Xinhua News Agency spent 1.122 billion to create the world's most expensive artificial idiot! 150 million US dollars exceeds the training cost of GPT-4—Revealing how "pure corpus" uses diversity collapse and the alignment tax to turn a repeater into a 21st-century cyber version of a horse pulling a train

This article provides a detailed analysis of the absurd phenomenon of Xinhua Net spending 1.122 billion yuan to build a compliant AI model, pointing out that its high costs are mainly consumed by content moderation, compliance filtering, and cloud services, lacking true algorithms and foundation training. Combining the principles of diversity collapse and the alignment tax, the article criticizes the project for degenerating into a cyber version of a "repeating machine" and a "horse pulling a train," and deeply discusses how this incentive mechanism prioritizing politics leads to a waste of resources across society, reflecting the structural dilemma facing the Chinese economy.

Video link:

Full text:

Everything as it is, rejecting fragmented understanding. Today, what we are going to talk about, I need to give you a heads-up in advance, because the absurdity of the whole thing can refresh your understanding of the lower limit of the four words "artificial intelligence." Right on the evening of June 4, 2026, the listed company Xinhua Net, well, that is Xinhua News Agency, released an announcement. What did the announcement say? It said they are going to throw 1.12 billion yuan to make an AI product, and among this, 741 million yuan is money originally used for other things that was temporarily reallocated. So the question arises, what kind of pillar of the state did this fucking Xinhua News Agency buy with its 1.1 billion yuan? This pillar, well, it might just be too heavy. Come, let's look at where this 1.12 billion yuan was actually spent, so everyone can know that the Chinese Communist Party's AI products cannot be called AI products, it's just how expensive the Chinese Communist Party's intelligence can be. Everyone look at this chart, it's particularly funny, because the largest item in its entire project expenditure.

What is it? It is cloud services, a single item of 323 million yuan. Then what about the second and third largest expenditures? One is system construction and R&D iteration, 160 million yuan; another is content construction, 153 million yuan. What does this content construction include? It includes building a corpus, well, hiring a huge gang of people to do content moderation. And what else? Stuffing Party clichés into it, in order to guarantee so-called political safety, you see. Otherwise, if this AI corpus opens its mouth and talks about "Journey to the West," we wouldn't be able to use it, right? So let me sort out this ledger for everyone again, and you will find a particularly magical thing, which is that the money spent basically on content moderation and compliance matters actually adds up to more than the money for its real system R&D. What does this indicate?

This shows that this fucking thing is not a tech project at all; it's a business heavily focused on compliance and auditing, while light on technology. Why am I talking about this today? Don't some people always think that the Chinese Communist Party (CCP) has some bullshit merit in China's economic development? Actually, what does the cost structure and cost focus of Xinhua News Agency's fucking AI reflect? It reflects the waste caused precisely by the CCP factor in the economy. A completely useless thing not only cost 1.1 billion, but inside it are all the compliance costs required by the CCP, and the corpus is also highly closed. Actually, from a macroeconomic perspective, if this cost is placed against the current background of China's economy, wouldn't it hold true for other industries as well? This is like doing something that incurs a lot of investments producing zero bullshit value, and what is the ultimate reason for spending this money? Is it just to satisfy the aesthetic of the Communist bandit? Then from an economic perspective, the cost here is a complete and utter waste. Some people might still say, spending 1.1 billion to make AI doesn't sound too exaggerated, right? Don't OpenAI and Google also burn money? This question is actually very good, but let's do a horizontal comparison: the issue is not the total amount, but what the hell that money was actually spent on. We don't even need to compare with OpenAI or Google; let's just compare it with DeepSeek, which the CCP itself has been bragging about to the skies. Didn't DeepSeek release that idiotic V3 model at the end of 2024? Its report states in black and white that this model used 2,048 H800 GPUs, consumed 14.8 trillion tokens, and what was the final computing cost for a single complete training run?

It was 5.576 million US dollars, which translates to about 40 million RMB. I'll put these two figures together—hey, take a moment to feel this. 40 million RMB, no matter how crappy this piece of junk is, at least it still works. Meanwhile, Xinhua spent 1.122 billion RMB to make a model that doesn't train any base model at all—meaning it has no algorithm. What's even more absurd is this: let's single out that 320 million spent on renting computing power and compare it further. How much did the training of GPT-4 cost? OpenAI boss Sam Altman himself admitted it exceeded 100 million US dollars. However, there are discrepancies because Stanford University's AI Index Report independently estimates it to be around 78 million US dollars. Google's top model, Gemini Ocean, had training compute estimated by Stanford to be about 192 million US dollars. And what was the total investment in Xinhua News Agency's fucking AI? It was 150 million US dollars. Everyone ponder this, ponder it carefully: this money is more than what fucking GPT-4 spent on a single training run, and this expenditure has already touched the threshold of top-tier models. Moreover, what are companies like OpenAI doing? They are opening technical mystery boxes; in other words, you don't know if the training direction is right until you finish training. What is the situation now? Whether in theory or in training path frameworks, things are highly mature. Then why does this fucking Xinhua News Agency have to spend so much money to make such piece of junk? What is a real AI arms race? A real AI arms race is throwing money towards the fucking ceiling of intelligence, while Xinhua's 1.1 billion is throwing money toward the floor of obedience. One goes upward and the other downward; the directions are completely opposite. Therefore, China's AI industry at best will just be the next electric vehicle sector, nothing more. The friction cost is simply too high. Well, now that we've talked about money, let's talk about something more essential. The CCP is just fucking relentless, why? Because they actually use.

Pure corpus to train AI. Using pure corpus to train AI is itself a very anti-intellectual joke that is destined to crash. Let's get one thing straight: why are current large models able to start doing deep reasoning? Isn't it because when a model's parameters, computing power, and most crucially, the scale and diversity of data simultaneously cross a certain critical point, hey, it suddenly unlocks a bunch of capabilities that small models completely lack? For example, multi-step reasoning and logical thinking abilities. What do models like GPT consume? They consume the corpus of almost the entire internet, which contains everything: high-brow papers and netizen bickering, left-wing viewpoints and right-wing cursing, truth and bullshit—a chaotic mix of everything. And precisely from this chaotic and pluralistic data ecosystem, models can compute something approximating intelligence. So what is diversity? Diversity is the nourishment of intelligence. Alright, now this fucking Xinhua News Agency goes against the grain; they want to purify the corpus until only the CCP's fucking propaganda is left, meaning only one voice and only one standard answer. In this case, what will happen to the AI? Ah, let's continue to look...

The first deterministic consequence is definitely a collapse in diversity, which goes without saying. What does that mean? It means that if you train this model indiscriminately with highly homogenized data, it will create irreversible flaws within the model, and the tails of the original data distribution will disappear. This is what we call degradation. Moreover, this kind of degradation occurs even in normal AI under standard conditions. Therefore, major AI companies are racking their brains to solve and fix this problem. For example, things like differential gradient rewards are used to minimize the mediocrity of AI. The Chinese Communist Party is fucking awesome! Right out of the gate, they spent 1.1 billion to handcraft a fucking retard.

What is the disappearance of the tails? Look at this chart here. The wide, fat curve on the left—well, it's normal, it's diverse, and it represents the distribution of the original data. The tails at both ends represent rare, minority information that actually exists. And this sharp, narrow shape is the distribution after purification. All the tails are completely gone; what's left? It's a lonely needle that only repeats the standard answer in the middle. Therefore, the purer the stuff you feed the AI model, the more monotonous it becomes, increasingly resembling what? A talking machine that only repeats the same sentence over and over. Yes, exactly a repeating machine, because this is a basic principle of mathematics, independent of human will. Let me give you an example so everyone can understand clearly. Suppose I now build a quantitative trading tool. What does the real market data distribution look like? It's close to this wide, short distribution state. Well, inside there is noise, emotion, all kinds of true and false information, as well as half-true and half-false news. At this point, I wrote a quantitative program. There is only one parameter in this program, and I also require all trading strategies to be based strictly on this single parameter. Do you think, if I use this thing for trading, what is the probability of me going to the rooftop? Wouldn't it be a fucking piece of junk where you can't even go out at noon because something is bound to go wrong sooner or later? Isn't this exact same logic identical to Xinhua News Agency's bullshit platform?

Then what is the second thing? It's called the alignment tax. Let's look at this. What is so-called alignment? It is using human feedback to bend the model in the direction you want, meaning it is forced into shape by you; it is not the result of natural training. Researchers found a pattern in experiments: the harder you align the model, the worse its performance drops in general capabilities such as translation and reading comprehension. In academia, this has a term called catastrophic forgetting. Everyone look at this curve. The horizontal axis is the intensity of its political alignment, and the vertical axis is its general intelligence capability. These two are completely inverse. The stronger the alignment, the lower the intelligence. Ordinary AI companies practice gentle alignment. To put it bluntly, it's teaching you not to use dirty words or teach others how to make bombs—anyway, stopping at a reasonable point, situated somewhere on the left side of the curve. And what is this retarded thing from Xinhua News Agency? It is extreme alignment. What they want is not a single character of political error, demanding absolute obedience and absolute loyalty, which pushes it directly to the far right corner of the curve. The alignment is indeed maxed out, but the IQ has also hit rock bottom. You know what, it can actually reflect Journey to the West. The only thing it reflects is the IQ of Journey to the West, because whatever aligns with it is equally aligned. Therefore, when you superimpose these two principles—on the left is the collapse of diversity, and on the right is the alignment tax—the conclusion emerges, and it is crystal clear. Xinhua News Agency spent 1.1 billion to create the most expensive retard in the world. This is like spending the money to buy a Ferrari and hiring the top engineering team, only for them to build you what? Well, they built a nuclear-fueled micro-electric scooter that can only travel on a fixed track at a speed of five kilometers per hour without ever making a wrong turn. Yes, that's this thing. So let's stop mocking the Qing dynasty's train pulled by horses; Xinhua News Agency's bullshit gadget is the 21st-century cyber version of a train pulled by horses. It's exactly the same, the logic is entirely identical. Rather than calling this thing AI, it's more like the most expensive Subor repeating machine.

Because of a very simple question: a real reasoning model needs to have basic reasoning skills and basic logical judgment capabilities, right? And Xinhua News Agency's process goes like this: you ask a question, this fucking retarded thing goes to search this pure corpus, and then the wrapped model constructs sentences based on the search results. It's merely piecing sentences together, not generating intelligent answers, and that's after going through layers of semantic review and layers of filtering.

Finally, it gives you that answer emphasizing absolute political security. Look at how many obstacles a fucking AI like me had to cross in the middle, and they were all artificially created obstacles. Every extra safety guardrail it added was like hacking another slash into an already small brain. What was left in the end was just a speaking, politically absolutely correct search box. I don't think it needs to be this much trouble; why doesn't Xinhua News Agency just directly acquire that deprecated Study Xi Strong Country app? Why bother making this thing and spending a fucking 1.1 billion? What's going on here? Do Xinhua leaders not have ad revenue lately, so they started scamming people to trick this fucking brother? Scam if you will, what does a dog biting me matter anyway? I'm happy to watch the show regardless. But this fucking job was done way too crudely; can't you do it a bit more artistically? Just look at this writing and review cost of 153 million; if these guys don't pocket 152 million, I'm a fucking dog with a conscience, leaving just one million to feed all kinds of garbage corpora and qualified texts into it, which I can tell you all wouldn't even be used. So I suggest investigating these fucking bastards at Xinhua; isn't this deceiving the emperor and defying superiors? Moreover, this matter is essentially just the fiscal left hand hitting the right hand. Everyone, take a look: the listed company Xinhua Net raises money from shareholders and the market, and then invests it in a large-scale repeater project led by Xinhua News Agency. This project builds its database entirely using data with the Chinese Communist Party background. Alright, who is this garbage product mainly sold to? It's sold to personnel in over 20,000 party and government organs, to people working in state-owned enterprises, and to these units. Whose money is being spent then? Isn't it still fiscal money? In the end, this procurement revenue turns back into figures in the listed company's financial report. This is the 20.35% internal rate of return disclosed by this fucking model, and the algorithm showing it can break even in 4.1 years. So don't say there is no exploitation within the Chinese Communist Party; grassroots civil servants and public institution workers alike cannot escape it. It amounts to these guys wanting from the start to make the money go in a circle and come back; this is a perfect closed loop of the left hand hitting the right hand. Its so-called market scenario is not built on any market mechanism at all, but rather the resurrection of a dual-track system and administrative command-style procurement. It's just that the group being forced this time are insiders who still have somewhat decent incomes. After talking about the cost structure of this thing, everyone can understand one thing: why state advance and private retreat along with administrative pricing will destroy the economy. Now let's switch perspectives again, what perspective? Let's treat the market as a circulatory system that was originally capable of intelligent self-regulation, but now, some fucking idiot insists on coming in and meddling, so the whole pricing will go off the rails. Well, talking up to here, you might think this fucking thing from Xinhua is already outrageous enough, but actually, under this social background, what I want to tell you is that this thing is actually not special at all, nor is it an isolated effect; this thing is now a logical specimen and a slice of the entire Chinese economy. You take the structure of this slice.

Open it up and you will find a very amusing thing: the play performed in many industries looks exactly the same. Let's sort out this logic first. What does such a long market ecosystem rely on to operate? Actually, just two things: profit and price. Wherever there is demand and money can be made, money and people will flock there. This is the most basic truth. And what nobody wants is waste. The invisible hand of the market is certainly not perfect, but it can at least firmly ensure one thing, which is moving in a direction useful to everyone. But once a fucking hand, for so-called political purposes, slams in crudely, it will obliterate and rewrite the entire set of signals. What happens then? Costs skyrocket—it is a very simple problem. Enterprises from then on no longer ponder how to deal with users, but ponder every day how to get this subsidy money, and will no longer spend on making products better, but on compliance audits and political security, pure, useless life-saving expenses. In the end, recovering this cost relies on the premium of administrative power. Xinhua News Agency is the purest sample of this mechanism; half of the 1.1 billion was burned on compliance and audits. Then, I also saw that this film is still written as under maintenance, which means that in the end, they made a gadget still under maintenance, and it's a search box that no normal person really needs. This is an inevitable result forced out under this distorted incentive environment. When the weight of political correctness is infinitely magnified, what happens to the weight of economic efficiency? An extension of this question is: is the alignment tax only levied on AI? In fact, the same applies to the economy, and its effect will be even more terrifying, because it will force the net return of the entire economy down close to zero, or even into negative numbers. In recent years, has the Chinese Communist Party (CCP) done any fucking fewer things to bleed the economy dry? Let's not look too far back; the closest to us is that throat-gagging, those three years. How many times did nationwide PCR testing gouge out dirty stares? Behind this isn't it still an astronomical-level business? Pouring hundreds of billions to trillions a year into it, until the medical insurance can't even be redeemed normally anymore. The drugs that medical insurance can reimburse for you now are just enough to kill you three times over, as if they are afraid you won't die.

Does an atypical rate hike by the Chinese Communist Party (CCP) get paid for by the people's medical insurance? Isn't this just a degradation of the entire economic system? Why? Because they have to align with the CCP? It's that simple. Then we must ask another question: Are the people in this economy stupid? Or is there a fucking idiot in power in this economy who is constantly collecting an alignment tax from it? The answer goes without saying. Also, look at the

Isn't the situation different now? The Chinese Communist Party keeps slashing all expenditures related to people's livelihoods, and then takes the stolen money to further degrade the entire economic ecosystem, while ramping up propaganda to the maximum. Look at how fucking awesome I am, spending 1.1 billion to build such a fucking piece of shit! I am so fucking rich, I spent 1.1 billion to build a repeater, can you do that? Of course, ordinary people really can't. At the end of the day, what the hell is this thing when you strip away the layers? It's not a large language model at all; it doesn't even have its own basic algorithms, it just puts a wrapper on someone else's work. It's a business of scamming and exploiting grassroots civil servants, just with an AI skin thrown over it. Just think about it, what is OpenAI doing right now? What is Dr. Andrew Ng doing? You can't compare them to this fucking piece of shit, doing so is honestly an insult. Hey, even if you've never eaten pork, haven't you seen a pig run? At least there's still DeepSeek, right? But this fucking thing, 1.1 billion dumped into it, and what is its main job? It's to vet you, to check if you've said anything wrong. At the same time, it's still a repeater currently under maintenance, doing nothing but political censorship work. If this thing can be called artificial intelligence, then my old piece-of-shit tape recorder at home—hey, press a button and it can recite the textbook word for word—shouldn't it at least be called an artificial intelligence entity? Then another day, I'll just drag this thing over, hey, and go ring the bell at the mall. And speaking of the name, Yujie, it's so fucking carefully chosen, sounds so cultured the moment you hear it, as if it's going to enshrine the CCP's political ancient texts word for word. Anyway, my intuition tells me that people inside the CCP system will have work to do now. What is the most likely correct usage of this thing? It's to copy the old path of Xuexi Qiangguo, forcing people inside the system to download it. At the same time, it will probably carry some hidden agendas, like monitoring or surveillance functions. Even if you are inside the system—hey, watching my show—you should keep your guard up. If these bastards insist on making you install it, just fob them off with some crappy old phone, and definitely don't install it on your daily phone. Let's just wait and see if China can break this record for the most expensive IQ test in the future. Only the CCP itself can break this record; nobody else really can, nobody else has that kind of audacity. Alright, that's all for today's broadcast. If you like the show, please click the button below; if you want to support me, you can join the membership channel to watch four in-depth episodes every week. Welcome to join the gang of fierce people, and we'll see you in the next episode.