Saturday, July 25, 2026

After watching a circle of science and technology news today, the screen was full of carnival of AI implementation, but real implementation and fake prosperity were mixed together, so we had to open it up and have a look.

There is a popular post saying that the product manager did not write a single line of code and relied on white AI to create a 3.2K Star open source tool. Just take a look at this cool article, don't take it seriously. Look at the other side. The brothers who engage in AI customer service in the enterprise are having a headache. They have changed from managing 100 people to managing 1000 AI concurrent operations. The evaluation system has completely collapsed. From semantic understanding to RAG retrieval, every link has to be re-quantified. This is called engineering reality. Now, FDE (Forward Deployment Engineer) positions have exploded, starting with an annual salary of 2 million yuan. Why? Because the universal model is not universal at all in enterprises, we must rely on human flesh to fill the pits at customer sites. Data infrastructure is the invisible ceiling. Even the AI interviewer has come out. It sounds bluffing, but in essence it is still RAG plus rule matching. It is still too early to completely replace human HR.

Let's talk about open source. Microsoft jointly issued a statement saying that the weight of open source is very important to the AI ecosystem, and American startups are also anxious, asking the government not to block China's open source AI. These capitalists are smart. Closed-source models are too expensive and easily stuck. Everyone is counting on open source prostitution to create ecology. However, the bottom line of the project cannot be lost. Today, there was news that a security camera's login page was directly hard-coded with a GitHub Admin Token. If this low-level error was placed in an enterprise-level AI system, data would be leaked in minutes. The Ministry of Industry and Information Technology has just reported 35 apps that collect personal information indiscriminately, including Songguo Travel, and supervision will only become stricter.

For ordinary developers, don't be fooled by those cool articles about zero-code creations. If you really want to rely on AI for a living, don't just focus on adjusting APIs. Either you can practice the internal skills of data infrastructure, or you can learn how to manage models instead of people. As for the open source model, it should be used, but you have to cover the security red line yourself, so don't wait until the system is hacked before crying.

The most intriguing thing about today's signals is not how strong the AI model is, but that AI is quietly changing from a "tool for doing things" to a "digital colleague." This transformation is the most worthy clue behind all the fragments today.

Look at the changes in the talent market. An FDE position has increased by 800%, with an annual salary of 2 million yuan. What does this signal mean? This shows that companies are no longer satisfied with buying an AI tool, but need people who can "deploy AI and make it really work." OpenAI has set up a $4 billion company to deploy, essentially exporting "AI employees." At the same time, Yuanqi AI launched "Super HR", which not only sifts resumes, but also goes to the recruitment platform to find, communicate, and score people like colleagues. This is no longer a tool, but a part of the process. Think back to Microsoft's joint release of "Open Source Weighting Model is Critical to a Healthy AI Ecosystem" and the founder of a startup on Hacker News's plea to the U.S. government not to cut off China's open source weighting model-the underlying logic of these signals is: AI is becoming the infrastructure, and the open source weighting model is the "factory setting" of this infrastructure. Whoever grasps this factory setting will grasp the "labor cost" of future digital employees.

Let's look at the breakthrough in brain-computer interfaces. Science Corp's PRIMA system has been approved for marketing in the European Union and can help patients with map atrophy restore some of their vision. This is more than just medical progress, it points to a more fundamental change: portals for information acquisition are moving away from fingers and eyes to direct interfaces with the nervous system. When AI, as a digital colleague, can help me handle 80% of my regular tasks, and brain-computer interfaces revolutionize the way humans obtain information, then human scarce resources will shift from "brain power" to "attention." LeBron James joined the 76ers for 8 million yuan in 2 years may seem like sports news, but it is actually the ultimate manifestation of the attention economy: when AI can do most of the work, the value of the top human athletes becomes even more prominent, because they provide irreplaceable, real, and story-filled focus of attention.

But the risks are also here. China Academy of Fine Arts assigns 30 sketches to freshmen. It may seem old-fashioned, but it is actually the most sobering confrontation in this era. When AI can generate any image, the Academy of Fine Arts instead emphasizes hand-drawn sketches, which is training something that AI cannot replace: the graininess of observing life and the ability to transform instant feelings into lines. This ability will become scarce in the future. The "word inflation" phenomenon in Chinese podcasts-native family, subjectivity, emotional value-is actually another anxiety: people are eager to explain themselves in new words, but AI can already understand you better than you do.

Half a year later, we will see more companies no longer recruiting "AI engineers" but hiring "AI managers". A year later, every company will have an "AI employee configuration table", just like the current human resources table. Those who benefit the most are those who can quickly use AI as a colleague rather than a tool, and those who suffer the most are those who think AI is just advanced Excel. When it's possible that your interviewer, your colleagues, and even your boss are all an AI, what you really need to prove is not what you will do, but why you, as a human being, are worth being hired.

In today's signal, the AI education white paper, the AI customer service evaluation system, and a large number of promotion of new models and tools have formed an obvious theme of "AI overheating". Tianli Qiming packages the "triangle of impossible" of education into an AGI solution, but does not provide quantifiable experimental data; similar articles often only list the scale of model parameters and computing power investment, but ignore the real cost of teaching costs, teacher training and student privacy. The past Internet bubble (1999 - 2000) and the AI investment peak from 2017 - 2021 both showed that technological concepts have been repeatedly amplified by capital, and the implementation benefits are often only 5 - 10% of the results can be realized. The current public model downloads (such as Solar‑Open2‑ 250B exceeding 500 times) are not equivalent to enterprise-level implementation. The real threshold is data governance, compliance auditing and continuous model maintenance. To this end, who is making money is particularly important: cloud service providers, labeling outsourcing companies and a few companies with large-scale computing power make profits from API call fees, computing power leasing, and model customization; while most universities, enterprises and ordinary users bear the risk of software and hardware procurement, training, and potential violations.

Looking at SpaceX's 13th starship test flight and the European Union-approved retinal brain-computer interface, both of which have been portrayed by the media as "technological revolutions." Each launch of a starship consumes hundreds of millions of dollars in research and development costs, but the rewards are still mainly government subsidies and a few commercial satellite orders; the marketing license for retina chips is still in the clinical trial stage, and long-term safety, surgical costs and subsequent maintenance costs have not yet been disclosed. Historically, similar high-investment projects (such as the nuclear fusion experiment in 2000 and the quantum computing prototype in 2010) have often attracted public attention in the short term, but left a large amount of outstanding debt in the gap between costs and actual benefits.

Taken together, what is hidden behind these hot spots is capital's pursuit of the concept of "high growth" rather than an objective assessment of technological maturity. Before deciding on investment, companies need to conduct detailed calculations of the full life cycle costs of the model, especially data annotation, compliance review and post-maintenance of model drift. At the regulatory level, recent export controls and personal information APP rectification show that policies are tightening and compliance costs will rise further. If these hidden costs and risks are not put on the table, the so-called technology dividend is likely to turn into another "bubble", with users ultimately paying for it, while the real profits will still be concentrated in the hands of a small number of capital and platforms.