Saturday, July 11, 2026

Today's most explosive signal is the simultaneous launch of GPT-5.6 and Grok 4.5. OpenAI and SpaceXAI are playing a clear card. GPT-5.6 is directly called "Opus level", and the pricing strategy has also been released, indicating that they are determined to compete with OpenAI for the high-end market. But if you look at the developer community, V2EX is full of posts promoting 5 yuan gifts, indicating that everyone is still using transfer station casings, and the scene of actually using these models for production-level applications has not broken out. The most important question ordinary developers should ask now is: Do I need to chase this version number? It is enough for 90% of people to use the GPT-4 level model, so don't let marketing lead you to waste money. Ali completely banned Claude and switched to Thousand Questions is a really tough move-forcing internally to eat his own dog food. Although this move is crude, it is indeed the only shortcut to breaking the product situation. You let your employees use competing products every day, and your products will never catch up. However, the risks are also obvious: if Qianwen could really fight, Ali would have taken the initiative to push it, but now relying on administrative orders shows that the team does not have enough confidence in its own model. In the short term, it is a broken arm to survive, and in the long term, it is a bet on the national luck. However, the cost of this bet is too high, and employee efficiency is inevitable to fall.

Another line worth digging deep is that the "de-Nvidia transformation" of AI chips is accelerating. France's antitrust investigation into Nvidia is coming to an end, and it may be possible to impose a fine of 10% of global revenue. This is no joke. At the same time, Lingruizhi has integrated hundreds of millions of high-performance RISC-V CPUs, and has created the first RISC-V core in China that supports SMT4, and it also says it will be mass produced. In addition, Byte and Nubia have joined forces to develop "bean-bag mobile phones", and Step Star is also going to launch AI smartphone phones. These signals, taken together, mean that the entire AI industry chain is shifting from "Nvidia alone" to "Diversified architecture + terminal-side reasoning". Don't think that RISC-V's performance is not even better than ARM and x86, but in scenarios such as AI reasoning that require high power consumption and customization, RISC-V's flexibility and open source advantages will be amplified. Lingruizhongxin even has a GPU company like Biren Technology in this round of financing, indicating that veteran players are also looking for spare tires.

Finally, Google requires AI ads to be labeled and NYC prohibits fraudulent subscriptions-these regulatory signals indicate that too much garbage has been mixed into the AI bubble. Tools that rely on AI to write resumes and generate content (such as the ai-job-search project) may seem useful, but they still treat job-seeking anxiety as business. What you should really be wary of is what said in the article "Medical AI Miracle": AI is packaged as a lifeline, but in fact it is just information integration, and may even replicate a bad blood scam. The technology community is accustomed to "releasing it first and then repairing it", but when it comes to life and health, bragging can kill people. Don't be deceived by demos and warm stories. Before landing, ask: What real problems can this thing solve? Where is the pit?

Today is July 11, 2026. From these signals, I can see two clear lines intertwined: one is that the AI industry is accelerating the construction of its own walled gardens, and the other is that AI agents are changing from conceptual tools to infrastructure., and began to trigger a rebound in regulation and trust. When these two lines come together, they will determine the industry landscape in the coming year.

Look at the walled garden first. Ali has completely banned Claude and vigorously promoted Thousand Questions. This is not a simple internal management, but the ultimate version of "eat your own dog food"-when your product is your employees, you can get the most realistic feedback loop. At the same time, GPT-5.6 was officially released, SpaceXAI's Grok 4.5 was also claimed to have reached the Opus level, and the arms race at the model level has entered a white-hot stage. But what is more noteworthy is that these models are rapidly being "hardwareized" and "terminalized": the second generation of bean bag mobile phones is exposed, the AI mobile phone of Step Star is about to be released, and the Chery Fengyun A9 is equipped with a 3nm cockpit chip and AI Lingxi Intelligent cabin 2.0. This means that AI is no longer just an API in the cloud, but is embedded in physical devices and becomes the entrance to daily interactions for users. France's antitrust investigation into Nvidia and Oracle's downgrade due to AI infrastructure all indicate that this trend of vertical integration is triggering regulatory nerves-whoever controls chips and terminals controls AI traffic and profits. On the other hand, Lingzhixin has completed hundreds of millions of yuan in financing to promote mass production of high-performance RISC-V processors. This is China's bet on open architecture at the chip level, trying to break the monopoly of Nvidia and ARM. In the next six months, we will see more "AI-native mobile phones" appear, but we will also see intensified geopolitical frictions at the chip level.

Let's look at the implementation and risks of AI agents. A number of new models appear on Hugging Face today, such as ThinkingCap-Qwen 3.6 for image text, HF Realtime Voice that supports voice, and Agents-A1 for agents. In the open source community, the ai-job-search project allows Claude to send and change resumes for you, openwiki automatically maintains code documents, and strix uses AI to automatically find vulnerabilities-these tools are no longer demos, but productivity that can truly reduce costs and increase efficiency. But the risks are equally obvious: the EU's adoption of Chat Control 1.0 means that all communications may be scanned by AI; warning articles about AI medical scams are being splashed, reminding us that when AI is packaged as a "miracle", the cost of failure cases is real. New York City's ban on fraudulent subscriptions is also a continuation of algorithmic manipulation. The "radius of trust" of AI agents is expanding, from changing resumes to writing code, to controlling budgets and conversations, but the boundaries of trust must be clearly drawn. Only products that provide transparency and controllability at the same time and do not rely on a single closed ecosystem will have long-term vitality.

I noticed that several signals can be strung together to tell a larger story, which is related to the current power swing in the AI industry.

The European Commission approved Chat Control 1.0, which is actually more alarming than the release of any new AI model. The bill essentially legalizes backdoors for encrypted communications and promotes large-scale surveillance in the name of "protecting children." Similar incidents in history, such as the confrontation between the FBI and Apple over unlocking the iPhone, ended in concessions on privacy protection. But this time is different. Chat Control is directly targeting the communication infrastructure of the AI era. If AI's underlying capabilities are used for content review and monitoring, the so-called "AI democratization" is a joke. What's even more ironic is that on the same day, OpenAI launched GPT-5.6, known as an "Opus level" capability. On the one hand, there is explosive growth in technological capabilities, and on the other hand, regulators are using 20th century methods to meet 21st century challenges. Looking at these two things together, the answer is already obvious whether future AI applications will move towards "free innovation" or "compliance cage".

Looking at Oracle's credit downgrade, this is the real bubble signal. Standard & Poor's downgraded Oracle on the grounds of structural risks posed by investment in AI infrastructure. Oracle's fiscal year 2027 capital expenditures are expected to be as high as $90 to $95 billion, and the free cash flow deficit could reach $42 billion. This is not unique to one company. Nvidia, Microsoft, and Google are all spending money crazily. But we have seen similar stories too many times in history. Before the Internet bubble burst in 2000, telecom operators frantically laid optical fibers and finally went bankrupt. In the cloud computing bubble around 2015, data centers were first built and then insufficient utilization was discovered. The current investment in AI infrastructure is essentially repeating this story. Who makes money? Of course it's Nvidia and its GPUs, and those who sell servers. Who is paying the bill? It is the boards and shareholders who are intimidated by the narrative of "not investing in AI or falling behind".

Ali has completely banned Claude and switched to Thousand Questions. Behind this signal is deeper anxiety about "supply chain security". On the surface, it may appear to be a product strategy of "eat your own dog food", but combined with France's antitrust investigation of Nvidia, this is actually a defensive retreat. If both Europe and the United States begin to regulate and restrict AI infrastructure, China companies must ensure that their AI capabilities do not rely on external models and chips. Ali's ban on Claude is not so much a product upgrade as a political risk avoidance. But the risk is that if Qianwen's actual capabilities fail to reach Claude's level, the developers will suffer. This narrative of "self-sufficiency" has also proved to be costly many times in history. For example, Huawei was forced to develop its own operating system. Although it was ultimately successful, the process was extremely painful.

The article questioning the miracle of AI medicine is also worthy of attention. It directly points out the core issues of the "AI Everything" narrative. The story of using GPT-4 to make a communication App for an autistic son may seem warm, but it essentially simplifies a complex issue. Real medical innovation requires clinical trials, data verification, and regulatory approvals, not a weekend of hackathon projects. This reminds me of Silicon Valley's "bad blood" scam. Theranos also deceived countless people with the narrative of "one drop of blood tests everything". Nowadays, the AI field is full of such "miracles", but if you look closely, many of them are "information integration" rather than "scientific breakthroughs." The risk is that when the public's expectations for AI are excessively high, once the actual results are found to be far less than expected, the trust of the entire industry will collapse.

To sum up, today's signals point to a core judgment: the power pendulum of the AI industry is swinging from "innovation" to "control." Regulators began to act, capital became anxious, and giants began to shrink their defenses. For ordinary users and developers, the best strategy is not to blindly chase new products, but to ask themselves three questions: Does this AI product really solve my problem, or does it just make me feel that the problem has been solved? Who is making money behind the scenes and who is taking risks? If supervision is suddenly tightened, will my dependence turn into a liability? These questions are the real risk assessment.