Saturday, June 27, 2026

Anthropic has repeatedly sued domestic manufacturers for model distillation, coupled with the U.S. government personally issuing access permits to GPT5.6 and Mythos models, and even wooed the European Union to form an AI circle. This shows that AI is no longer a toy in the geek circle, but a practical strategic weapon. For us ordinary developers, the biggest pitfall is the risk of technology outages. Don't dream of being able to rely on shell or ash distillation of top overseas closed-source models. Compliance and bottom-level self-control are life-saving talismans. A lot of model hot-ups must be made in the business architecture. Don't put eggs in a basket that may be unplugged at any time.

Looking back at embodied intelligence and autonomous driving, Da Xiao's robot dog has begun to patrol the field around the clock. The ideal car is stuck with Tesla FSD to create a closed-loop data computing power. The country is also eager to come up with standards for intelligence and embodied intelligence. This sends a strong signal that the industry has moved from rolling PPT and parameter quantities to rolling project landing and data flywheel deep water areas. The era of creating a cool Demo to scam financing is over. Now the battle is about who can truly run the closed-loop hardware, models, and data in the dirty and tiring physical world. The engineering pitfalls are all in the long-tail scenario, which cannot be solved by adjusting a few APIs.

These two trends are actually two sides of the same thing. The upper level is engaging in technical blockade and ecological fragmentation, and the lower level is fighting to break through the closed loop of engineering in the physical world. For technicians, focus less on those fancy shell applications and more on how to reduce reasoning costs, how to do multimodal data cleaning, and how to actually stuff Agents into robot dogs and cars. This is the hard currency that will help you secure your job in the next five years.

Today, these signals spell out a very clear picture: AI is moving from "model arms race" to "infrastructure sovereignty war" and the AIization of the physical world is accelerating simultaneously, with a hidden causal chain between the two.

Let's look at the core conflict first. Anthropic sued four China AI companies one after another in four months. This time it was Ali's turn, citing "model distillation." This is no longer a pure intellectual property dispute. Combined with another signal-the United States is opening up the Mythos model to some companies, and OpenAI announced that the release of GPT-5.6 Sol will be up to the U.S. government to decide who can use it-you will find that AI models are being explicitly incorporated into the framework of national security. Models themselves have become strategic resources, the open source versus closed source debate is outdated, and the real question is: who has the right to use the most advanced models and who has the right to decide who can use them. The name of the Qwythos-9B model on Hugging Face speaks for itself-it stitches Ali's Qwen and Anthropic's Mythos together, an instinctive response to lockdowns in the technology community, but also a gray area with extremely high legal risks. Half a year later, I think "model distillation" will become a standard diplomatic issue, similar to today's chip export controls.

The second thing is that the integration of AI and the physical world is accelerating, but it exposes a serious crisis of trust. The signal of "AI faking the World Cup" is very typical-the beauty in the stands is crying, but the traffic is real. This is not just entertainment fraud. It belongs to the same type of problem as the bean-bag watermark appearing in the illustrations of Lanzhou University papers: the proliferation of AI-generated content is systematically eroding the credibility of public information. When academic papers, news sites, and sports events can be forged at low cost, the cost of verification for the entire society will soar. On the other hand, robot dogs patrol the west coast of Shanghai 7x24 hours a day, Ideal Car develops its own batteries and uses the "bottom" strategy to counter the brand lock-in of the Ningde era, and Zhiyuan Robot's clever hand is valued at US$1 billion in five months after its establishment-all these are telling you that AI is gaining control of the physical world. On the one hand, the authenticity of the digital world is disintegrated, and on the other hand, the autonomy of the physical world is enhanced. This asymmetry will create a huge contradiction: we are increasingly relying on AI to manage reality, but it is increasingly difficult to judge whether the reality presented by AI is real.

The third clue is hidden in a seemingly inconspicuous place. Apple raised prices by 20% across the board due to soaring storage costs. Why are storage costs rising? Because AI training and reasoning require massive amounts of data, global demand for memory chips has been pulled up, and supply chains are being restructured. The United States wooing the EU to engage in an AI semiconductor supply chain security alliance is essentially redividing the ownership of chip production capacity. This directly affects the daily life of developers-behind the posts on V2EX that discuss "should we run" and the complaints in the developer community about the decreasing number of copper coins are anxiety about uneven distribution of technical dividends. When large companies and countries compete for AI sovereignty at the top, what ordinary developers and small and medium-sized enterprises feel is rising costs and narrowing opportunities.

My judgment is that we are entering an "AI Berlin Wall" era. In the coming year, the AI technology stack will be cut into two or even more parallel systems from chips, models to applications. This is not a prediction, but what is happening. Anthropic's lawsuit, the U.S. review of model users, and the European Union's supply chain alliance are all bricks building the wall. On both sides of the wall, there will be their own "high road"-blocking, distilling, raising prices, and looking for alternatives. The harm this division has on technological evolution itself may take several years to fully manifest itself.

Looking at today's noisy science and technology news, all I smell is the gunpowder smell of geopolitics and the sour taste of bubbles fueled by capital. The so-called AI technology revolution is becoming a land enclosure movement jointly managed by a few giants and the government.

The U.S. government now has to personally decide who is qualified to use GPT-5.6. At the same time, it also opens Anthropic's Mythos model to some companies, and also pulls the European Union into an AI partnership. Don't think that this is technological universal benefit. This is naked technological blockade and camp division. When model distribution rights are transferred from companies to governments, AI becomes a strategic weapon. The EU's fear of being kidnapped by the United States is by no means a worry. They have seen through the colonial logic behind this partnership. In this game, all non-American technology ecosystems pay the bill, and what is lost is the diversity of technology open source.

In China, the background of the prosperity of big models is even more questionable. Anthropic has sued four leading China AI companies for infringement in four months, pointing the finger at model distillation. Look at the names on HuggingFace that blatantly copy Claude and Mythos 'open source models, and you will know how much self-research is. Many companies use investors 'money to do the business of adjusting APIs and washing data, forcibly packaging technical barriers. Once the bottom-level manufacturers turn off the taps legally and technically, the valuations of these companies that rely on distillation to survive will instantly return to zero. Who makes money? It's the PPT teams who tell stories. Who is paying the bill? They are corporate customers who are superstitious about domestic substitution and takeovers in the primary and secondary markets.

What is even more ridiculous is the chaos at the application level. Using free AI to help more than 10 million candidates fill in their volunteers is simply a joke on life. Voluntary filling is a high-risk decision with extremely low fault tolerance rate and requires dynamic games and extremely accurate data. Nowadays, AI can't even understand the bean bag watermark in a static paper illustration, and can't even guarantee the most basic authenticity. What can we use to predict the admission probability? Packaging this immature technology into a free tool to grab traffic will ultimately destroy the future of ordinary families. As for using AI to fake the World Cup with beautiful women crying to swindle traffic, it is just the tip of the iceberg of technological evil.

In today's AI circle, truly hard-core breakthroughs are concealed under endless shells, fraud and policy games. History tells us that in any wave of technology, those who ultimately survive are people who can solve real problems and have the ability to independently produce blood, rather than those speculators who take advantage of the situation.