Saturday, July 18, 2026

The two things worth mentioning in today's signal are the two things: AI Agents are finally starting to do practical things, and the open source model has taken to new heights.

Let's talk about Agent implementation first. The biggest change in WAIC this year is not which booth is more dazzling, but that the focus of the topic of the entire industry has shifted significantly backward. Last year, we were still comparing whose parameters were bigger and whose Demo was more amazing. This year, everyone began to ask a simpler question: Can this thing be incorporated into real business processes and be verified by results? Yu Linyi of Senbo Technology said that AI applications focus not only on technology, but also on an evidence-effective business closed-loop. This is on the point. Anthropic's internal team used Claude Cowork's case to be more convincing. One of the two people reduced the reporting work of two days a week to two hours, and the other directly built the activity line and handed it to the Agent to run it automatically. This is what an Agent should do. It's not chatting with you, it's working for you.

But don't get excited too soon. The real pitfalls for agents lie in boundaries and reliability. If you let it run automatically on the assembly line, which step is hallucinating and which step is wrong, who will take the lead? Codex's hidden functions may sound cool, such as browser import, computer control, and memory management, but when put in a production environment, each of them is a potential security risk. Microsoft is developing Mythos-like AI Vulnerability Detection Tools, which shows that the industry itself is aware of the security issues of AI code generation. Agent can be used, but there is still a distance from using it safely.

Look at Kimi K3 again. The Dark Side of the Moon directly open-source a 2.8 trillion parameter model, 1 million token contexts, and natively supports visual understanding. Someone on V2EX used it to write a web version of macOS, and the effect was quite good. This shows that the capabilities of open source models are indeed rapidly approaching closed source. But how can ordinary developers use a model with 3 trillion parameters? Unless you have enough graphics cards or use cloud APIs. What's interesting is that someone on Show HN used SSD streaming to run Qwen 3.6 MoE on the 16GB M1 Pro. This idea is worth paying attention to. The hardware is not enough for SSD to make up for it. Although it is not elegant, it is indeed a realistic solution for the transition period.

Finally, I would like to mention AWS's $1.7 billion billing error. There are almost 1000 likes on HN. The billing complexity of cloud services has reached a point where even AWS itself cannot figure it out. This is a wake-up call for all teams that rely on the cloud: there may be a real problem with your bill, so check it quickly.

Among today's technological signals, the most worthy topics are "The War on AI Terminals" and "The Anxiety of Landing of Agents." Behind these two clues lies the same larger story: the critical point of technology from showmanship to practicality. The point is coming, but who can cross it is far from clear.

Let's first look at the war on AI terminals. The appearance of STEPX Neo at WAIC was a landmark event. This "large-model native smartphone phone" is no longer a simple hardware + software superposition, but an attempt to redefine the paradigm of human-computer interaction-"trustworthy, controllable, reversible, and visible." This positioning directly compares OpenAI and Google's hardware layout, but more subtly, the logic behind it contrasts sharply with Anthropic's "don't make hardware" strategy. Anthropic gambled that "the model is strong enough, there is no need to occupy the entrance," while STEPX does the opposite, trying to lock in users 'usage scenarios through hardware. The core of this war is not technology, but business models: hardware entrances mean traffic and data, but also huge costs and risks. STEPX has great ambitions, but the challenges it faces are also obvious-how to carve a way through the cracks of giants such as Apple, Huawei, and Xiaomi? More importantly, can it truly solve the pain point of the implementation of "intelligent agents", or is it just another gimmick for "AI mobile phones"?

This directly leads to the second theme: the anxiety of landing agents. From Claude Cowork to Codex's hidden features, to Tencent's two-line war (Mixed Origin vs. WeLM), we see AI changing from "being able to chat" to "being able to work." But the problem is that these tools are glamorous in the Demo, but they face the embarrassment of the "last mile" when they actually land. Anthropic's case demonstrates the potential of Agent workflows, but it also exposes a cruel reality: Most companies don't need "full automation," but rather "controllable semi-automation." Behind WeLM's technical blog and Xiaohongshu account is Tencent's anxiety about "business closed-loop"-no matter how strong AI is, if it cannot be embedded into real workflows, it will be just another fancy tool. Yu Linyi, chairman of Senbo Technology, has a very representative view: "AI applications are not just about technology, but also an empirical and effective business closed-loop." This means that AI entrepreneurship in 2026 is no longer a model war, but a "know-how" competition. Only who can understand industry pain points faster and embed AI into specific business processes can survive.

The general trend behind these two themes is that AI is shifting from "technology-driven" to "scene-driven". In the past few years, we have seen an arms race in model parameters (Kimi K3's 2.8 trillion parameter is an example), but starting in 2026, technology's "performance" is no longer the only measure, or even the most important criterion. Whether STEPX's "reversible visibility" or Anthropic's "real working scenarios", they both emphasize the same thing: controllability of the user experience. This is similar to the mobile Internet in the 2010s-the war at that time was not about "whose mobile phone was faster" but about "whose ecosystem was more complete." But AI is far more complex than mobile phones because it involves ethical and legal issues of "human-computer symbiosis." For example, Codex's "computer control" feature may seem cool, but if an Agent manipulates files on your computer on his own, who will be held accountable? The "reversible" design emphasized by STEPX may be to deal with this risk.

The impact of this change on the industry landscape will be far-reaching. First, the boundaries between hardware vendors and model vendors become blurred. Companies like STEPX may become "Apple in the AI era", locking users in through hardware, while companies like Anthropic may become "Intel in the AI era", focusing on providing underlying capabilities. Second, the window of opportunity for startups is narrowing. In the past year, financing for embodied intelligence has soared fivefold, but the report "2026 Most Investors Focus on Artificial Intelligence/embodied intelligence Enterprises 50" also mentioned that the lack of a "business closed loop" is the biggest risk. This means that companies that only show off their skills and have no ability to implement scenarios will face large-scale elimination in 2027. Finally, the advantages of giants will be further expanded. Tencent's two-track war (Hunyuan +WeLM) is an example-giants have enough resources to explore multiple paths at the same time, while startups can only all-in one track.

Half a year to a year later, we may see the following key changes: First, there will be a bubble of "pseudo-innovation" in AI terminals. Many companies will launch "AI phones" and "AI glasses", but most of them are just a simple superposition of hardware + APIs, and there may be only 1-2 companies that can truly form an ecosystem. Second, the implementation of agents will shift from "tool-based" to "platform-based". Currently, most Agents are single-point tools (such as Claude Cowork), but in the future there will be an "Agent middle platform" that can flexibly combine different Agents to solve complex tasks. Third, there will be "AI fatigue" in the industry. When AI fails to deliver on its "revolutionary" promises, users and investors become more cautious, which will lead to a wave of startups that fail.

In this process, the biggest beneficiaries may be those companies that "understand scenarios" rather than those that "understand technology." For example, although companies like Senbo Technology may not be the most advanced in technology, they understand the pain points of specific scenarios such as marketing and customer service, and can embed AI into real business processes. The biggest losers may be those "technology-first" startups that may shine in demos but struggle to move forward in the real world. In addition, traditional hardware giants (such as Apple and Huawei) may become the biggest variables-if they can deeply integrate AI capabilities with existing ecosystems, they may redefine the landscape of this war.

In the end, the outcome of this war may not depend on technology, but on who can better understand "people." STEPX's emphasis on "trustworthiness and controllability" and Anthropic's emphasis on "real working scenarios" are essentially answering the same question: In an era when AI is ubiquitous, how can humans maintain a sense of control over technology? The answer to this question will determine the next decade of AI.

Among today's signals, the most eye-catching ones are the leap into the mobile phone battlefield, the open source of Kimi K3, and the five-fold surge in tailored smart financing. They seem to be telling different stories, but when put together, they spell out the most dangerous picture of the AI industry in 2026: everyone is fighting for the "entrance", but no one is willing to admit that the "entrance" itself may be a false proposition.

Let's talk about steps first. A model-making company suddenly released the AI-native mobile phone STEPX Neo on WAIC, claiming "human-computer symbiosis." It sounds sexy, but it reminds me of the "Internet mobile phone" bubble in 2015-at that time, everyone who did content, e-commerce, and social thought they could make a mobile phone. What happened? Hardware is an extremely cruel industrial chain. Every link in supply chain management, quality control, channels, and after-sales is a bloody pit. The stepping model ability may be good, but the mobile phone doesn't just stuff the model into the shell. The real risk is, when a company makes models, terminals, and ecology at the same time, to what extent will its energy be diluted? Historically, Google has built Nexus and Pixel, but it has never really shaken the mobile phone landscape;OpenAI was rumored to be making hardware last year, but in the end it was silent. Why doesn't Anthropic make hardware? They bet that "the model is strong enough, there is no need to occupy the entrance"-this judgment is clear.

Look at Kimi K3, with 2.8 trillion parameters, the world's first 3 trillion open source model. The scale of the parameters is indeed shocking, but the 175 billion parameters of GPT-3 were already considered to be "ridiculously large" and now they have doubled 160 times. Here comes the question: Who can run? How much does the training cost? What is the cost of reasoning? In a real application scenario, when a model with 2.8 trillion parameters and a model with 28 billion parameters answer the task of "help me write an email", is the experience gap worth the cost? I noticed that the most popular models on Hugging Face are instead a small 27B model like Bonsai-27B. The open source community is voting with its feet: it's not that bigger is better, it's that it's enough. If the Kimi K3 just "can run but can't afford it", then it is more like a technology demonstration than a product.

Finally, look at the physical intelligence. Raising 93.5 billion yuan sounds like a golden track. But pay attention to one detail: The 36Krypton report specifically mentioned that Zhiyuan Research Institute bluntly stated in the "Top Ten AI Technology Trends in 2026" that "the business of specific intelligent startups is seriously homogenized." 92.9 billion yuan was invested in 322 companies, with an average of nearly 300 million yuan per company, but how many companies can make robots that can be truly mass-produced? How many companies just change the color of the robotic arm in the demo? I remember the commercial robot bubble in 2017. Countless companies relied on PPT to raise funds, but in the end, they couldn't even build a robot that could stably deliver food. History does not simply repeat, but rhymes.

Objectively speaking, there is only one logic behind these actions: the AI industry is shifting from "model competition" to "scenario implementation." But in the process of turning, too many companies regard "occupying entrances" as "creating value." Mobile phones are not the entrance, but what users are willing to use every day. Parameters are not products, but users 'ability to be willing to pay. Financing is not verification, users are willing to repurchase.

Today's signals tell me that the AI industry is experiencing a "bubble within a bubble"-a capital bubble superimposed on top of the technology bubble. There is only one question that needs to be answered most now: Who is really making money, and who is just using investors 'money to buy traffic?