A repost of a long post by Tang Jie, Tsinghua professor and founder of Zhipu AI: why technology now carries unprecedented weight in the AI era, how vision sets the ceiling, and why cognition matters more than either.
Repost note: The author is Tang Jie (Tsinghua professor, Zhipu AI founder, Weibo @唐杰THU). The original was first published on Weibo at 13:32 on June 29, 2026. This site republishes it with the original paragraphs and wording intact, without edits. Copyright belongs to the original author.
Many people believe a company’s success depends on management. That is right, and used to be very right, but the AI era may have changed it; everything has changed: end-to-end every day, flat every day, removing middle management every day. So many things have become AI-driven, and more anxiety comes from the fear of falling behind technologically. Engineers suddenly have their moment, and technology has become more important than management. Management is still useful, but if you do not understand technology—or even go deep into it—you do not know how to manage. What exactly are you managing?
In the AI era, almost every technical advance is exciting: pretraining infrastructure (loss balancing, indexpool, kernel fuse), Agentic RL, self-judge, OPD, fully self-training, and especially the paradigm shift in AI learning. A new technique often floods X and spreads around the world overnight. No earlier era had this density and intensity: you no longer wait years for one landmark paper; you chase a wave of new things every few days. That is also why technological backwardness is no longer “falling behind slowly”, but something that happens overnight—and why technology’s weight in the AI game has risen to an unprecedented level.
The size of a person’s vision has always been a measure of their ceiling. A broad vision does not guarantee that you will accomplish something great, but a narrow one certainly prevents it. The AI era often depends on imagination and positioning: grasp the main contradiction and make bold bets. Without vision, the plate becomes small and you naturally fall behind. “Let’s contract first, survive, and copy once someone else has done it” does not work in the AI era—by then you will already be off the table.
But cognition is the most important thing in AI. The core of AI’s development is technology. This is different from the business-model game of starting a company 20 years ago and the product game of 10 years ago. In the business-model game, you found a new social need, spent money to own the model, and kept others out; capital mattered most, and successful founders were often financiers or management experts. In the product game, you designed a psychological or physiological loop that kept users—games, video, or chat, anything that retained them and grew DAU; product managers and reading human nature mattered most, and most winners were product managers. Today’s AI game is different again: OpenAI and Anthropic repeatedly leapfrog each other, showing that the essence of the AI era is rapid technological progress. If you pause to polish a product, you may discover the next day that the underlying technology is obsolete and nobody uses the product; if you pause to think about the business model, the AI world has already been overturned again.
The endgame of AI is AGI, a dragon-hunting game. Along this road, actions aimed at hunting rabbits are not fundamental—unless you are truly starving, you should focus on the essence and the main line. But slogans are nowhere near enough. AGI is a vision, yet there has never been a standard answer for where it is, how to define it, or how to achieve it. The traditional academic sequence—define it slowly, design algorithms, run experiments, validate, prototype, then productize—certainly will not work. Nor can it be achieved by the old “beat him in a flash” approach, with a few people working overnight. AGI requires extreme enthusiasm and conviction about AGI (strong first-principles thinking), deep technical accumulation and control of details, and a group of simple, pure people who do not consume one another’s energy or dither. AI people in this era care less about so-called “experience” and more about being pure and free of internal friction. But being young does not equal being pure; surface-level youth can hide the essential importance of purity.
Technical judgment in AI is another challenge. Many people like ML and CS because they offer elegant theories and mathematical derivations. But AGI needs to be more original and native, and the mathematical derivations of AI have already surpassed most CS graduate students and most CS programmers. CS people once thought they stood above everything, that they were the creators of AI and all other fields would collapse. In reality, it may be our own knowledge system that is being comprehensively overturned. That is why CS seniority suddenly stops working: everyone in CS is flattened and managed by AI. Only by becoming an IC with forward-looking cognition for the AI era can you survive: iterate your knowledge quickly and improve your cognition quickly (Can you foresee AI five years from now? No. Two years? One year?).