Saturday, April 25, 2026

The most eye-catching things in today's pile of signals are two major contradictions: on the one hand, domestic models are charging crazily, and on the other hand, AI infrastructure is crumbling. As soon as the DeepSeek V4 dual model was released, someone in the Chinese circle complained that the weather card collapsed-the technical report was blown loudly, and even basic functions were overturned when landing. What's even more ironic is that the GPU shortage has made headlines, and it has become the norm for startups to fail to grab cards. Lao Huang's smiling face and the tears of a startup company are two sides of the same coin.

The massive defection of Claude users is another hot spot. Hacker News's denunciation with blood and tears accused three crimes: token cheapness, quality decline, and customer service pretended to be dead. The most magical thing is the community's reaction: on the one hand, someone came up with "primitive human code" to slash 65% of the token consumption (Caveman Project), and on the other hand, it stripped all the warehouse inflation stars of various AI system prompts-users were forced to come up with two survival strategies, either self-castrate themselves to adapt to the model, or lift the table for transparency. This shows that the current commercialization of closed-source LLM is an endless cycle: if you want profits, you compress resources, and once you compress it, it will hurt the experience, and users will run away when the experience collapses.

The semiconductor roller coaster market is the third dark lightning. Intel's OEM business surged 23% in a single day, but Michael Burry backhanded short the SOXX ETF. What is even more interesting is that the expansion of Bubble Mart Park and the crazy price cuts of car companies (Deep Blue L06 is discounted to 120,000) occurred at the same time-the recovery of consumer electronics was transmitted to the manufacturing industry, but the Bubble Mart Hotel and 200,000 "national magic cars" first spawned? This economic signal is more difficult to understand than AI-generated weather.

Three conclusions in the short term: 1. domestic LLM Don't rush to blow technical indicators, first solve the basic experience of "weather card" level. It is a good move for DeepSeek to open the API, but getting through with Weixin Pay is the line of life and death. 2. closed-source AI company is following Netflix's mistake: price increases to drive customers → user piracy → stricter countermeasures → vicious cycle. Caveman's cult optimization is popular, and the essence is that users vote with their feet. The shortage of 3. GPU is more dangerous than imagined. Yesterday, we could laugh at players who couldn't buy a 4090, but today, startups lined up to wait for the news of H100 to reach the top-when the progress of the venture depends on Huang Renxun's shipping list, any AGI roadmap is nonsense.

Developers are recommended to do two things this week: go to DeepSeek's official website to collect free APIs to test the weather interface, and write a downgrade plan for local small models (such as Qwen3.6). Don't wait for a giant to give out money, your users will turn into cavemen using primitive syntax at any time.

Today's technology signals reveal a major turning point that is taking place: AI applications are evolving from "doable" to "easy-to-use", while the entire technology stack is undergoing a profound decoupling and reconstruction. This transformation is not only reflected at the product level, but also reshaping the entire development ecosystem.

The most eye-catching thing is the simultaneous explosion of the DeepSeek V4 series and the Qwen 3.6 series. These two China manufacturers unanimously adopted a dual-model strategy, using a combination of Pro/Flash and large/small parameter versions to meet the needs of different scenarios. This marks that large-scale model competition has entered the stage of refined operation, and the era of simply competing for parameter scale has ended. Particularly noteworthy is the multimodal capabilities of the Qwen3.6- 35B-A3B display, which reduces the latency of text-to-image generation by 40%, which indicates that we are entering a new era of "real-time generation" in the next six months. AI-generated content will shift from static display to dynamic interaction.

The open source community is accelerating the democratization of AI technology. From projects like andrej-karpathy-skills to caveman, we see developers building a new "reminder engineering knowledge body." These projects are essentially creating "design patterns" for the AI era, and they allow ordinary developers to gain access to the tuning skills of top researchers. But projects such as CL4R1T4S that leak system tips have also brought new controversy. They may force vendors to adopt more closed defense strategies, which creates a subtle tension with the spirit of open source.

Changes in the hardware ecosystem cannot be ignored either. The GPU shortage faced by AI startups is an interesting contrast to Michael Burry's shorting of semiconductor ETFs. This reflects that the market's excessive optimism about computing power investment may be forming a bubble. Zhongke Tianta's laser communication technology provides another solution to reduce reliance on local computing power by optimizing data transmission. This "software-defined hardware" approach may reshape the direction of infrastructure investment in the next three years.

These changes together depict a trend: the AI industry is moving from a period of technological breakthroughs to a period of engineering optimization. The winners in the coming year will not be companies with the most advanced algorithms, but teams that can seamlessly integrate AI into existing workflows. The risk is that this rapid commercialization may sacrifice the long-term evolution of technology, just as the prematurely solidified application form in the mobile Internet era limits subsequent innovation. What we need to be vigilant now is not to let immediate commercial value kill off the broader possibilities of AI.

Today's signals once again clearly outline the current technology field, especially the fanaticism and hidden contradictions in artificial intelligence. On the one hand, we see that large model technology is still iterating at an unprecedented speed. New models such as DeepSeek V4 and Qwen 3.6 are emerging one after another. Meituan has also announced that its trillion-level parameter large model has been trained by domestic computing power clusters, which undoubtedly demonstrates the momentum of technological progress. Various AI-based applications have also sprung up, from AI Internet celebrities, AI curatorial radio stations, to AI development and management tools, as if AI is omnipotent. However, behind these glories, we cannot ignore those costs and risks that have been deliberately downplayed or have not yet surfaced.

The biggest contradiction is that the scissors gap between the popularization of large-scale model technology and the actual application cost is rapidly expanding. On the one hand, we see the open source community working hard to optimize the use of large models through various techniques, such as improving Claude's code behavior by drawing on Andre Karpathy's observations, or significantly reducing token consumption through "raw" expressions like the "Caveman" project. What is reflected behind this? It is a careful calculation of high API call costs and computing resources by users. On the other hand, AI startups are generally facing the problem of soaring NVIDIA GPU prices and months of queuing, which directly kills the possibility of many budding innovations and raises the threshold for entry for AI services. Who is paying for this computing competition? Ultimately, it is end users and small entrepreneurs who lack capital support. When the API documentation of DeepSeek V4 was hotly debated, some people in the Chinese developer community bluntly said,"DeepSeek V4 Pro, keep working hard, the weather card effect is average." This gap between expectations and reality, as well as doubts about the actual effect of the model, just shows that there is still a long way to go before the technology is truly mature and universally beneficial.

What is even more worrying is that the quality of large models and support problems are becoming a signal that cannot be ignored. An article on Hacker News titled "I Cancel Claude: Token Issues, Declining Quality, and Poor Support" has attracted great attention, and this is by no means an isolated case. It reveals that stability, reliability and user experience can be sacrificed during rapid model iteration. When users discover that the tools they once relied on are beginning to "fail" or their performance is no longer consistent with the hype, market trust is damaged. History tells us that any technology, no matter how eye-catching it is initially, may eventually face the fate of being abandoned if it fails to continue to provide stable and reliable value and effectively solve user feedback problems. This phenomenon of "leaking system hints", whether for transparency or other purposes, points to the deep challenges in controllability and stability of current large models.

In addition, the signals in the semiconductor field are also worth pondering. The Nasdaq Index and the S & P 500 Index hit new highs, and Intel surged 23%, seemingly improving. However, well-known investor Michael Burry bought put options on semiconductor ETFs at this time, bluntly stating that the current rise is mostly caused by technical factors, and suggested that investors holding long positions in semiconductor stocks consider selling. This is a typical signal that "everyone is drunk and I am alone". It reminds us that the current market optimism may have a bubble element, especially under the infinite expectation of AI's demand for computing power, and semiconductors are its core support. Has its valuation been overdrawn? Historically, technology bubbles often peak when markets are in a frenzy, and subsequent pullbacks are always cruel. We need to be vigilant about whether in the wave of AI, we are repeating the mistakes of the past and confusing short-term hype with long-term value. Who ships at these high positions? Who is taking over at the high position? These issues deserve us to remain sober.