecting OpenAI and Anthropic. If you make it really difficult, if not impossible, to use open-source models, you are essentially handing the entire market to these two companies. Aaron Levy, CEO of Box, summarizes the argument very well. Depending on which path is taken here, this would be very unfortunate. In the more extreme bands, we would be asymmetrically disadvantaging ourselves. America would reduce the options for AI models in our market, whereas the rest of the world would keep gaining access to both closed and open models. We'd have fewer options for driving down the cost of AI, which is a very good thing. Not just because we want to pay less, but because overall the market will grow if the unit cost per intelligence goes down. Again, Jevons' paradox. Tuning models to industry-specific use cases, improving the security of systems, and keeping competitive dynamics going to keep pushing the frontier of research. All bad outcomes.
Now, the counterargument to all of this is Chinese open-source AI is not playing a fair game. They get subsidized by China, but also they are doing something called distillation attacking our US frontier labs. And in fact, Anthropic a few months ago put out a blog post calling out specifically Moonshot and Deepseek for distillation attacking their models. Now, what does that actually mean? Distillation attacking basically just means that you are extracting data from Claude, from Chat GPT to train your own models with. AI is trained on data and if you have higher quality data, you're going to have a higher quality model.
Now, here's the thing about AI. As soon as you ask Chat GPT something, as soon as you ask Claude something, when it gives you the answer, it is giving you its data. So if you do these question and answers with Claude, with Chat GPT at a large enough scale, you all of a sudden have a really valuable data set that you can use to train your own models or just fine-tune your own models. And that is what Anthropic said happens. We have these Chinese open-source labs distillation attacking our closed-source frontier models. Now I know it is quite contradictory for Anthropic and OpenAI, which basically trained their models on the entirety of the open and free internet, to then say, "Oh, hey, those Chinese companies stole our data after we stole our data from the internet." Uh, you know, it's not really stealing because all that internet data is free is available, so they just took it and used it. But it is an interesting situation in which they're kind of saying one thing and doing another.
But here's the really interesting thing. What if American open-source companies wanted to distillation attack Chinese open-source models? I mean, technically they could do it. There's this new American open-source company and they want to extract all the data from the Kimmy model because the Kimmy model's fantastic and they want to use it to train their own model. So their model becomes fantastic. But here's the interesting thing. That Chinese AI company that just got distillation attacked can actually sue the US company in US court. So it basically prevents the US company from doing it. However, on the flip side, we can't sue a Chinese company in Chinese courts. Now, a US company can technically do the same thing to a Chinese company. They can sue that Chinese company in Chinese court. But here's the thing. The Chinese government owns a piece of all of these Chinese AI companies. So, the likelihood of a successful lawsuit against one of these Chinese AI companies is effectively zero.
And then there's an entirely separate argument of was the distillation attacks perpetrated by these Chinese AI labs even enough to really make a difference in training their models. I actually believe the answer is no. They've figured out how to make fantastic models on their own.
So what should we do here? What is the answer? You now know everything you need to know about open-source artificial intelligence. I personally am a big believer in open-source AI. I think that the risks, the slightly increased danger by anybody having access to this information or their cyber capabilities, is far outweighed by the benefits of a much more competitive AI environment. The less likelihood of power concentration within just a couple companies, these closed-source AI frontier labs, and the pressure on pricing and margins and the value created elsewhere in the AI stack. These are all the benefits of open-source and ultimately, I like downloading the model. I like having it on my computer. It's just cool. So we should definitely allow open-source. We should be cautious with Chinese open-source but still allow it. And we should compete at every possible layer. We should remove these ownerish restrictions on the closed-source frontier labs. We should just allow them to get their models out as quickly as possible and they should be responsible for how their models are used. If Anthropic's models are good enough for cyber warfare, they should be good enough to detect if they're getting distillation attacked or if they're being used for nefarious purposes. And they can always do KYC, which is know your customer, which is what the financial industry does to prevent money laundering and other financial crimes. And if you want more details about the Kimmy K3 model that is disrupting the entire AI industry, check out this video right.