Monday, July 6, 2026

The signals worth talking about today actually point to one thing: the AI industry is moving from an arms race of "whose model is smarter" to a pragmatic stage of "whether this thing can be implemented at all?" You see, it's very interesting to string together several signals. Ali has completely disabled Claude Code internally because of the discovery of backdoor risks. This is not a minor security vulnerability, but a direct hit on the dead hole of enterprise-level deployment. GPT-5.6 has been revealed to be released next week. It has three major sub-models with external acceleration dial functions, and OpenAI is still struggling to show its control. On the other hand, the MiniMax M3 's narrative has shifted from "our models are smarter" to "we redefine our values." Taken together, these three signals reflect that the AI industry is experiencing a critical turning point: the marginal benefits of model capabilities are diminishing, and real competition is beginning to shift to reliability, controllability and practical implementation capabilities. Ali's ban on Claude Code is a wake-up call for ordinary developers. The Codex API is really cheap, and promotional posts are posted on V2EX. However, if the security review is skipped, what awaits you may be the compliance red line of the entire company. Relying on third-party closed-source AI tools for core development processes is much more risky than you think. What is more noteworthy is that Ali recommends an alternative solution Qiaoder, which means that domestic manufacturers are accelerating the construction of their own AI development infrastructure, and the signs of ecological fragmentation are already obvious. In the future, people working on code may have to prepare two sets of tool chains, one for domestic use and the other for foreign use. This is the reality. Let's look at the hardware direction. The OASIS voice input ring has become popular, but the logic is correct-smart rings have always lacked a killer scene, and the concept of voice input plus Vibe Coding is very clever. But you have to think about it, can the domestic team making similar products handle the latency and power consumption of speech recognition? Once this category explodes, Apple and Huawei will definitely follow suit, leaving small teams with no more than a year. As for the financing news of exoskeletons and embodied intelligence, Osha, Light Elephant Technology, and Memory Technology are all taking money, but the maturity of these tracks varies greatly. Exoskeletons can already organize outdoor hiking activities, and the commercial path is relatively clear, but the delivery of embodied intelligent robots in industrial scenarios, to be honest, is still a way to go before truly making money stably. The story of brothers starting a business at Tsinghua Vehicle College is very good, but being able to get angel wheels does not mean that large-scale delivery can be achieved. The hottest Agent safety barrier design in the technology circle is actually the most practical value to developers. The security of Agent products cannot be solved by simply adding an audit layer. Dynamic routing and risk stratification are needed from the product design level. This is the hard nut in engineering, and it is much more meaningful than staring at the model parameters. There is another incident that is not related to technology but worth mentioning today, that is, the girl who scored 699 in the college entrance examination was sexually assaulted online. Discussions in the technology circle often focus on efficiency and advancement, but the lack of underlying logic is the real risk. By choosing to study medicine, she can solve more practical problems. If technical practitioners only focus on code and not society, sooner or later they will step into the trap. Overall, today's signal tells you that model capabilities are no longer a moat, and project implementation, safety compliance, and scenario definition are the keys to the next stage. Don't chase after the parameters of GPT-5.6, first see if the Agent you have at hand can avoid accidents.

In today's signal, there is a hidden chill spreading. Ali has completely banned Claude Code internally, citing the "risk of embedding backdoor"-this is not an isolated incident. At the same time, OpenAI quietly released codex-plugin-cc, allowing Claude Code to call Codex for code review or task delegation. In the open source community, the usestrix/stripe project tops the trend list. It uses AI hackers to proactively discover and fix application vulnerabilities. The three lines point to the same judgment: the safety barrier of AI agents is no longer a chapter in the product manager's document, but is becoming an arms race. When companies begin to ban employees from using a certain AI tool, it is not because the function is poor, but because the trust system has not been established, which means that the delivery model of the entire AI tool chain must be restructured. In the next six months, we will see a large number of enterprises building or locking in specific AI security middle layers. Similar to SSL Certificates for the Internet, AI trust layers will become infrastructure-level businesses. Scenarios such as code review will quickly switch from "manual review" to "AI-to-AI attack and defense" mode.

Another noteworthy hardware trend is hidden in rings and mechs. The popularity of the OASIS voice input ring has finally broken the narrative of the smart ring's "reduced version of the bracelet"-it does not perform health monitoring, but becomes a direct AI interactive portal. This is in line with my previous observation that "AI hardware is de-screening." When Vibe Coding brings fire to voice input, and when Osha's exoskeleton turns outdoor hiking into a human-computer collaborative experience, we see two different scales of the same thing: human-computer interaction is returning from finger touch to more natural voice, proprioception, and gesture commands. What the founder of Osha said about "redefining the human-machine relationship" is not rhetoric, but hard logic-when AI models can understand human movement intentions in real time, the exoskeleton is no longer just a rehabilitation device, but the next generation of personal enhancement equipment. The explosion point of such products depends on two factors: battery energy density and lightweight materials. The continuous rise in the price of electronic cloth and the monopoly of high-end looms by Japan until 2029 indicate that the bottleneck in the supply chain is spreading from semiconductors to new materials, which will slow down the pace of hardware innovation, but it will also become a definite window of opportunity for domestic substitution.

There is also a set of signals that can easily be treated as isolated commercial disputes, but together they reveal the fragmentation of technological globalization. Wahaha was detained by the FDA for sodium cyclamate. It clearly met China standards but was defined as a "violation." This is not a simple trade friction, but a decoupling of the standard system. Infineon and Innoseco had wins and losses on gallium nitride patents. China courts banned Infineon, and German courts ruled in favor of Infineon. The legitimacy and commercial destiny of the same product in the two judicial systems are completely different. Behind the continuous rise in electronic fabric prices is the surge in demand for AI computing power, which has driven high-end substrate materials, and key equipment completely relies on Japan. The narrative behind these signals is that the technology supply chain spawned by AI and new energy is shifting from global division of labor to regional autonomy, and each country is accelerating the establishment of its own standards, patent barriers and production capabilities. Risks at the enterprise level have changed from "technological competition" to "compliance arbitrage", and multinational companies that can control multiple sets of rules at the same time have gained asymmetric advantages. Half a year later, we will see more "geo-designed AI products." We will not only customize language and interface for the China market, but also security architecture, data sovereignty and authentication paths.

The most noteworthy thing in today's signal is not which model swipes the list, nor is it how much money which company has raised, but two things: First, the AI industry is undergoing a round of collective skin change of "value narrative", and second,"ownership" and "security" issues from games to Agent products are quietly deteriorating.

Let's look at AI first. The MiniMax M3 report that "it's no longer just about how smart the model is, but about redefining the company's value" sounds very advanced. But to put it bluntly, this is the last round of arms race that competed for parameters and benchmarks that could not continue. Everyone found that no matter how big the model is, no money can be made. The gross profit margin of the business model relying on API calls is bleak, and ToB's customization projects are heavy and slow. So now everyone is starting to talk about "redefining value", which is essentially looking for the next starting point that can tell stories, raise funds, and support valuations. There are two risks involved: First, the so-called new value is often about putting new wine in old bottles, packaging models into platforms and products into ecosystems, but the actual realization path is still blurred. Second, new concepts such as embodied intelligence, Agent, and AI hardware are rapidly being foamed-look at today, several embodied intelligence companies with Tsinghua background have integrated hundreds of millions of angel wheels. The model is surprisingly consistent, but the landing scene is still stuck on the concepts of "automobile industry" and "rehabilitation exoskeleton". History has repeatedly proved that when a large number of homogeneous startups and cross-definition financing are poured into a track at the same time, the bubble is not far from bursting.

Another warning signal focuses on the "safety and ownership" line. Ali has completely banned Claude Code internally on the grounds that there is a risk of embedding backdoors. This actually reveals a terrible reality after AI Agents enter the production process on a large scale: When your development work begins to rely on AI-generated code and AI agents to perform operations, the security of the system will be completely entrusted to a black box that you cannot completely review. Disable Claude Code today, and it may be any company's turn tomorrow. This is not an isolated incident, but a microcosm of the entire industry's serious lag behind functional expansion in the construction of safety barriers. At the same time, the article on digital game ownership on Hacker News and the report on price increases in the game industry can be seen in conjunction-players are losing actual control of "purchased content", while companies are using "AI empowerment" and "subscription system" to accelerate this deprivation. Users are increasingly not users, but tenants of products.

As for social hotspots such as "poor girls being cyber-abused", it seems to be an emotional event, but its deep signal is that the equal information brought by technology has not made society more rational, but has instead made moral trials and group violence more easily amplified. Tsinghua University's choice of "Zhuoyi Class" was questioned. Behind it was distrust of the flow path of classes, which was further solidified under the recommendation of the algorithm. Technology companies talk about "social responsibility" every day, but few people are willing to face up to the cognitive divisions their platforms are creating.

Judging in one word: There are not many really useful things in today's technological signals, but those narratives that are packaged as "positive"-whether it is the reshaping of the value of AI companies, the angel wheel with embodied intelligence, or the so-called new hardware popularity-it is worth asking you more: Who is working hard to promote, who is paying the bill, and who will cover the final bubble?