Thursday, July 9, 2026

OK, after reading these signals today, there are two things worth talking about most: First, the battle for talent in the AI circle has come to the surface, and second, AI applications have finally evolved from "making a toy" to "really helping people." It's the stage of working."

Let's talk about ICML 2026 first. Tsinghua University won the best paper and DeepMind won the time test award. These are routine operations at the technical level. What is really interesting is the "talent hunt" organized by China's major manufacturers in Seoul-booths, coffee bars, cruise ships, and all the tricks they can use are available. Companies such as Ali, Byte, and Xiaomi are throwing out diamond and platinum sponsorship not to publish a few papers, but to those who can implement it. Think about it, big factories now do not lack money or computing power. What they lack are researchers who can truly run models and solve engineering problems. Academic conferences have become job fairs, which shows one thing: the pure academic bubble is receding, and the talents who can be recruited are hard currency. This is good for ordinary developers. If you have real skills, you are likely to receive a call from a headhunter in the second half of this year.

Looking on the other side, the trend in the developer community has changed. The OpenWiki project can automatically write documents, Strix can automatically find vulnerabilities, and OpenAI and Claude Code have created a plug-in to call each other. These tools all solve the same core problem: allowing you to do less meaningless work. It's not the kind of fancy job of "helping you generate a poem", it's a real job-after you finish writing the code, it helps you make up documents, check security, and do reviews. This is what AI should look like when it comes to implementation: it's not to replace you, it's to take care of those annoying finishing tasks. Combined with that article about botsitting, in the future, every team must have someone "watching AI work." This is not a science fiction concept, it is a reality that will appear at your workstation in the second half of this year.

Finally, the TikTok layoffs article is particularly ironic when viewed together with the previous ones. On the one hand, big factories are spending money to grab AI talents, and on the other hand, they are laying off trust and security teams. What does it mean? This shows that capital is clearly divided into "cost center" and "profit center". If you can write code and build models, you are the emperor; if you engage in content review and compliance, you can leave at any time. This structural imbalance will only intensify in the short term. If bottom-level developers do not want to be internally optimized, the best way to protect themselves is to turn themselves into the one who can "take the lead". Don't just know how to adjust APIs, learn to see through the black box of models, and learn to help the business fill the hole. This is the core competitiveness in the next three years.

What deserves most attention today is the intersection of two forces: one is the interstellar expansion of AI infrastructure, and the other is AI's deep reconstruction of the job market. SpaceX applied to deploy a third-generation constellation of 100,000 satellites, and made it clear that it and the Starmind Space Data Center project are two separate projects, which means that the underlying logic is not a "satellite Internet" at all, but a pipeline for computing power transmission in the AI era. When Musk tried to use 100,000 nodes to turn the global data flow into a low-latency loop, Google's Pixel 11 family canceled the 128GB version, and Mercedes-Benz's pure electric GLC fully switched to AI. These seemingly independent signals were actually pointing to the same direction-AI is no longer just about the software layer, it is reshaping the entire physical infrastructure from chips to terminals, from space to wheels. Behind these events are mass production of Apple's folding screens, 90% layoffs in Douyin Indonesia, and a possible rise to US$2 billion for World Cup broadcasting rights in the United States. These events are all the same game about "capacity": whoever controls transmission and computing capacity will control the pricing power for the next decade.

The battle for talents has entered the white-hot "Law of the Jungle" stage. At ICML 2026, China's major manufacturers moved their booths and coffee shops to Seoul to systematically hunt for top AI talents. However, in the past, Huawei's talented teenager Li Bojie publicly complained about DeepSeek interviewer questioning his plagiarism. The internal examination has spread from technical competition to recruitment culture-When companies urgently need people who can "fight", interview standards have become distorted and defensive. This is exactly the same as the paradox of TikTok laying off employees globally but at the same time increasing the number in key positions. What's even more interesting is that there are two opposite trends in the developer community: on the one hand, the AI job search tool ai-job-search allows Claude to automatically send your resume, and on the other hand, AI hacking tools such as strix begin to automatically look for code vulnerabilities. This shows that AI is becoming both a "job seeker" and an "interviewer", and the self-iteration speed of technology far exceeds the adaptability of organizational systems.

The emergence of the new word botsitting reveals the real truth about employment: Microsoft and Morgan Stanley have begun to set up AI trainers, and white-collar workers spend an average of 6.4 hours a week doing "AI care". Graduates in 2026 will be more popular because they understand AI better. This is not like AI replacing humans, but more like AI creating a large number of "translation layer" positions-humans no longer write code, but correct code written by AI; they no longer do content review, but teach AI to do content review. But the hidden danger lies in that when OpenAI launched GPT-Live real-time voice and Grok 4.5 was released almost simultaneously, the generalization speed of AI capabilities was shortening the shelf life of these "translation layer" positions. Once the model itself can correct its own output, botsitting will disappear as quickly as typists did back then. The star-chain laying of infrastructure and the strong shocks in the talent market are essentially two sides of the same coin: the sooner people realize that "AI is not a tool but an ecology", the more they can find their place in the next round of reconstruction.

To be honest, when taken together, today's signals reveal not the excitement of a technological breakthrough, but a costly shift of anxiety. Who is paying and who is making money? I'm beginning to have the answer.

Let's start with Google Pixel 11 canceling the 128GB version. Storage costs are dropping at a rate of 15%-20% per year, and flash memory grains are now close to the floor price. However, in 2026, Google told consumers that there is no option for 128GB. Starting at 256GB, the price will naturally rise. This has nothing to do with technological evolution, but is a typical de-risk pricing strategy. The real cost is passed on to users, and the reasons behind this are the volume expansion of the AI functional system and the caching requirements of the end-side model. But do users really need to pay a few hundred dollars more per year for these so-called "smart" features? History tells us that when mobile phone manufacturers eliminated removable batteries and headphone jacks ten years ago, the reason given was also "for more advanced functions." What happened? The parts market has made a lot of money, and users 'migration costs and maintenance costs have doubled. Now it's time to store, the same script.

Let's look at ICML 2026's China factory stealing people. The venue of 11,000 people, with booths, coffee bars, and cruise parties, is like the madness of the mobile Internet bubble in 2015. At that time, he was an iOS engineer, but now he is an AI researcher. But the core question remains unchanged: What substantial profits are these high-paying talents creating? At today's AI application layer, most products are still based on free or subsidized models-GPT-Live is free, Grok 4.5 is open, and a large number of Hugging Face models are free for commercial use. Companies use venture capital and capital market money to pay these talents 'tens of millions of annual salaries, but the business model is still vague. Historically, the 2017-2019 autonomous driving talent war left behind a feather: star startups closed down, technical routes were falsified, and investors lost their money. Today's AI talent competition replicates the same path dependence, with profits flowing to conference sponsors, headhunters and high-end real estate agents rather than shareholders.

What alerts me most is the simultaneous launch of SpaceX's 100,000 satellite application and the AI data center plan. The two independent constellations together have more than 1.1 million satellites, a bet that Earth's orbital resources can be used without restrictions. But in reality, most prediction models seriously underestimate the manufacturing costs, launch costs, on-orbit maintenance costs, and de-orbit cleaning costs of a single satellite. The lesson of history is that low-cost satellites bring not only communications progress, but also orbital debris crises. The collision between Iridium and Cosmos 2251 in 2009 caused a surge in debris density in orbital areas that are still used at high frequencies. There are now more than 6000 satellites in orbit in the Star Chain, and thousands of collisions must be avoided every day. The operational risk of expanding to 100,000 units cannot be managed by human resources at all. It relies entirely on AI automated scheduling, and AI's fault tolerance rate in this field is almost zero. What makes money is SpaceX's launch services, while what pays for is the future aerospace industry and global astronomy research-affected observatories are already complaining about star chain tailing problems. This risk is seriously underestimated.

To sum up, today's technology news is about the same thing: using scale and technical narratives to cover up cost transfers and risk spillover. Storage costs are transferred to consumers, talent premiums are transferred to capital markets, and orbital resource risks are transferred to the entire space ecosystem. There were few people who really broke the game, but the people who paid the bill had already lined up.