La Xina tancarà els seus models d’IA? Què diu realment el rumor i què està canviant

Let's talk now about the bad news, friends, coming from China regarding artificial intelligence. And we have been very happy and content for a long time thanks to the open models that have come from China. The United States also has open models like the ones Meta has released, but especially in China this has been a lifeline not only for users like us, but also for many companies. There are companies that have found this helpful in fine-tuning these much less expensive models, thus saving a lot of money. Even Shopify itself has switched to using Queen 3 instead of Open AI. Here's Shopify using Qen 3 to cut inference costs by a factor of 75. In other words, he was paying millions of dollars and now he's only paying thousands of dollars because he switched from using the OpenAI API to using Quen 3 and is saving a lot of money. So, the bad news, folks. It turns out, folks, that the people of China might be considering limiting foreign access to their most advanced artificial intelligence models. The original news, reported by Reuters, states that Beijing is considering whether it could block foreign access to its most powerful Chinese models. This is according to a source who says that there have been meetings between Alibaba, B dance, and Zi with Chinese authorities, and some officials suggest that the law should be much stricter with people who steal artificial intelligence R&D because it is a matter of national security, and that the new restrictions will also address who can use or invest in domestic startups. So, for example, the United States cannot invest in a Chinese company. And this has already actually happened because you don't know that Meta had bought Manus. Manus china and now China has blocked the operation. So you can see that it seems China wants to continue blocking this. This comes somewhat in response to what has happened with the United States. The United States has for some time blocked access to Fable and GPT 5.6 artificial intelligence models abroad, but mainly with an eye on China. And what's happening on the other hand, in China, China says, is that, of course, what's happening is that our models are open and people, companies in the United States, are benefiting from that. And not only that, but there are also problems regarding national security, they say, right? They say, wow, even

if we're giving them this advantage by getting closer, it could happen that this advantage we had disappears, right? So what they want is to cut back a bit, because of the brutality, right? Oh, watch out, be careful with this because I've been looking into this a lot since this news seemed a bit extreme to me. I'm telling you it's a " could," in fact I've also seen it on Time, I've been reading it. See? China could restrict access to its most powerful artificial intelligence models , meaning that it might allow some, but not others, I don't know what, I don't know how much. Officials are talking about criminalizing any leaks or thefts of already patented technology under China's strict national security law, according to one of the sources. They also raised the possibility of implementing new measures to restrict who can finance domestic startups, the source added. The extent of the possible restrictions is still being debated, according to two sources who added that they might only apply to future models. It was not immediately clear when it would come into effect, or even if it would. Okay? So, they've been asking a lot of people in all the media if this could be true, if it's not true, I don't know what. But nobody has answered, not Alibaba, not Bite Dance, not ZI, nobody has answered about this. Nobody says anything. So I've been really looking into where this is coming from. I say, "Holy crap, but where is it?" Because I'm telling you, it's not a sure thing. Even on Reddit, when they're talking about this topic, when they're discussing this topic here, look at Reddit, we have it here, what they're saying is that the news isn't true. Beijing says it is not looking to restrict access. He says this is like they're saying this is a lie. Because? Because it turns out that what Reuters says, according to them, is based on this document. So, Midu, how do they manage to steal models? Some kind of social intelligence? It's not about stealing models; what they do is use another model to train their own. What they do is use another model; what you do is distill your own. And so what you're doing is like stealing their intelligence. It does n't steal his intelligence, but if you make enough calls, what you can do is kind of copy

how he behaves in different situations to understand his train of thought. This article, which as you can see is all in Chinese, is actually a document that comes from the Chinese government department , go. Okay? And in that document, precisely what it talks about is, they were talking about this, uh, it's not that it's talking about blocking access to Chinese models, but what it was commenting on was more about the acquisitions that some Western companies were making in China, foreign investments and things like that. Uh, so nothing, it's more or less about this. Likewise, although they say this, what is happening is that on Twitter there do seem to be rumors, and there are rumors from good sources, that many artificial intelligence laboratories are commenting at some conferences that they are going to pivot to closed-source models. These companies will face pressures they simply won't be able to withstand. Now is the time to build a true open-source AI from scratch. It is time to reclaim our sovereignty. Well, Jun says this , but he's not the only one. Notice that, well, there are already some, some like Ryan Lee. Ryan Lee is the Head of Death Rel of Mini Max, one of the most important Chinese models, who has already said, "Look, at least at Minimx we are going to continue publishing frontier open weight models, that is, the quite powerful models." But I can tell you that at least on Twitter there are quite a few rumors from different people saying that in the corridors it is being discussed that there are Chinese companies that are already warning different companies that they are going to close the model, that is, their model. I'm going to say what I think will happen because I don't think they're going to close it 100%. I believe that what is happening is that there are many American companies taking advantage of the fact that these are open weight models , which they can use freely for commercial purposes, and I think that is what the Chinese side does not want to happen. In other words, what they don't want is for it to be used for financial gain. So, I think the models will continue to be open, but there are already Chinese models that are actually, because look, open Chinese models, but that are commercially restricted. Let's ask HG GPT, but there are already

some Chinese models that are actually open for home use, but commercially, of course, they're kind of open source, but with... Ah, no, rather, I'm not saying [snort] that you should look for them, for example, Kimik 2.7, if I'm not mistaken, has this license, a license that tells you it's open, but for commercial use you have to ask for permission and you have to pay. And this, I think, is something that we're going to be seeing more and more of. I don't think they'll stop being open because I think that's their strategy to reach more people, but I think what will happen is that they'll limit them for commercial use. I'll show it to you when I'm finished, while we look at other things because I have to tell you about this Minimax thing. Minimax has said that they will remain open and so on. Well, Minx, among others, has secured a super important investment. This is according to Skyler Miau, who is the Head of Engineer at Minix AI. And a letter written by the CEO says that markets will fluctuate, external noise will come and go, but our direction remains unchanged. Being at the forefront of this industry, we have a clear understanding of the pace. He says, "From today until the day we reach AGI— they're already thinking about AGI—I will no longer receive any compensation from the company; that is, my salary will be eliminated . For the next four years, I will allocate shares equivalent to 4% of the company's total capital from my personal holdings to reward team members who choose to build this journey with us long-term and create value together." Damn, not bad at all, huh? This has surprised me. This has surprised me because it's something you don't usually see . I will allocate 1% of my shares to establish a dedicated fund that supports continued growth in open source communities and the broader AI ecosystem. I will dedicate all my time, energy, and resources to this mission. This is my long-term commitment to the founder, to our company, our team, and the future we are building. And that's because Minx has secured an investment of millions and millions of dollars. In other words, it's not bad. And what is this going to bring? Well, the rumor is that Mini Max is planning to launch a 2.7 trillion parameter model, which would be, I

think, the largest, if I'm not mistaken, the largest Chinese open-source model. If it's not the biggest, it must be one of the biggest. 2.7 trillion parameters, open source model and would be launched in Q3, that is, it would be launched in September. This would be it, you see? larger than any other model and would be six times larger than the Mini Max M3 that just came out. And this has been discussed among employees, within Mini Max. Look, these kinds of rumors usually blow up pretty fast , but imagine a model at that level. Minimx 2.7, any commercial use requires authorization. To use it commercially, you must apply for another license. Bauchan, I don't know what. Well, they're old, but you see, Kimi says yes. Allows normal commercial use. Normally, a restriction applies when the product exceeds 100 million monthly active users or $20 million in monthly revenue. Sure, for very large companies. You know that all these models really have fine print when they say they are open models, that they really have a little bit of slack, that you have to see, we do n't, because we don't have so many millions of users or earn that much money per month, but it is true that sometimes you have to be a little careful, but this doesn't stop because we already have the next models of Quen. There are rumors that we'll get Quen 3.8 this August, remember it's currently at 3.7, and Quen 4 would arrive in September. Quen 4, well, 3.8 would already be better than GLM 5.2, but the problem is there are also rumors that GLM, which is currently the most powerful Chinese model, is on par with Opus and such, okay? They speak very highly of this GLM 5.2 model, which is the new one and has some amazing designs and all that. Well, it seems they're about to release version 5.5. In August they're going to release GLM 5.5, that is, the speed at which they're coming out. Then DeepSek is going to release its version 4 this week, the final one, because as you know, the one that had been released until now was the preview version. In other words , we're going to have a really hot summer . And then, also, recently, in case you did n't know, if you missed it , Long Cat has come out, which maybe you have

n't heard of at all, but I'm going to tell you about it because it's going to blow your mind on many levels. Long Cat 2.0 was released, another Chinese model that is currently one of the largest models released, with a total of 1.6 trillion parameters, okay? 1.6 trillion, which is quite a lot for this type of model. We're already talking about very, very large models, which obviously aren't meant to run the business, but I'm surprised that they're at least making them open and that they can train on this. So what's new about this model? Why is this model important? Besides how big it is, size isn't everything. You can rest assured. It's for several reasons. The first is due to the comparison with the other models. Notice the long cat that's at the same level. It's a shame he doesn't mention names, okay? But it's there, oh yes, it says so here. The legend is down here, sorry. But here we can see that it is obviously below opus 4.8, 4.7 and so on. In some cases it's there, there, okay, but it's a model that's on par with the big models. But the most amazing thing about this model is that it has been trained 100% with Chinese cards. I mean, if you really look into how they've been doing it and all that, it turns out they've been using all Chinese technology. I mean, it surprised me, I mean, wow, well it's quite interesting because I thought , oh, well no, here it says that you're made with the, well, this is a lie, right? Because it says here that if this is how it has been trained, deploy. Ah, no, this is how to run it, but how it was trained, because that was one of the news items, it was trained using completely Chinese hardware chips and not with Nvidia processors. Of course, that's what I had read in the news. had used Huawei cards. So, that was the big news, that it was one of the first super-progressive models that had been trained entirely with national chips with a 100% Huawei card. In other words, I hadn't used anything from Nvidia. Sure, it surprised me, but the thing is, I've seen this here, the B300 and all, but what it says here is how to run it, okay? how to run it. But it turns out that

this Long Cat is the first model that has been trained 100% without Nvidia chips, using only domestic chips. Things are coming, friends, things are coming. How this is slowly evolving at a crazy speed, and I don't know if it will stop with all these rumors about bad news from China about them closing the models and so on, but in reality, suddenly, at least right now , what we are seeing is the complete opposite. More models are coming out than ever before, more models are coming out than ever before. This Longcat 2.0 is coming out, and the most beastly version of this model... the website is awful, it's this one . This is the official website, okay? This is the official website, but the worst part, look, here it is, is that the training and inference was done with 50,000 domestic cards. So , there you have it. This is the official website of this model that makes you say, "Holy crap, they didn't even use their own artificial intelligence to make it." But it's not even accessible. But there's something even worse about all this, and that's who 's behind this model, the Longcat model. That's the most brutal thing. Well, it turns out that behind this model is this company here, mayuan.com. And this company does more than just that; it's famous for operating a platform of local services, including food delivery. I mean, friends, the largest domestic card-trained model in Chinese history, it was done by the people who would make a balloon. An Uberit does more things. Ultimately, it's like an intelligence agency, I mean, it's a technology services company , okay? It has more things, hotel bookings, instant e-commerce, it has more things like a rapi, okay? I mean, imagine that the Chinese delivery service has trained the domestic model even more, I mean, it's pretty amazing, which makes sense because in the end, obviously all these kinds of companies are going to use artificial intelligence massively, internally, and they won't want to depend on any external ones, and they have to have a lot of processes to improve robotics within their companies and so on. But, well, it's still surprising that it's not merely an artificial intelligence company , but a technology company with services, for example, food delivery, that is getting into the world of artificial intelligence. This is just the beginning, friends, just the beginning,

that more and more companies are getting involved in this kind of thing. And here's one of the things I want to share with you so you don't lack work, because I've had a vision and I want to share it with you. Okay, the business of the future, if you're a programmer, I'm going to tell you what I see as the business of the future so you won't lack work. and I shared it. A business of the future for many programmers is helping companies migrate from OpenAI and Anthropic to open source models because the savings can reach 80%. So, you go to a company that's spending millions of dollars a year and you have the know-how, you know how to make [music] step by step and migrate business logic processes from a very expensive model like OpenEI Anthropic and you migrate it to an open source model, I'm not even saying Chinese, okay? But open source, watch out, be careful, because there's a mistake here that many people make, which has amazed me. Look, now in the chat someone else says, "But the hardware is extremely expensive." But let's see, a lot of people here, a lot of people were telling me the same thing. The value of hardware is that hardware is I don't know what. Using open source models is one thing, having your own infrastructure is another. They are two completely different things. And I'm going to explain why. The point is, for example, that in Asure, to name one , because there are more, Azure actually has open source models, let's say Deepsic, to name one, you can use Deepsic V4 Pro and V4 Flash from Asure's servers, that is, Microsoft's, without needing to use the Chinese Deepsic servers. And let's also consider that it's a European company, well look, the Poland region can be on servers in Europe as well as South America, everywhere, look, in Italy, I mean, in Europe, in the north, in the south, I mean, in different places, they don't really have to be, look, you also have Kimi, eh, because maybe they say, "No, it's Kimi." In other words, it's true that having your own infrastructure can be profitable or not, but it doesn't have to be that way. And why do I say this is going to be a job? Because here's the second problem. Many people tell me, "Well, sure, but the

CHGPT and Anthopic models are better." Of course , but look at the models, for example. And this is an API, huh? What I'm showing you here at the end, this is an API that you use like you would use the OpenAI API . For example, here you have Kimica 2.6 and the price you have here is perhaps five times lower than the price of Open I Anhropic. What's going on? The problem is that when these models are worse, you usually need to manage them better; you need to have a better harness to better control the agent, give them better context, and better control the prompt. In other words, many times what you have to do is control the model better yourself. Things that in OpenAI and Anthropic might either happen by magic or have already been trained in a certain way, or they have more ease in many cases. In these kinds of things, you have to direct it much better, okay? In that case, once you've done that—and that's the beauty of it, that's where your work comes in—your work isn't simply changing a string of text, but hey, I'm surprised that there are people who don't even know this, and that you can also do fine-tuning on all of these, okay? There are models that you can fine-tune, and you have, for example, look, this is what Shopify has done. What Shopify has actually done is not to have its own infrastructure, but rather to fine-tune Queen 3 to avoid using GPT 5.4 and has saved $20 million, that is, per year. It's just barbaric. Instead of firing, well, how much is that? Whether it's 200 people or 2000 people, what they've done is migrate this and they've saved a lot, a whole lot of money. So it makes perfect sense. It makes perfect sense. And how do I know this? I know because I've been talking to some companies who have told me their invoices. They told me, "No, it's because we're paying X." I'm not going to say the name, but there's a fairly famous e-commerce company in Spain that I've been told was paying $60 million a year for artificial intelligence. That includes everything related to subscriptions, all the business logic, all the processes; it's like the complete bill, but I said, "But what are you saying?" He says, "Yes, yes, because then, well, he explained the

infrastructure to me a little , blah blah blah blah." And I say to him, "Haven't you thought about using Amazon Bed Rock?" Because I get the feeling, of course, because in the end they only used Open and Anthropic. And I say to them, "Haven't you thought about using something like Amazon Bedrock, which is essentially similar to Microsoft Foundry, to use the models you need only where it makes sense?" Because a lot of money, right? Oh, I didn't even know it existed, did I ? It's just that I don't know what, I don't know how much. And I, dude, like, and it all makes perfect sense . And besides, you certainly don't need borderline models in all processes . I believe that people who have this knowledge, who are very clear about it, can go to a place and say, "Hey, look, I can save you 2 million dollars a year." He wo n't be short of work. How much are they going to pay you? Imagine you get paid $100,000 a year. That's because you're still saving him $900,000 a year. I think it's amazing . So there you have it. I'm telling you, it surprises me, so you can get an idea of why it's important, why I think it 's interesting. It's because there are many people who, when you tell them to use open source models, notice that many, many, many people, many people here, many, many people take it for granted that you see, you still have the problem that the hardware is at prices I don't know what, I mean, everyone takes it for granted that it's your infrastructure and it's curious, huh? Because then it's like, wow, there are people who don't even know it, that's why what happens happens. And that's why, look, someone was saying, "They've invested 60 million in artificial intelligence and haven't been able to invest a million in getting advice." That's awful, Guikadep, how many times has this happened? For example, as I mention in the post, this has already happened in the cloud. These eyes have seen this. I mean, it already happened in the cloud, first everyone went to the cloud like crazy, then they saw the bills and said, "Damn, we have to optimize this, right?" Or, "How many times has this happened with cybersecurity?" Oh, I mean, these things happen every day, totally. he [music] [music] oh