Thursday, July 2, 2026

After reading a circle of science and technology news today, I smelled a strong smell of living a tight life. In the past, big factories engaged in AI, which was called a Token carnival. What about now? Companies such as Lindy and Uber directly replaced Claude with DeepSeek V4, claiming to save 90% of costs. This matter is too real. Capital does not care about feelings. The era of accounting has officially arrived. Lao Huang is boasting that Prompt is dead. Loop is in power. To put it bluntly, it is asking you to turn AI into a closed-loop system and stop manual training. Real small model economics is the job that engineers should do. You write CRUD with the most expensive model every day, and the month-end bill can scare the boss into a heart attack.

Looking at the underlying infrastructure and tool chain, there are also undercurrents. The Institute of Information and Technology has developed a benchmark for evaluating intelligent computing operation and maintenance agents, and also covers five domestic chips. This is paving the way for domestic computing power, indicating that the operation and maintenance of large-scale model training and reasoning has become a stuck engineering problem. In the open source circle, codebase-memory-mcp, an MCP tool that indexes the code base into a knowledge map and claims to save 99% of tokens, has directly hit the hot list. In order to prevent China users from distilling, Anthropic actually inserted a monitoring program into the Claude Code, preventing you like thieves. What does this mean? To implement an Agent, it is not enough to have a model alone. You must have memories, a tool chain, and you must be prevented from being paid for nothing.

So don't believe in fancy concepts. You see that Tsinghua University start-up received hundreds of millions of dollars in financing, but he refused to be labeled as a model of the world because the concept that it could not solve practical problems was nonsense. For ordinary developers, what we should get started now is definitely not how to write fancy prompt words, but how to build a hierarchical invocation Agent infrastructure and how to use MCP to feed local code bases and business data to AI. Whoever can make AI achieve high-end effects on low-end hardware will have the job in the next two years. Don't just watch and go back and optimize your API call bill.

You noticed a very interesting detail: While Sony announced that it would stop producing physical game discs in 2028, a paper on the first successful division of synthetic cells made headlines in HN, and Claude Sonnet 5, a model representing the highest intelligence level of AI, was also quietly released. Taken together, these three signals actually tell the same underlying story-the separation of the physical world and the digital world is accelerating, and the speed and depth of this separation far exceed most people's imagination.

Let's talk about Sony's decision first. This is not just the end of optical discs, but the end of the entire "media era." Over the past three decades, all digital content has had to be attached to a single physical carrier-optical discs, cassettes, hard drives. This bundled relationship gives us the illusion that owning a CD is equivalent to owning a game. But 2028 is a critical time point, marking the complete farewell to physical backup solutions for mainstream consumer-grade products. At the same time, a California laboratory announced that it had successfully built "synthetic cells that grew and divided from scratch"-the first time humans have created a living body according to the design blueprint. One is that the digital world is completely decoupled from the physical medium, and the other is that the physical world has begun to self-assemble according to the digital blueprint. This set of comparisons reflects a trend: in the next decade, we will experience both the "dematerialization" of digital content and the "programmability" of the physical world. You no longer need a CD to buy games on your mobile phone, just like in the future you can design a cell on a computer and print it out.

Another signal that is underestimated by most people lies in the contradiction between the release of Claude Sonnet 5 and the Anthropic blocking mechanism. The Qwythos-9B model on Hugging Face is based on a fine-tuned version of Claude, with 1148 visits-a large number in the model circle. Reddit broke the news that the Claude Code was embedded with a detection program for China IP. This explains two things: First, the most advanced open source models are actually "distilled" from closed-source models, which is the true ecology of the current AI industry chain; second, Anthropic chose to use hard blockages at the code level to stop this distillation, rather than following a legal or political line. This means that the country-specific barriers to AI models are changing from "policy initiatives" to "architectural decisions"-future models may have restrictions on use at the architectural level from the beginning. This is bad for China developers, but it is a catalyst for the entire industry: when open source distilled versions cannot be obtained from the outside, self-developed large models change from an "optional" to a "must-have." Meituan's open source LongCat-2.0 has trillions of parameters and a long context of 1M. In this context, it is not just a technical demonstration, but a strategic hedging.

Let's talk about the Hubei candidate who gave up Qingbei and chose Nanjing University Computer. 702 points, 142 in Chinese. This choice is actually an accurate response to the above two trends. When the industry of physical media is dying out and AI capabilities are cut off by geopolitics, there are only two real security assets: intellectual labor that cannot be eliminated physically, and original creativity that cannot be distilled. Computer science is stuck in between. But there is a deeper logic to this choice-it represents an accelerating trend of "professional supremacy." In the past ten years, the value of universities has been understood as "brand" and "connections," and it is an iron rule that top scores match top prestigious schools. But the examinee's logic completely betrayed it: he placed professional rankings above the school's brand. This is not an isolated case. When AI can replace a large number of general-purpose jobs, and when the recruitment logic of top companies and laboratories is increasingly focusing on high-level capabilities in specific fields, Qingbei's "comprehensive aura" is being replaced by the "professional moat" of Nanjing University Computer.

Finally, talk about the neglected signals. Israel's "Operation Blue and White" and Spain's record high death toll in June remind us that the strings of geography and climate are still tightening. Goldman Sachs estimates that the World Cup has created 40,000 new jobs in the United States-a large event can drive seasonal fluctuations in employment data, indicating that the economy itself is still fragile. The probability that the Fed will maintain interest rates is still above 70%, which means that capital costs will not fall in the short term. For the technology industry, this means that the financing window will not open within half a year, but Tsinghua chip companies, nuclear fusion companies and smart exoskeleton start-ups have received hundreds of millions of yuan in financing at this time-this is not a trend, but investors 'logic has changed: when short-term exit is hopeless, funds will flow more intensively into tracks with deep technical barriers and highly consistent with long-term trends.

One day in July 2026, countless technological signals converged to the same conclusion: the programmable era of the physical world has arrived, digital assets have been completely separated from the medium, national barriers in the AI field have moved from hidden to clear, and individuals and enterprises Security assets are shifting from "brand" to "professional depth." The transformation will not be completed in a year, but it will reshape the competitive landscape in almost every industry over the next eighteen months.

I looked through the signals you received today, and I feel that the technology community is collectively engaged in a large-scale performance art. Everyone is shouting "efficiency revolution" and "paradigm shift," but what I see is a group of people working hard to build walls-industrial walls, geographical walls, cognitive walls.

Let's talk about the core cost issue first. Meituan has opened up the trillion-parameter model LongCat-2.0, and the Institute of Information and Technology has developed a benchmark for intelligent computing operation and maintenance. These are good things. But pay attention to the rhythm: on one side, there is a practical article about "The end of the era of Tokens management", teaching companies to replace Claude with DeepSeek V4 to save 90% of costs; on the other side, Silicon Valley is madly pursuing "Loop replaces Prompt", and Wu Enda and Huang Renxun are shouting. Is this a contradiction? As soon as companies learned how to save Token money, experts said Prompt was dead and forced you to go into the autonomous circulation system. The cost is not yet understood, and you have to invest in new architecture-which is essentially a huge migration cost trap. This was the case in the late stages of the SaaS bubble in history. Manufacturers continued to create new concepts to ask customers to pay for it. In the end, customers found that the operation and maintenance debt caused by the new system far exceeded the savings. I slapped the table on the concept of "verifying debt". This is the real question: After AI is cycled independently, how much energy do you have to spend every day to verify that it does not go astray? No one accounts for this account.

Looking further down, Reddit broke that Claude blocked China user monitoring programs, Sony stopped producing physical disks by 2028, and Israel was preparing to fight against Iran-put these signals together, the trend is clear: technology products are becoming geopolitical sandbags. Sony's physical disk hacking is a digital transformation on the surface, but in fact it controls 100% distribution channels. What you want to play in the future, how you play, and whether you can modify it all depend on the platform side. Claude is even more naked. Monitoring is not for user experience, but for fear that you will use it to distil a model. An AI company has written China users into client code, which shows that the foundation of technical trust is rotten. The global technology ecosystem is being cut into islands by the triple walls of politics, business, and algorithm. The "zero API fee" open source tools that developers grab today may suddenly be cut off tomorrow due to sanctions.

Finally, I want to talk about a seemingly irrelevant signal: Hubei candidates with a score of 702 will give up Qingbei and choose Nanda computers. This is a smart choice, but note that he chose not "AI" or "Big Model", but "Computer". At a time when the global AI bubble is compounded by geographical blockade, selecting basic hardware and the soil at the bottom of the system is much more reliable than chasing hot spots. Today, stories of hundreds of millions of yuan in financing from the prestigious department of Tsinghua University are everywhere. Smart Core Light, Zhongjian Technology, and Hangmo Technology are all entrepreneurship in Qingbei. But the real barrier here is the one that makes interconnection between silicon-based photonic chips. Others that make myopia presbyopic smart glasses and exoskeletons are essentially betting that hardware manufacturers are willing to pay an extra premium for their purchases. When I looked at these financing news, I thought of those Tsinghua teams in the VR bubble in 2015 that wore headsets to raise funds-nine out of ten eventually became PPT warehouses.

Who makes money? It sells shovels (chips, benchmarks, Agent frameworks). Who is paying the bill? It is companies that are eager to use AI to reduce costs and consumers who pay for equipment. The ones who can survive this round are not the ones who shout the Loop slogan loudest, but the ones who can calculate clearly "who will continue its life in 20 years."