Saturday, April 18, 2026

Today's pile of signals may seem messy, but when you carefully pull them off, you will find that there are actually two words: Agents are surging.

Take a look at Trending Code, where there is a water of agent frameworks, agent assistive tools, such as "growing agents","design systems for agents","plug-ins for Claude's memory", and the "agency skills framework". Looking at Product Hunt, Claude Opus 4.7 emphasizes "agency coding", OpenAI directly upgraded Codex to "running applications and automating tasks", and even a new product called The Factory directly branded "Agent-native software dev". These people really couldn't sit still, for fear that they would not catch up with this wave of agent craze if they were too late.

But the question arises. Agent sounds beautiful, but how can it be used in practice? Look at V2EX. There is a brother who bluntly said,"RAG is not satisfactory." RAG is one of the cornerstones of the agent. If even RAG cannot be fed enough, no matter how good the agent boasted, its execution ability would have to be questioned. There is also the "caveman" project, which allows Claude to speak like a primitive in order to save tokens. This is simply a major irony of contemporary AI. With the development of AI to this day, in order to save some money, developers still have to teach it to "cut the crap". Isn't this a fight between engineering "stingy" and ideal "intelligence"? This shows that the so-called "agents" are still dancing on the tightrope between cost and efficiency. These agents may seem cool, but how many of them are backed up by accumulating tokens and burning money, and how many of them truly have the ability to generalize problem solving? These are all real engineering challenges.

Another point worth noting is that the war of AI has spread to all corners. In China, NVIDIA stepped into quantum computing and directly sent the CEOs of quantum companies to the throne of billionaires. This shows that once a giant enters the market, its certification and driving force for an industry is devastating. Although quantum computing is still far from the general public, Nvidia's move is really aimed at the next decade. On the other hand, Meta spends billions of dollars on Broadcom every year to make AI chips. This once again proves that in the AI era, self-developing chips is the core competitiveness, and buying others 'products is ultimately controlled by others.

Another point is that the implementation of AI is undergoing a process of "miniaturization" and "scene-based". Although the marketing of "AI PCs" and "AI phones" may sound a bit gimmick,"end-side AI" is indeed a trend. When the cost of cloud AI is high and latency is obvious, delegating some capabilities to the device can not only protect privacy, but also improve the experience. Of course, the direction of end-side AI has not yet completely converged, but devices like mobile phones and PCs that everyone has are undoubtedly the best carriers. However, whether the so-called "AI PC" is a real AI or an OEM IQ tax requires a clear eye.

In general, the tide of AI is still surging forward, and agents are the hottest wave at the moment. But don't just look at it, the money burned behind it, the engineering challenges it faces, and whether it can really solve practical problems are what we who do technology should pay more attention to. As for those grand narratives, just listen. The key is whether your code can run, whether it runs fast, and whether it saves money.

Today's technology signals reveal an accelerating turning point in the industry: AI is evolving from a tool assistant to a digital organism with sustained memory and growth capabilities. The explosive growth of projects such as Hermes-agent and Claude-mem shows that developers are no longer satisfied with single-session AI interactions, but are pursuing digital partners who can accumulate experience over the long term and form a personalized knowledge base. This transformation is similar to the leap of mobile phones from feature machines to smart machines, indicating that AI applications will enter the era of "personal data assets." But this has also brought new issues of data sovereignty. For example, the debate on GitHub about "whether AI should have API access" reflects the industry's anxiety about AI's sense of boundaries.

The Chinese technology circle shows a clear trend of "AI hardware." The glorious shrimp farming book, the Krypton 8X 1400-horsepower electric car, and the Ean N60's lidar smart driving system are all turning AI from cloud algorithms into a touchable physical experience. The rapid iteration of this hardware carrier is creating "AI Moore's Law"-AI's physical penetration doubles every 6 months. But the risk is that the market could fall into a configuration arms race, like the quantum computing ETF demonstrated today, which surged 30% in a single day, reflecting that capital's speculative pursuit of hard technology could create new bubbles.

There is an interesting "atavistic phenomenon" in the developer ecosystem. The explosion of the Caveman project proved that when AI is too complex, humans instinctively seek minimalist interactions. This project, which reduces the use of tokens by 75% in primitive language style, is essentially a rebellion against the current over-engineering of AI. Similarly, the popularity of the Pyra tool chain reflects developer weariness with bloated technology stacks. This trend may spawn a new generation of "subtractive innovation"-just as the iPhone disrupted smartphones with steamed buns in 2007, the killer AI products of the future may be those that can help users reduce operating steps rather than add features. But the challenge is how to simplify interactions without sacrificing the depth of AI's capabilities.

In today's science and technology news, there are several signals that deserve special attention. They are about AI agent technology, quantum computing, and the development of new energy vehicles. The latest developments in these areas not only reveal the potential of technological progress, but also expose potential problems and risks.

AI proxy technologies such as NousResearch/hermes-agent and Julius Brussee/caveman, as well as Claude Opus 4.7 and The Factory Desktop App, are emphasizing improving development efficiency and user experience through AI proxies. These projects and products look very promising, but it is questionable whether they will actually increase productivity as significantly as advertised. Historically, many new technologies have been overstated in the early stages, such as early AI programming assistants and automation tools, but the actual results have often been far below expectations. In addition, the widespread use of these AI proxy technologies may bring new security and privacy issues. For example, automatically capturing and processing data during the development process may reveal sensitive information. In terms of cost, although AI agents can save some manual operation time, the cost of developing and maintaining these systems cannot be ignored. Whether the company is willing to pay high fees for this remains unknown.

The latest developments in the field of quantum computing, especially the open source quantum AI model Ising launched by Nvidia, have attracted strong attention from the market and industry. Although Nvidia's entry has indeed brought new vitality to the quantum computing industry, we need to remain calm. Quantum computing has made some breakthroughs in theoretical and laboratory environments, but it still faces many challenges in practical commercial applications. For example, the stability and error rate issues of qubits, as well as the special environment and hardware support required for quantum computing. The launch of the Ising model may speed up the processing of certain tasks, but it remains to be seen whether it can solve a wider range of practical problems. In addition, the high threshold for quantum computing means that only a few large companies and research institutions can truly benefit, while most small and medium-sized enterprises may have difficulty keeping up with this wave of technology. Market optimism about quantum computing could also spark a bubble, similar to previous artificial intelligence and blockchain craze.

In the field of new energy vehicles, the release of Krypton 8X and Aon N60 demonstrates the competitive landscape and technological progress in China's new energy vehicle market. These vehicles have not only reached new heights in performance, but have also made significant improvements in intelligence. However, the fierce competition and high research and development costs in the new energy vehicle market have put many companies under huge financial pressure. The price of Krypton 8X is as high as 356,800 yuan, which is not an easy price for ordinary consumers. Whether the market can absorb these high-end models and whether companies can continue to make profits under the dual pressure of technology and market are issues that need attention. In addition, infrastructure construction for new energy vehicles, such as charging stations and power exchange stations, remains a bottleneck. Although many companies are actively promoting the construction of these infrastructures, the speed and coverage of progress still need to be improved.

Overall, although today's technological signals are exciting, the bubble elements and potential risks behind them cannot be ignored. Breakthroughs in these fields of AI agent technology, quantum computing and new energy vehicles need to balance the relationship between technological innovation and practical applications to avoid disappointment and market bubbles caused by high expectations. At the same time, companies need to pay more attention to cost control and sustainable development to ensure that these technologies can truly benefit the majority of users, rather than become a luxury for a few.