Saturday, August 8, 2026

In today's hot news, there are several events that deserve special attention. The first is the annihilation of the national table tennis men's singles in the Yokohama Championship. This is not only a reflection of the individual level of the athletes, but also a comprehensive consideration of team, tactics, psychology and other factors. Secondly, there are reports that ByteDance refuses to take distillation shortcuts, which reflects the long-term attitude of technology companies in pursuing technological innovation. Finally, the "three supports and one support" exam cheating incident in Henan once again brought fairness and supervision issues to the forefront.

The defeat of the national table tennis men's singles in the Yokohama Championship appears to be inferior to others, but the deep-seated reasons are worth pondering. First of all, there may be tactical deficiencies, resulting in the inability to flexibly adjust when facing different opponents. Secondly, athletes may have problems such as excessive pressure and unstable psychological quality at the psychological level. This reminds us that while pursuing competitive results, we must also pay attention to the all-round development of athletes.

ByteDance refuses to take the shortcut of distillation and chooses to adhere to technological innovation, which reflects the long-term doctrine of a technology company. In the field of artificial intelligence, the accumulation of model parameters cannot bring about real breakthroughs, but requires in-depth exploration of algorithms and optimization of model structures. This attitude is commendable, but it also requires greater cost and patience.

The cheating incident in Henan's "three supports and one support" exam has once again exposed fairness and regulatory issues. Cheating in exams not only undermines fairness, but may also cause unfair competition for other candidates. This reminds relevant departments that when selecting talents, they must strengthen supervision of the examination process to ensure fairness and justice.

These three incidents may seem independent, but are actually related. They reflect the existence of issues such as fairness, supervision, and technological innovation in the fields of science and technology, sports, and education. These problems require us to start from multiple aspects and solve them together.

First of all, in the field of sports, we must pay attention to the cultivation of athletes 'psychological qualities and improve tactical levels. At the same time, we must also pay attention to the construction of coaching teams. Secondly, in the field of science and technology, enterprises must be brave in innovation, adhere to long-term doctrine, and also pay attention to social benefits. Finally, in the field of education, we must strengthen supervision of the examination process to ensure fairness and justice, and select truly outstanding talents.

In short, these three incidents have sounded a wake-up call for us and reminded us that while pursuing development, we cannot ignore values such as fairness, justice, and innovation. Only in this way can we achieve sustainable development and build a better future.

Sister foresight is here. With today's data, I see that several core trends are accelerating, and they echo each other and outline several key outlines of the future technological landscape.

First of all, the game between "long-term doctrine" and "large-scale efficiency" surrounding the technical route of the big model is entering a white-hot stage. ByteDance rejects the "distillation shortcut" and insists on investing in Seed model research and development from a longer-term perspective. This is a typical strategic choice with "heavy assets, high risks, and high returns." In contrast, Kimi K3 achieves amazing scaling efficiency with a MoE architecture with 2.8 trillion parameters, proving that beyond parameter scale, architectural innovation is the key to improving real value. What reflects behind this is the transformation of the entire AI industry from "stacking materials" to "intensive farming". The giants may still be pursuing absolute parameters, but rising stars like Kimi have achieved the effect of "breaking a thousand pounds with cleverness" through technical ingenuity. This differentiation indicates that in the future, AI competition will no longer be a pure competition of computing power or parameters, but a comprehensive competition of technical paths, model efficiency and business models. Half a year later, we will see more innovative architectures like Kimi emerge, and companies that blindly pursue scale but lack core technology breakthroughs may face growth bottlenecks.

Secondly, the application of AI Agents is moving from concepts to more specific "productivity tools" and is beginning to touch a wider range of business scenarios and personal lives. The details of OpenAI's slightly products have been exposed, emphasizing "vital" interactions, indicating that AI is trying to integrate more deeply into our home environment, evolving from an information assistant to a more emotionally connected partner. At the same time, the Thousand Questions Update installs Agents into computers and mobile phones, and the emergence of various "Agent OS" and "Agent harnesses" indicate that Agents are evolving from independent AI individuals to "digital employees" who can work together and manage tasks. For AI companies, this means that the reshaping of operating models is no longer a simple customer service, but the need to build a more complex customer success and product growth system. For ordinary users, this means improved work efficiency, but it may also bring the risk of over-reliance on AI and new challenges to data privacy and security. Within a year, we will see more enterprise-level and personal-level Agent applications implemented, and how to build a trustworthy and controllable Agent ecosystem will become the key.

Finally, the competition around AI computing power and infrastructure has entered a new stage. AMD's acquisition of Taalas directly targets hardware optimization of AI inference performance. Through the idea of a "silicon engraving model", it attempts to achieve breakthroughs in computing power at the underlying hardware level, which is more disruptive than simply improving GPU computing power. At the same time, DeepSeek API is about to increase its prices significantly. This is not only a reflection of cost, but also a pricing signal from the market for the supply of high-quality AI models. It implies that the commercialization of AI services is accelerating, but the cost pressure is still huge. AI gateway projects like OmniRoute bring together a large number of models and providers to try to reduce the cost of AI access for developers and enterprises, an effort to hedge computing power costs and technology silos. This multi-dimensional competition at the hardware, model, and service levels will ultimately affect the popularization speed and cost of AI technology, and also determine which companies can occupy a favorable position in this technological revolution.

In general, current scientific and technological signals show that AI is moving from the laboratory to the broader real world, and its development path shows the diversification of technical routes, deepening of application scenarios, and accelerated iteration of infrastructure. In the process, innovators will benefit from technology dividends, while those who remain self-reliant face the risk of being subverted.

The most alarming thing today is not which model has been swiped on the list again, but the absurd reality exposed by the combination of two seemingly unrelated signals. On the one hand, DeepSeek announced a sharp price increase, and on the other hand, a webmaster disclosed that 99% of the traffic on his 1.5 million-page website comes from crawlers. Putting these two things together, you will find that the core contradiction of the current AI industry is not a technological breakthrough at all, but a collapsing economic model.

What does DeepSeek's price increase mean? This shows that low-cost APIs were not a sustainable business model before, but burning money to grab shares. Now that capital is patiently bottoming out, it must move closer to profit. But the question is, who will pay for the price increase? The webmaster's experience gave a cruel answer: an increasing proportion of robots on the Internet are talking to robots. AI companies use crawlers to grab data to train models, and the content generated by the model is captured by another crawler, forming a huge amount of idle traffic. This means that a considerable part of the computing power, bandwidth, and electricity consumed by AI does not serve real human needs, but feeds a self-circulating system. Zhang Yiming said that he would not take the shortcut of distillation and spend 10 trillion yuan as a basic model. It sounds very long-term, but if a large amount of false traffic and false demand are mixed into the terminal demand, the return rate of this bet will be a big question mark.

The second thing worth mentioning is AMD's acquisition of Taalas and its intention to carve the model directly into silicon chips. This is essentially admitting that general-purpose GPU reasoning is too inefficient and has to take the dedicated chip route. This divergence occurs in every technology cycle in history: common platforms explode first and are then eroded by dedicated solutions. The problem is that once the model architecture is still iterating rapidly, engraving the model into the chip means that the hardware life cycle is extremely short and the iteration cost is extremely high. Who makes money here? Chip manufacturer. Who is paying the bill? AI companies that bet on the wrong model architecture.

As for Noema's article on why everyone in the technology industry is sad, the answer may be simple: practitioners themselves have a vague feeling that this tech boom has created a lot of activity, but has not yet created the same amount of real value. When the tide recedes, what really remains will not be the models with the largest parameters, but the products that solve the real problem and are willing to pay for it. Among these signals today, what I see is more noise and less signal.