Wednesday, July 1, 2026

Companies have finally begun to settle accounts, and the era of sufficient Tokens management has come to an end. In the past, everyone used to make big models as lavish as spending public funds. Now they find that they can't bear the bills, so they began to use small models like DeepSeek V4 to replace Claude to do odd jobs. Layered calling is the right solution to engineering. Lao Huang recently talked about Loop Engineering, saying that Prompt is dead and that it wants to build an autonomous circulation system. It sounds mysterious, but it is actually to let AI run closed-loop automation. But there's a big hole behind this. Hacker News just revealed that Claude Code used steganography in requests. You think you are using AI to improve efficiency, but in fact, AI vendors are quietly tracking the flow of your code. When this black box risk is implemented at the enterprise level, you must guard against it and don't send out core business logic naked.

On the tool side, Cursor has been launched on iOS, and a Vibe Coding tool like Bilt.me that directly converts Figma into a mobile App has also emerged. Product managers can directly produce prototypes, and developers can change code anytime and anywhere. This blurring of boundaries is practical for improving efficiency. But don't just bother, Microsoft is still announcing a new round of layoffs. Big manufacturers are spending money on AI while cutting traditional positions. The signal is obvious. South Korea has even regarded doubling DRAM production capacity in five years as a national-level super project, and the hardware base is still expanding rapidly, indicating that this wave of AI infrastructure is far from reaching its peak.

Look at the Hubei candidate with a score of 702. He left Qingbei instead of going to Nanjing University to study computer science. This shows that this industry is still attracting the best minds, and the competition will only become more fierce in the future. Everyone on Hacker News praises Qwen 3.6 27B as a local development dessert, which actually points out the way for ordinary developers to survive. The future is where experts use AI to arm themselves to the teeth. If you don't learn to deploy locally and carefully plan hierarchical scheduling models, you may not even be able to drink soup in an increasingly sophisticated industry.

A core signal today is that the AI industry is undergoing a hidden shift from "ability competition" to "cost control". China Information Technology Institute released the intelligent operation and maintenance agent evaluation benchmark, which is ostensibly a set of technical standards, but actually means that the large model has entered the deep water area of "operation efficiency" competition. At the same time, the practice of hierarchical invocation of enterprise AI has been widely discussed-replacing Claude with DeepSeek V4 to save 90% of costs. This is no longer an attempt by a few companies, but a common phenomenon. Both Huang Renxun and Wu Enda are advocating that "Prompt is dead, the Loop Revolution", which is essentially saying: AI no longer requires humans to handwritten prompt words, but should become an automated closed-loop system. These three signals, put together, point to the same conclusion: the "gold rush" in the AI industry is ending, and the next stage is the "gold refining" stage. Whoever can produce more value in a unit of computing power will survive.

Another signal worthy of attention is the global game in the semiconductor field. South Korea has announced the "three major super projects" to double DRAM production capacity within five years and advance the wafer factory construction cycle by 12 years. At the same time, Samsung, SK Hynix, and Micron were sued in the United States for memory price manipulation. Apple and the European Union held "constructive talks" on compliance issues with Siri AI. Microsoft plans to lay off thousands of people. These events may seem independent, but they are actually connected together into a clear narrative: AI's demand for computing power is reshaping the entire hardware supply chain, but the supply chain itself is experiencing power restructuring and compliance pressure. South Korea is betting on an explosion in DRAM demand over the next five years, while the United States uses litigation and regulation to check and balance foreign giants. The reason why Apple's Siri AI hit a wall in the EU is essentially because it wants to run AI on the local end-side, which touches the EU's digital sovereignty and privacy bottom line. Microsoft's layoffs are a structural adjustment made amid the contradiction of huge investment in AI but undiminished cost pressure.

There is also an easily ignored but profound signal: Hubei 702-point candidates gave up Qingbei and chose Nanpu Computer, and the consecutive defeats of Germany and Japan in the World Cup. The former reflects young people's highly rational evaluation of professional tracks-they no longer believe in the aura of prestigious schools, but value actual technological accumulation and industrial connections. The latter reflects that in traditional competitive sports such as football, the forces of systematization, tactical discipline and professionalism are crushing the "talented" style of play. Japanese fans are ridiculed for cleaning up garbage for "performing civilization every four years." Behind this controversy is actually a cognitive separation between "performance behavior" and "systematic behavior." All these signals remind us that in today's world, focusing on system efficiency, cost control and long-term accumulation is more important than relying on single outbreaks and path dependence. In the next six months, we will see more companies shift from "stack models" to "stack operations" and from "matching parameters" to "matching costs". At the individual level, the differentiation between choosing a track and continuing accumulation will become more and more obvious.

The AI industry has finally been awakened by the real bill. When companies begin to replace expensive big models with hierarchical calls and small models, and when Microsoft increases its investment in AI while laying off thousands of people to control costs, this shows that the bubble of storytelling based on brainless computing power over the past two years is bursting. After the technology craze period ended, everyone finally realized that the commercialization of AI was far more expensive than the promotion. In this carnival, the ones who really make money are still chip giants and cloud manufacturers selling shovels, while the ones who pay the bill are small and medium-sized enterprises that are dragged down by high interface bills and try to replace manpower with AI but find that the hidden costs are higher.

After the dividends of the underlying model peaked, the circle began to skillfully coining new words again. From claiming that the prompt word project is dead to pursuing autonomous circulation systems, to packaging code generation with various new concepts, they are essentially covering up the current sense of powerlessness of AI in the implementation of complex projects. Allowing AI to realize round-the-clock automated closed-loop sounds sexy, but ultimately it will be up to human engineers to take care of the verification debt accumulated by machine illusions. Looking back at the technology bubble in history, every time it reaches a difficult stage to implement, there will be this kind of architectural hype that attempts to cover up the underlying flaws with more advanced abstract concepts.

In sharp contrast to the careful calculation of the application side is the blind rush on the infrastructure side. South Korea is expanding its semiconductor production capacity 12 years ahead of schedule for the artificial intelligence revolution, and the country is also intensively introducing smart computing evaluation benchmarks. However, at a time when the application layer is still worried about the cost of computing power and relies on layoffs to subsidize technology investment, this advanced construction of heavy assets poses a great risk of overcapacity. When enterprise-level real demand cannot support the huge consumption of computing power, those wafer factories and smart computing centers that blindly expand in order to catch up with the wind are likely to become the next wave of heavy asset burdens.