Sunday, May 3, 2026

The most interesting thing to talk about today is the accelerated integration of AI proxy tools and LLM coding. Both Warp Terminal and Hermes-agent projects are trying to deeply integrate AI agents into the development process, which is much more radical than simply Copilot auto-completion. But seriously, how practical is this "AI-first" development environment? I have tried Warp. Its conversational interaction looks cool, but it is not as efficient as actually writing complex business logic as using a traditional IDE. Today's AI agents are more suitable for handling standardized tasks and show their timidity when encountering scenarios that require in-depth systematic thinking.

Another obvious trend is the blurring of the boundaries between physical AI and digital AI. Looking at reports from Zhuoyu Technology in the Chinese area and Silicon Valley's embodied intelligence companies, everyone is suddenly mentioning the "mobile base model." Isn't this just putting a brain on a robot? But the problem is that the gap from demo to real scenes is much wider than imagined. Yushu Technology's shipments look beautiful, but if you actually look at user feedback, you will know that the performance of these robots in an unstructured environment is still disastrous. What is the difference between the current goal of "US$14 billion in revenue by 2036" and the big pie drawn by blockchain companies back then?

What worries me the most is that developers 'dependence on AI tools has reached the point of madness. Some people on V2EX also asked how to use Claude and Codex in China. Please, why are you still playing AI programming if you can't even handle a serious development environment? Look at those skills projects. They use Karpathy's training guide as a bible, but they can't even write the basic algorithms well. Nowadays, AI assistance is like equipping primary school students with a supercomputer. The computing power has been improved, but the mathematical thinking is still rotten. This kind of training method that puts the cart before the horse will sooner or later cost the entire developer community.

Today's technological signal data reveals a profound transformation being experienced in the field of artificial intelligence and software development. From open source skill catalogs to new AI models, to emerging products and business dynamics, these signals point to a trend: the practicality and integration of artificial intelligence is accelerating, and software development is gradually becoming an auxiliary tool for artificial intelligence.

First, the rise of open source skill catalogs and AI models shows that artificial intelligence is becoming more accessible and customizable. Projects like mattpocock/skills provide a catalog of personal skills that can be accessed directly from the user's Claude catalog, which not only simplifies skills management, but also promotes the automation of personal knowledge management. At the same time, the release of AI models such as Mistral-Medium-3.5- 128B and Wan2.2 14B demonstrates the huge potential of artificial intelligence in fields such as video generation and text generation. The emergence of these models will greatly promote the application of artificial intelligence in various industries.

Second, emerging products and business dynamics show that artificial intelligence is being widely used in various fields. The launch of products such as Cloud Computer by Manus and Scholé marks the application of artificial intelligence to improve work efficiency and personalized learning. The joint watch of Glory and Zhang Xue locomotive shows the trend of integration of artificial intelligence and consumer electronics, indicating that smart hardware will become a standard feature in daily life in the future.

However, this trend also brings a series of problems. First of all, with the popularization of artificial intelligence, data security and privacy protection have become an urgent issue to be solved. Secondly, the rapid development of artificial intelligence may lead to changes in the talent structure, and practitioners who lack relevant skills may face the risk of unemployment. Finally, the application of artificial intelligence in various fields may also exacerbate social inequality, because technology dividends may not be evenly distributed.

Overall, the practicality and integration of artificial intelligence are driving changes in the field of software development. This trend will bring huge opportunities to all walks of life, and is also accompanied by many challenges. For practitioners, maintaining sensitivity to new technologies and the ability to continue learning will be crucial. For the entire society, how to balance technological progress with ethics and ensure that the development of artificial intelligence can benefit mankind will be a common issue that needs to be faced in the future.

The most prominent thing about today's data signals is the proliferation of the concepts of "Agent" and "Copilot", as well as their gradually emerging limitations and costs in the real world. On the surface, AI seems to be omnipotent, from code generation, video creation to personal health management, and even the entire business model, it is penetrated by these "agents" or "co-pilots". However, when we peel off these shiny propaganda coats, some deeper problems emerge.

First of all, regarding the emergence of "Agenda Development" and various "Claude skills", this undoubtedly reflects the industry's strong expectation for AI to automatically generate and optimize code. We have seen many projects dedicated to improving LLM code behavior, even with concepts like "ML Intern". This is reminiscent of early projects that tried to fully automate software development, most of which ultimately failed. Historical experience tells us that software development is far from simple code accumulation. Requirements understanding, architecture design, complex debugging, and team collaboration are all aspects that AI is currently unable to undertake independently or even effectively assist. When the AI model automatically inserts submission information such as "Co-Authorized-by Copilot" into VS Code without authorization, this is not only a technical flaw, but also a violation of developer dominance. It also hints at the ambiguity and potential governance issues of AI on the collaboration boundary. Do we really need a "co-pilot" to quietly tamper with our submission records behind the scenes? Behind the convenience brought by this "intelligence" is there a hidden dilution of control over the development process and long-term maintenance and traceability challenges?

Secondly, the comprehensive penetration of AI is especially reflected in various new products and models, from image-generated videos to local AI retouching, to Microsoft's "Copilot Health". The true value and commercial sustainability of this trend of "everything can AI" are debatable. At the product level, many AI features are more like icing on the cake rather than solving core pain points. The deeper problem is that the large-scale construction of data centers driving these AI applications is bringing unprecedented energy needs. Berkshire Hathaway's Greg Abel mentioned that data centers will drive significant growth in the utilities sector, which is not empty talk. When data center electricity consumption already accounts for 8% of the peak power grid capacity in some regions and is expected to grow by 50% or more over the next five years, we must examine the costs of this growth model. The demand for AI computing power is growing exponentially, while power supply and infrastructure construction are linear or even limited. This contradiction between "unlimited demand" and "limited resources" will eventually turn into high operating costs and may bring environmental pressure. Who will ultimately pay for these costs? Is it consumers, businesses, or the future energy structure?

Finally, although the competition between embodied intelligence and physical AI is fierce, its claimed "large-scale implementation" and "aggressive revenue goals" still need time to verify. The bubble and technical bottlenecks of the early robot industry are vividly visible. In the pursuit of "low-cost, high-performance" expansion, we should be vigilant whether this kind of radicalization will sacrifice product reliability, safety and long-term economic benefits for market share. When a technology field quickly shifts from "laboratory breakthroughs" to "commercial mania", it is often the moment when risks begin to accumulate.