Friday, May 8, 2026

There are two things worth talking about most today: the scuffle at the AI application level and the arms race at the bottom of the computing power. Let's start with the first one. Now the entire AI application ecosystem has fallen into a weird carnival state. Look at Vidu's promotion of "a video will be released after a word on WeChat" and a bunch of "vibe coding" tools on Product Hunt, which are essentially using AI to wrap old concepts. The fatal flaw of these tools is that when big manufacturers 'model capabilities are iterating every week (such as the crazy upgrade of Byte Seed), the moat of application-level products is a joke. The AI video track may seem lively right now, but I bet 80% of startups will not survive for 18 months-either being crushed by big manufacturers or replaced by open source.

Then there is the computing power competition. Anthropic threw 20 billion yuan at Google Cloud this time, and Wuwen Xinqiong raised another 700 million yuan in financing, indicating that the industry consensus has been formed: the next three years will be a stock war for GPUs. But there is a huge risk here: everyone is betting that GM AGI will arrive as scheduled, but what if there is a deviation in the technical route? Now these investments are like optical fiber infrastructure in 2000-when we spend money, we feel that it is not enough, and only when the bubble burst did we discover that there is a serious overcapacity. What is particularly ironic is that while everyone is frantically accumulating computing power, Chrome actually secretly stuffing 4GB model files into users 'computers. Isn't this the most realistic contradiction of computing power? If the cloud computing power is not enough, the terminal's wool will be wiped out. Sooner or later, this kind of naughty operation will trigger a regulatory earthquake.

What worries me most is that the AI development ecosystem is splitting into two parallel universes: on the one hand, the myth of "AI agent" pursued by VCs (look at those fancy tools like Hermes and Phrony), and on the other hand, engineers are still honestly adjusting transformer parameters. The LLM financial trading frameworks in GitHub trending and Karpathy's skills list are truly valuable explorations. The danger now is that too many people are confused by the media's rendering of "PPT in one sentence" and ignore that what AI engineering requires is a solid data pipeline and verification system. It's like the story of the unlucky developer who was in demand-when the entire industry is chasing the wind, the ones who suffer the most are always the people who actually work.

Today's technological signal data reveals several important trends, the most interesting of which are the rapid development of artificial intelligence infrastructure, the rise of multimodal AI models, and the application of AI in consumer-grade products. These trends not only represent cutting-edge progress in technology, but also reflect profound changes in the industry landscape. In the next six months to a year, these trends will continue to evolve and have an important impact on different stakeholders.

Anthropic's large-scale investment in computing power and Wuwen Core's huge financing show that AI infrastructure construction is entering a new stage. The characteristic of this stage is that technology companies no longer only focus on the optimization of algorithms and models, but invest more resources in the construction of computing power and hardware facilities. Google Cloud has become Anthropic's main partner, which means that cloud service providers will occupy a more important position in future AI competitions. Cloud services not only provide AI companies with powerful computing resources, but also help them achieve efficient collaboration in data storage, processing and transmission. For the entire industry, this means that the threshold for AI applications will be further lowered, and more small and medium-sized companies can quickly enter the market with the help of cloud services and promote the popularization and application of AI technology.

At the same time, the rise of multimodal AI models is also worthy of attention. The Doubao-Seed-2.0-lite and mini versions released by Volcano Engine support a unified understanding of video, images and audio, which marks the development of AI technology from a single modality to a multimodal fusion. This integration not only improves the comprehensive capabilities of the model, but also opens the door for more application scenarios. For example, Vidu Claw generates high-quality videos through WeChat conversations, demonstrating the huge potential of AI in the field of content creation. In the next six months to a year, the performance of multimodal AI models will be further improved and application scenarios will be more extensive. Content creators, media companies and advertisers will be the main beneficiaries of this trend, using these models to quickly generate high-quality content and increase productivity and innovation capabilities.

However, with the rapid popularization of AI technology, privacy and compliance issues have become increasingly prominent. Google Chrome quietly pushed a 4GB local AI model file, raising user concerns about data security and privacy protection. Although this helps improve the user's experience in the browser, downloading without the user's consent may violate relevant privacy regulations. In the future, as AI technology becomes more widely used on personal devices, this compliance risk will become more common. Technology companies need to pay more attention to user privacy and establish transparent data use mechanisms to avoid potential legal disputes and social trust crises. Regulatory authorities should also strengthen supervision of these technologies to ensure that the development of technologies does not harm the legitimate rights and interests of users.

In short, today's signals show that AI infrastructure construction, multimodal AI models, and the application of AI in consumer products are entering a new stage of development. These trends not only bring new opportunities for technology companies, but also pose new challenges to cloud service providers, content creators and user privacy protection. In the next six months to a year, we may see more innovative applications and technological breakthroughs, but at the same time we also need to pay attention to compliance risks and social ethical issues to ensure that the development of technology can benefit more people.

The heat of today's technology ecology reminds me of the situation before the Nasdaq crash in 2000. There are three structural cracks hidden under the apparent prosperity, which deserve the vigilance of every participant.

The first hidden danger lies in the excessive arms race in AI infrastructure. Anthropic has invested US$200 billion to gamble on the expansion of computing power, and Wuwen Core has invested another 700 million yuan to sprint "AI-native infrastructure." This collective bet on GPUs is very similar to the 2018 blockchain mining bubble. When all players are betting on the "shovel business", what is actually exposed is the difficulty of commercialization of the application layer-36 Krypton's AI video Agent report put it bluntly: big factory models iterate every week to squeeze the living space of third parties, practitioners admitted that "you can only make quick money." What is even more dangerous is the cost-passing chain: cloud vendors collect computing power rents, start-ups use VC money to pay, and the final takeover may be secondary market investors. Historically, the outcomes of this model have been very similar: referring to the hardware carnival of HTC Vive and Oculus during the VR boom in 2016, the content ecosystem has never kept up.

Privacy compliance has become a ticking time bomb. Google Chrome's forced push of the 4GB native AI model is essentially building an unauthorized "spy base station" on the user's device. This is worse than the Cambridge Analytica incident-back then, Facebook required at least user click authorization, but now AI models reside directly on the hard drive. European privacy expert Alexander Hanff's accusation is no fuss: When the General Data Protection Regulations require "limited purposes" of data processing, multifunctional models such as the Gemini Nano are essentially compliance black holes. Apple's $250 million compensation for Siri's false publicity is just a prerequisite. I predict that the EU's fines for AI stealth data collection will exceed 10 billion next year.

The most hidden risks lie in the false prosperity of the developer ecosystem. Those Agent frameworks on GitHub that have gone crazy, such as TradingAgents, which won 2000 stars in two days, don't even have basic backtest reports in their documents. The V2EX forum was overwhelmed by the promotion of "$10 for comments", which is similar to the 2021 cryptocurrency airdrop craze. When developers like mattpocock use the personal Claude prompt as an open source project operation, it means that the industry has fallen into a "tool for tool's sake" loop. The real irony is that the "vibecoding" movement-known as "using voice to command agents to write code to free hands"-essentially uses a more complex technology stack to solve non-existent needs, just like using a rocket engine to drive a baby carriage.

These signals piece together the truth: the AI industry is using capital violence to cover up product hollowing out. The gap between infrastructure investment and business returns is growing, user privacy has become the default victim, and developers use false prosperity to paralyze themselves. When Volcano Engine promotes a "full-modal understanding model" in high profile but avoids the cost of reasoning, and when product managers boast of "using Agents to generate PPT without touching the computer" without saying that it takes three hours-these selective narratives will eventually be exposed by reality. The lesson of history is clear: before all technology bubbles burst, the last rave was always the shovel seller.