Tuesday, July 14, 2026

Regarding AI voice, GPT-5.6's Live voice mode does have something. The full-duplex architecture and task delegation finally give voice interaction a little "human", not the kind of stupid question-and-answer dialogue. But don't rush to climax, there are still many pitfalls in the landing project of this thing. Real-time voice puts much greater pressure on delay, packet loss, and context management than text. Moreover, the layered system of OpenAI sounds beautiful. In fact, if you think about it, a voice command needs to be broken into three pipelines: intention recognition, content generation, and Text To Speech. The synchronization must be ensured. This complexity is beyond the reach of an average team. Apple's SpeechAnalyzer API is quite pragmatic, directly benchmarking against Whisper, and may focus more on local end-side reasoning, which is in line with Apple's consistent privacy card. But to be honest, for most developers, using these APIs to make products now, don't expect to reproduce the "feeling of life" in the demo. They are all carefully trained cases. If you take it to do customer service, the user will turn over with a dialect and background noise.

Looking at the ecosystem of agents, interesting things have come out in the past few days. OfficeCLI, an open source project, allows AI agents to directly read and write Office files, does not rely on Office installation, and only binary files. This is true pragmatism. Giving AI the hand to operate documents is more practical than doing fancy multimodal things. There is also a live data market like AgentKey, which helps you feed real-time data to agents. The MCP monetization tool allows you to monetize an MCP server in 5 minutes with zero commission. This is all about building infrastructure for agents, which shows that the industry is beginning to seriously consider the commercialization of agents and is no longer just playing with concepts. However, although the "Self-Regulation Convention on the Protection of Personal Information for Agents" has been signed by 31 companies, to put it bluntly, it is just a compliance shell to prevent harsh supervision. The actual effect? Wait until something goes wrong.

The article on manufacturing AI is quite accurate. Don't brag about unmanned workshops all day long. First, solve the earthy scenarios of employee daily reports filling in indiscriminately and quality inspection data lagging. Using the lightest tools, such as OCR+ form recognition, you can save a lot of manpower by digitizing offline data. When AI is implemented, it is most afraid of starting a large model reconstruction process as soon as it comes forward. Bosses will be tricked into spending money, and in the end, they will not even tighten a screw. The really reliable way is to first find the most painful, dirtiest, and most repetitive link in the factory and use AI to turn it into automation. Even if you just write the Excel formula correctly, it is better than creating an "AI middle platform".

Finally, say something else. Today, the developer community is full of posts about losing money in buying a house and the pressure to raise a child, which is in sharp contrast to these technological orgies. The technology circle is always chasing new hot spots, but most people's lives are still daily necessities. Don't be carried away by the news about AI agents, voice models, and autonomous driving. Learn what you need to learn, but don't get all in illusions. Look at Byte, all of them are engaged in autonomous driving, but the route is the World Model of the Seed team. It is not to build a car directly, but it is still necessary to find a landing scene. No matter how powerful the technology is, in the end, we still have to solve real problems, such as asking the factory owner to send two fewer WeChat messages urging goods, or letting your child do his own homework (just kidding). Be pragmatic and don't wait for the bubble to burst before regretting it.

The most noteworthy hidden line today is that AI is completing its extension from the brain to the limbs. The feeling and explosion brought by the GPT-5.6 Live voice model and the breakthrough of Apple's SpeechAnalyzer API are not only an improvement in speech recognition, but a leap in the human-computer interaction paradigm from request response to full-duplex real-time coexistence. When ByteDance prepared for autonomous driving through the Seed World Model team, and even the US military used unmanned boats at sea to attack Iranian facilities in actual combat, we saw the same underlying logic: the big model is no longer a dialogue box on a web page, it is becoming a real-time operating system for the physical world. Real-time processing of voice and vision allows AI to directly take over vehicles and weapons. Half a year later, this full-duplex interaction will completely change the form of intelligent hardware. Applications that only make text API casings will be eliminated, and model manufacturers and hardware integrators who truly master end-to-end real-time control capabilities will take away the biggest cake.

At the same time, the rapid digestion of hardware is reshaping the logic of manufacturing. The average age of new energy vehicles is only 1.8 years, and the replacement cycle is on par with that of mobile phones. This is not only a change in consumption habits, but also a dimension reduction blow to asset-heavy industries by the speed of technological iteration. But this also means huge waste of resources and supply chain pressure. SK Hynix CEO warned that 2027 will face the largest storage shortage in history, and even consider a memory-as-a-service rental model. This is by no means groundless. When new energy vehicles, AI data centers and end-side AI PCs are all madly competing for computing power and storage, the shortening of hardware life cycles and the rigid limitations of underlying production capacity will inevitably collide violently. A year later, we may see the hardware iteration speed forced to slow down due to the peak of physical production capacity. Companies with upstream resource allocation capabilities and recycling technology will benefit, while new car-building forces that blindly pursue high-frequency replacement will face the risk of capital chain rupture.

Behind these grand narratives, AI's commercial closed-loop is being implemented in an extremely rustic way. Whether hardware factories use AI to process chaotic forms and quality inspection orders, or Douyin e-commerce incorporates bean buns into the shop settlement sequence to achieve AI diversion attribution, it all shows that AI has passed the conceptual stage. Those companies that are no longer obsessed with the fantasy of unmanned workshops, but use lightweight tools to solve the problems closest to money are the truly pragmatic ones. When 31 companies signed the Convention on the Protection of Personal Information for Agents, it also marked the transition from barbaric growth to rule establishment. The winners of the future are no longer the companies with the coolest demos, but the ones who can seamlessly weave AI into the most boring business processes and achieve profitability within physical boundaries where computing power is scarce.

Among today's technological signals, the two most worthy of digging into are the "fast digestion" trend of new energy vehicles and the real dilemma of AI implementation in the manufacturing industry. The two may seem unrelated, but in fact they both point to the same question: Who is really benefiting from technological progress and who is paying for the bubble?

Let's first say that the average age of new energy vehicles is only 1.8 years. The logical chain behind this data is very clear: subsidy policies drive sales surge, capital chases the wind to create overcapacity, and then car replacement behavior is packaged as a "youth consumption upgrade" through marketing rhetoric. But think about it carefully, what does 1.8 years old mean? This means that a large number of vehicles are eliminated at the most severe stage of battery decay, and the cost of recycling and processing these batteries is currently not included in the industrial chain. What's even more ironic is that when we replace electric vehicles like mobile phones, we are actually exchanging higher resource consumption for the so-called "sense of technology." Historically, this "pseudo-upgrade" phenomenon has occurred in any industry during a period of rapid expansion. For example, the wave of functional machine replacements in the 2000s left only e-waste and consumer debt. The difference between new energy vehicles is that they carry too many expectations for environmental protection and energy revolution, and this expectation is being used by capital and turned into a new consumerism trap.

Then there is the implementation of AI in manufacturing. The hardware factory owner mentioned in the article may know the truth better than any analyst: AI's gimmicks are far greater than their actual value. In reality, those "unmanned workshops" and "intelligent manufacturing" that are frequently promoted often become simple digitization of basic data such as daily reports and quality inspection orders. This is not a technical issue, but a business model issue. AI companies need to sell solutions, and manufacturing bosses need political achievements and subsidies. The two sides hit it off on the concept of "high-quality", but the real improvement in production efficiency has been ignored. What is even more dangerous is that this superficial digitalization may conceal the real problems of the manufacturing industry: the industrial chain relies too much on low-cost labor and lacks core technology accumulation. Historically, Japan and Germany's manufacturing industries relied on lean management and craftsmanship spirit rather than "black technology." The implementation of AI we are seeing now is more like a new kind of "digital packaging" used to cover up long-standing structural problems in the manufacturing industry.

What these two phenomena have in common is that technological progress is reduced to a consumer behavior or a capital game. The rapid iteration of new energy vehicles and the apparent prosperity of AI manufacturing are actually consuming social resources, and the real beneficiaries are capital and policy makers. For ordinary consumers and manufacturing workers, they may just be paying for someone else's story. What is even more alarming is that this frothy technological advancement will obscure the real question: Where is the battery recycling system for new energy vehicles? How to achieve the accumulation of core technologies in the manufacturing industry? When we focus all our attention on the concepts of "car replacement" and "AI empowerment," these truly important issues are ignored.

Technology is never neutral, it always serves specific interest groups. When we see an industry or technology overhyped, ask: Who is driving the story? Who is paying for this? History tells us that any technology bubble that is divorced from actual needs will eventually burst, leaving only waste of resources and social mistrust.