Wednesday, June 24, 2026

Recently, after reading a tour of technical developments, my biggest feeling is that the "big steel refining" during the AI bubble period is finally coming to an end, and the engineering world is beginning to return to pragmatism of meticulous attention.

Look at the popular projects on GitHub now, such as headroom that compresses tool output and feeds logs to large models, claiming to save 60% to 95% of tokens. There is also a guy called ponytail, which directly makes AI Agents "not write code if they can" like an old fox. What does this mean? This shows that developers are scared of the token bills and illusions of the big model. Nowadays, when developing AI applications, we no longer focus on how big the model you call is, but on how you "feed" the model. If context management is not done well, no matter how strong the model is, it will be a money-burning machine. The Bytefa Bean Bag 2.1Pro is also the same way. It doesn't brag about "enterprise-level stable operation" and "project delivery". To put it bluntly, it just tells bosses that this thing can work without causing trouble. Ordinary developers really don't need to go to the bottom layer now and study more on how to compress contexts and how to use MCP to remember code base. This is the real skill that can be implemented.

Looking at the commercialization of large factories, it is becoming more and more practical. Huawei Smart Driving not only raises prices, but also provides a "bottom-up" guarantee. This move is too cruel, which is equivalent to directly turning smart driving from a marketing gimmick into a responsible subject. In the past, accidents were blamed on the driver, but now Huawei dares to say that I will take care of accidents in legal areas. This technical confidence is one aspect. The more important thing is to build a commercial moat to force friends who dare not take the bait into a corner. Apple's iOS 27 has also learned to be smart. Instead of telling you how smart Siri is, it will directly engage in non-influencing AI, such as automatic account sharing and automatic password updating, and stuff it all into the bottom of the system. Users cannot perceive the existence of AI at all, only feel that the mobile phone is easy to use. This shows that the end result of end-side AI is not to build a chat robot, but to quietly solve trivial pain points.

So stop being superstitious about subverting the world. Now, you are the real boss who can write down the token bill and quietly solve system-level pain points.

Today's technological signals have pieced together a clear underlying trend: artificial intelligence is shedding the cloak of dialogue boxes and fully diving into the bottom of the operating system and the physical world. Apple deliberately circumvents traditional conversational AI in iOS 27 and shifts to system-level non-influence intelligence. This is essentially the same thing as Huawei's courage to carry out accidents against high-end smart driving and the breakthrough of the universal cerebellum model of humanoid robots. The end of AI is not to become a smarter chat partner, but to become the autonomous driving system of the digital and physical worlds. When Apple automatically handles complex tasks at the bottom level and Huawei assumes driving responsibilities in the physical world, it means that AI is changing from a consultant to an executor or even a responsible person. Half a year later, companies that still use AI as a dialogue box will be marginalized. The real winners will be infrastructure players who can seamlessly weave intelligence into their lives and workflows.

At the same time, the marginal effects of the single-body model arms race are declining sharply. Google lost two AI trump cards in a row, causing its market value to plummet. SpaceX issued bonds for its AI ambitions but was voted with its feet. These all imply that the myth of the single model that vigorously works miracles is being shattered. Instead, OpenRouter's multi-model collaborative equalization and a large number of agent memory compression and token optimization tools have emerged in the developer community. The industry's focus is shifting from building the largest brain to building the smartest neural network. One year later, the stacking of computing power and parameters will no longer be the core barrier. How to efficiently route, orchestrate and schedule multiple models will become a new moat at the application layer. Those model dispatchers with architectural instincts will receive a higher business premium than giants who simply pile up computing power.

Seeing that Huawei said it would provide smart driving with the bottom line, SpaceX evaporated US$600 billion in three days, and Google dropped US$1.5 trillion overnight due to the departure of two AI giants. I suddenly had the feeling of returning to 1999. It's not that these companies will go to the same end as dot-com back then, but that this craze of "one piece of news can make the market value fluctuate violently" is too familiar.

Huawei's smart driving is on the surface responsible to consumers, but behind it is the cruel reality of industry competition. Smart driving packages have increased from 32,000 to 36,000, and the landing price has increased from 12,000 to 15,000. This increase is tied to the new "protection and service rights". To put it bluntly, Huawei is using price increases to cover the risk cost of accident compensation. The question is, how to define the so-called "legal driving roads and accidents occurring within the area"? Where are the boundaries of division of responsibilities? If something goes wrong, is it Huawei's system, the integration of car companies, or the driver's judgment wrong? If these details are not clarified, it will be a confusing account. In the end, either the premium cost will be passed on to consumers, or the responsibility determination will be entangled. What is even more noteworthy is that if all smart driving manufacturers follow suit, will this industry become a war of attrition in which "whoever cannot bear the cost of claims first will be eliminated"?

SpaceX's stock price collapse and Google's brain drain point to the same weakness of the AI bubble: the human factor. SpaceX fell 23%, and its market value evaporated by US$600 billion. The trigger point was that it wanted to issue bonds to support its AI business, and the market panicked. What does this mean? Investors are beginning to question whether the leveraged AI narrative can last. Google is even more straightforward. The father of Transformer and the Nobel Prize winner both left. No matter how much computing power is piled up, if the core architectural genius cannot be retained, what is the significance of the arms race? A very cruel reality is that the scarcest thing in the AI industry is not GPUs, but the dozens of people who truly understand the underlying architecture and can make breakthrough innovations. The Zhao family said that "the most expensive thing in the 21st century is talent." After looking at Google's experience, I think this sentence may need to be changed in the AI era and changed to "the most expensive thing in the 21st century is the two or three people who can make your stock price drop by 1.5 trillion yuan."

Another signal that is not too obvious today but worth pondering is OpenRouter's "multi-model collaborative equalization". Fable 5 was banned, and it immediately used a combination punch to top it, claiming to have made the "smartest composite AI." The essence of this is that when the single strongest model is not available for political or business reasons, the industry is using systems engineering ideas to circumvent restrictions. Multi-model collaboration sounds great, but the actual effect is likely to be to reach or even surpass single models on certain specific tasks, while showing cowardice on coherent tasks that require deep reasoning. However, if this direction really works, it will have a big impact on Nvidia's computing power monopoly narrative. By replacing the sky-high supermodel with a cheaper model combination, the person who calculates the accounts will calculate faster than anyone else.

At the end of the day, the picture these signals pieced together is that AI has moved from a technology race to a cost race and a trust race. Huawei is betting that consumers are willing to pay for the bottom, SpaceX is betting that the market will tolerate it to leverage, Google is betting that its organizational culture can still attract talents, and OpenRouter is betting that combinatorial AI can break monopoly. In every gambling game, someone is making money and someone is paying the bill. At times like this, what needs to be most vigilant is not the technology itself, but the narratives that package complex questions into simple answers.