Sunday, August 9, 2026

There are two things worth talking about most today. One is that AMD acquired Taalas to engrave the model into the chip, and the other is that byte refused to use the distillation shortcut to strengthen the ten-trillion-parameter model. These two things may seem to be incompatible, but in fact they tell the same story: the competition for AI has changed from "who can refine models" to "who can kill whom at the cost of computing power."

AMD made a tough move. Engraving the model directly into the silicon wafer for reasoning essentially solidifies the software into hardware, and the reasoning delay and energy consumption are directly pulled to the physical limit. This idea is completely different from general-purpose computing like GPUs, which takes the path of application-specific integrated circuits. But the pit in the engineering is also obvious. Once the model is engraved in, it cannot be changed. If the architecture is iterated, your chip will be scrap iron. So this thing is only suitable for those basic model scenarios that have converged and high-profile usage, such as specific tasks on the reasoning side. There will be no impact on ordinary developers in the short term, but this shows a trend: the lower limit of reasoning costs is being redefined by hardware, and the cost of running models in the future may be an order of magnitude lower than you think.

The byte side is more interesting. Zhang Yiming refused to distil and asked himself to train the model with ten trillion parameters from scratch. This attitude is regarded as a contrarian in the current circle of distilling everywhere. Distillation is really fast. Use other people's models as a teacher and you can reach the 80% level in a few weeks, but the ceiling is also sealed by others. What Byte wants is that the Seed model can enter the world's first echelon. This goal will never be achieved through distillation. However, the training cost of ten trillion parameters is astronomical, and any loss in computing power, data, and engineering stability will be a disaster. If you pass this path, you will be the next OpenAI-level player, but if you fail to pass, you will lose tens of billions. But then again, when you come to a company of this size, not betting on this is the biggest risk.

Let's talk about the hot post on HN,"Code is never the most difficult part." This is true and not true. For senior engineers, code is really not the most difficult thing. Requirements understanding, system design, and cross-team collaboration are. But for a large number of middle and junior developers, writing code itself is a bottleneck, and AI programming tools are helping them overcome this hurdle. Today, GitHub trending has a bunch of peripheral tools for Claude Code, OmniRoute has created a unified gateway with more than 290 providers, and book-to-skill has directly turned technical PDF into Claude Code's skill pack. The explosion of these tools shows one thing: the AI-assisted programming ecosystem is rapidly taking shape, but the real value lies not in the code generation itself, but in the rearranging of developers 'workflows. Tools that only do code completion will soon be eliminated, and tools that can deeply understand the project context and remember the history of team decisions will have a future.

My judgment is that the AI infrastructure layer is undergoing a hard-landing cost reconstruction. Whether it is chip-level solidification or ten trillion parameter training, it is pushing the threshold of AI to both ends: the bottom layer is getting more and more expensive and more professional., the application layer is getting cheaper and more inclusive. For ordinary developers, the most important thing to do now is not to train deconvolution models, but to figure out which links in their business scenarios can be completely reconstructed with AI, and which links will cause trouble if used. Don't be fooled by those stories about reports being produced in seconds. There is another article that said that it is still too late for AI to produce reports, because the bottleneck is not in the generation speed at all, but in the short action window of the business process itself. No matter how fast the tools are, it will be useless if people cannot keep up.

Judging from today's signals, the two core trends are "the competition between AI scale and long-term doctrine" and "traditional industries are torn apart by new infrastructure." At the internal meeting of ByteDance, Zhang Yiming publicly refused to use distillation to catch up with cutting-edge big language models, emphasizing that the model should be ranked among the world's top echelon during the Seed stage; at the same time, DeepSeek, AMD and OpenAI continued to invest heavily in model parameters and chip integration, and AMD even acquired Taalas, which specializes in carving models into silicon wafers. Behind it is a clear underlying trend: AI research and development is moving from "short-term commercialization" to "extreme scale" and "hardware collaboration". In the short term, the cost advantages of distillation technology can still keep some platforms competitive, but in the long run, only companies that form barriers in computing power, data and model depth can dominate the red sea of generative AI. For the industry, this means that capital and talent will further gather in large model research and development. If startups want to survive, they must find breakthroughs in vertical segmentation scenarios, data governance or model security, rather than blindly following the superposition of model capacity.

Another significant clue is the phenomenon of traditional heavy asset industries being torn apart by new infrastructure. There is a sharp contrast between Hughes 'bankruptcy and SpaceX's soaring performance in the satellite Internet field, indicating that the low-orbit satellite satellite chain has subverted the broadband landscape originally centered on ground base stations; similarly, car dealers generally lose money and reduce profits, behind which is the rise of direct sales of new energy vehicle companies, software-defined vehicles and autonomous driving platforms. The entry of Huawei's electric drives and lidar into GAC Aian's product lines is a concrete manifestation of the integration of software and hardware and the reengineering of supply chains. Half a year later, the integration of these two tracks will become more obvious: the large-scale vehicle-mounted model generated by AI will be directly deployed on low-orbit satellites to achieve low latency for the entire link of the Internet of Vehicles; accordingly, if traditional dealers do not transform into Data service providers or localized AI application providers will face further marginalization.

At the value level, the technology monopoly brought about by long-term investment in AI will provide excess profits to companies with complete closed loop of software and hardware; the collaboration between satellite Internet and new energy vehicles may spawn a new ecosystem of "aerospace-vehicle connectivity" and attract cross-border capital. The risk is that the computing power requirements and energy consumption of large-scale models are still costs that cannot be ignored, and regulatory review of model security and data privacy may tighten within a year; at the same time, the transformation costs of traditional industries are high, employee retraining and The painful period of supply chain restructuring will lead to short-term unemployment and industry fluctuations. Overall, in the next six months to a year, those companies that can deploy ahead of schedule at the intersection of AI infrastructure and new infrastructure will become new industry leaders; while players who rely on old business models and lack technological upgrades may be quickly marginalized by the market.

In today's science and technology news, we can see several topics worthy of attention. They not only reflect the current trends in scientific and technological development, but also reveal potential risks and challenges.

First of all, the T1 e-sports club raised the second team's first HLE match in jungler Painter, revealing the fierce competition and constant change characteristics of the e-sports industry. As an emerging sport, e-sports is as competitive as traditional sports. However, behind this fierce competition, is there a hidden neglect of the health and long-term development of players? After all, e-sports players have relatively short careers, and how to balance competitive performance and player well-being is a question that the e-sports industry needs to ponder deeply.

Secondly, ByteDance refuses to take distillation shortcuts and adheres to a long-term development strategy, which reflects the firm stance of the company in the face of market competition. In the field of artificial intelligence, technology updates are extremely fast, and competition among companies is extremely fierce. ByteDance chose to adhere to its own technical route, demonstrating its confidence in technological development and pursuit of long-term value. However, whether this long-term approach can bring the expected returns will still need to be tested with time.

Moreover, car dealers have left the market in batches and the profits of the automobile industry chain have declined, reflecting the challenges facing the automobile industry. With the rise of new energy vehicles, the traditional automobile industry is undergoing a profound transformation. In this transformation process, car dealers are under increasing pressure to survive. Behind this is the competition between new energy vehicles and traditional vehicles in terms of technology, cost, and market. For the automotive industry, how to adapt to this change and achieve sustainable development is an urgent issue to be solved at present.

To sum up, today's science and technology news reflects the following characteristics and problems:

1. is fiercely competitive: There is fierce competition in e-sports, artificial intelligence, and automotive industries, and companies and individuals need to continuously improve their capabilities to meet challenges.

2. Long-term Doctrine: In the rapidly developing field of science and technology, adhering to a long-term Doctrine development strategy will help enterprises achieve sustainable development.

3. Changes and Challenges: The rise of new energy vehicles has brought huge challenges to the traditional automobile industry. How to adapt to changes and achieve sustainable development is a question that the automobile industry needs to ponder deeply.

4. Risks and Opportunities: In the process of technological development, risks and opportunities coexist. Enterprises and individuals need to be risk aware, seize opportunities, and realize their own value.

In short, today's science and technology news provides us with rich information and allows us to see the trends and challenges of scientific and technological development. In future development, we need to pay attention to these trends and challenges and actively respond to them to achieve harmonious coexistence between science and technology and society.