Original title: 《LatePost Exclusive | Tang Jie’s internal memo says it will launch the Touch High plan, “If you don’t reach the summit, it’s failure”》
Original source: LatePost team

LatePost learned exclusively that today, Zhipu’s founder Tang Jie released an internal memo outlining Zhipu’s understanding of the next phase of competition for AGI. In the memo, Tang Jie said that Zhipu will continue to follow the so-called “anti-intuitive” path, launching the “Touch High（摸高）plan”—that is, continuing to focus on AGI research rather than short-term commercialization and value capture.

On the road to the AGI endpoint, there are a few mountain peaks that must be crossed—and they are also where today’s technological wave is most ferocious. The four peaks Tang Jie listed are:

-Long Horizon Task

-Autonomous Agent System

-Fully Self Training

-Extreme Safety Governance

Among them, extreme safety governance is especially emphasized. Zhipu plans to invest resources at the hundred-million-yuan scale to tackle mechanical interpretability head-on. This means clarifying the neural logic behind model decisions and pushing black-box systems toward transparent and interpretable systems.

On January 8 this year, Zhipu listed on the Hong Kong Stock Exchange as the first IPO stock for large models, with an offer price of HK$116.2. Over the following six months, Zhipu’s share price surged to as high as HK$2,980—more than 24 times the offer price—while its market cap once exceeded HK$1.3 trillion.

On July 8, more than 25M shares held by 11 cornerstone investors were scheduled to unlock. Shares with a market value of over HK$40 billion entered circulation, and the market’s original expectation of selling pressure failed to materialize—Zhipu’s share price rose instead of falling. The next day, Zhipu announced a share placement at HK$1,588 per share, a discount of about 13%, raising approximately HK$31.4 billion—this was the highest single-record share placement this year among Hong Kong-listed AI companies.

According to Zhipu’s official explanation, the raised funds are mainly used for base model R&D, building compute infrastructure, expanding commercialization, and laying out a global ecosystem. In the open letter, Tang Jie said its judgment about the leap in the “upper bound of intelligence” “is also the recognition we most want to deliver to everyone.”

In a series of comprehensive evaluations, Zhipu’s GLM-5.2 model has been widely recognized as already touching the capability boundary of the overseas leading-edge models. Moreover, due to its open-source nature, it is welcomed within the technical community.

Below is the full text of the internal memo:

《The Great Wave Has Come》

— To every person at Zhipu and our partners who care about the future of AI

Who we are: “Essence, anti-intuition, and focus”

Zhipu is never a company that chases the latest hype. It grew out of a lab, carrying the lab’s two decades of methodology. This methodology can be summarized in three words: essence, anti-intuition, and focus. Think deeply enough, and you dare to choose bravely enough; choose bravely enough, and you must endure long enough.

Looking back, almost every key choice we made once seemed “anti-intuitive.” In 2006, we sat on the bench with an academic search system running on desktop machines because we wanted to figure out that “mining the evolutionary mechanisms of disciplines” was something worth answering for a decade; from 2021 to 2022, when “making machines think like humans” was viewed by most people as a crazy plan like a moonshot, we pulled resources and bet on a trillion-parameter model, delivering GLM-130B—fully 18 months before ChatGPT detonated the world; and on the day of Zhipu’s H-share listing on January 8, 2026, we treated it as an entirely new starting point, firmly returning to foundation model research across the board and pushing hard for the next generation of models.

While others strike the bell, we return to zero. This is not a posture—it is a belief. Since the endpoint is AGI, short-term benefits or industry headwinds are only scenery along the path to the final destination.

What has carried us from then until today is a kind of extreme focus and a pure, idealistic commitment to truth. We took ten years to move our academic search system from a single desktop machine to ten million users. For large models, we’ve worked for nearly ten years as well—and we will continue to deepen and go on. Today, Zhipu is made up of people who are willing to question the essence, dare to act anti-intuitively, and can stay focused enough to do things through to the end—this is the source of Zhipu’s core competitiveness.

How we view this era: the upper bound of intelligence is being rewritten

If there’s one thing we learned over the past two decades, it’s this: real business opportunities never lie in fine-tuning products and business models, but in leaps in the upper bound of intelligence. This is our most fundamental judgment about the current AI transformation, and also the recognition we most want to pass on to everyone.

This transformation, at its core, is not a product innovation or a business model innovation—it is a technological revolution that raises the “upper bound of intelligence” itself. Whoever can push that bound up by even an inch first will be able to redefine the capability boundaries of every industry. All next-generation AI companies that focus on first principles are competing for that inch of breakthrough.

And the evolution of the upper bound of intelligence follows a clear path. Artificial intelligence is completing the leap from perceptual intelligence to cognitive intelligence. Machines are no longer just “seeing” and “hearing,” but beginning to “understand” and “reason.” And the next step points directly to AGI.

We have a simple yet stringent definition of AGI: AGI is not the intelligence of any single genius, but the sum of humanity’s intelligence level. It should possess the ability to create original knowledge at a “relativity theory” level—which is the only standard we use to measure whether we have truly reached the peak. On the road to this endpoint, there are several mountain peaks that must be crossed—and they are also where today’s technological wave is most intense:

The first peak: Long Horizon Task capability

The most exciting breakthrough today is teaching models to complete an extremely long task—not instant Q&A, but planning and execution spanning weeks, months, or even years. For example, a model can tirelessly search for vulnerabilities in software—at its core, it is learning the way of thinking of a top security expert, and then amplifying it through machine endurance.

The second peak: Completely autonomous agent systems

On top of long-horizon tasks, an agent collective that can independently drive, collaborate, and run 7×24 hours will become a new form of productive force. We once mentioned “OPC—one-person company,” but technical progress is moving faster than expected—we are heading toward a “fully automated company NPC.” The three problems that were once believed to require a paradigm shift to solve—Memory, Continual Learning, and Self-Judge—are now gradually being resolved under the dual drive of both technology and application. Long context and Retrieval-Augmented Generation (RAG) are approximating the shape of memory; increasing the iteration frequency of models itself is approximating continual learning; and leading models have already shown early sprouts of self-judgment.

The third peak: Self-Evolving

This is the hardest—and also the most alluring—peak. Training AI with AI has already taken shape: models write code themselves, clean and synthesize training data themselves, and train themselves. This may consume some compute, but it saves the most precious human labor and time. And in the era of large models, speed is the most important thing: rapid iteration directly widens the generation gap in cognition. When overseas top companies begin building compute clusters on the scale of one million and even two million chips, their real use might be to make models train themselves.

After crossing these three peaks, what happens?

AI will start learning what “me” is, and what self-awareness is. Beyond that, it will touch human emotions. Further still is consciousness itself. From perception to cognition, from cognition to generality, from generality to superintelligence (ASI)—this path has already been laid out. The great wave has come, and it is irreversible.

This is not just our view. In its report 《From AGI to ASI》, Google DeepMind made a cold judgment: even if the capability of a single model forever stays at human levels, as long as compute continues to grow, superintelligence could be forced into existence. They reason that if the number of globally runnable AGI instances grows at a tenfold annual pace, it will reach 13k after five years. These intelligent agents share the same underlying “brain,” achieve a hundredfold improvement in thinking efficiency, and replicate experiences at zero cost—at the group level, they are equivalent to ASI. In other words, moving from AGI to ASI requires breakthroughs at the algorithm level as well as the aggregation of ultra-large compute resources.

This irreversible trend will penetrate the entire technology stack from the top down. When AGI arrives, today’s applications may need to be rebuilt as AI-native, and even might no longer require those applications. Operating systems could be rewritten. In the future, when you open your computer, what you see may be an “LLM OS,” with all functions generated on demand. Deeper still is the challenge to the von Neumann architecture that has run for 80 years. Finance, law, e-commerce, the internet—no industry will be left out. Many friends come to me wanting to transform their enterprises and catch up with AI, but only a few truly recognize that this irreversible transformation has already begun.

The direction we are devoting all our effort to: “Touching High”

Once the trends are clear, what remains is choice. Zhipu’s choice, as always, is “anti-intuitive”: when the industry is generally accelerating commercialization, we decide to break upward.

We name this strategy the “Touch High（摸高） plan.” At the historical turning point where AI moves from perception and cognition to fully general intelligence, Zhipu will adopt a “Touch High” posture to challenge the physical and algorithmic limits of current technology. Over the next two years, we plan to invest strategically—not chasing short-term application monetization, but directly targeting the next high ground for AGI.

This investment will focus on four core engines:

First, long-horizon tasks. Move AI from “instant Q&A” to “grand engineering,” develop a new generation of memory architecture, enable models to run through the entire project lifecycle “learn while doing, do while learning, and remember while working,” and possess top-level capabilities to autonomously break down grand goals (such as “designing a new anti-cancer drug molecule”) into thousands of executable sub-tasks.

Second, autonomous agent systems. Move from “smart assistants” to “digital employees.” Build a society of agents containing thousands of different professional “personalities” and “skills,” enabling them to independently debate, collaborate, review code, and schedule resources—achieving “digital production power” at the level of “autonomous driving.”

Third, fully self training. As human high-quality data is about to be exhausted, convert compute into evolutionary fuel: build a high-quality synthetic data factory, achieve “knowledge from nothing” through AI vs AI adversarial game (Self-Play), and grant the system the capability to restructure its own code within a safe sandbox—so that the pace of evolution breaks free from the physical constraints of human engineers.

Fourth, extreme safety governance. This is the one I most want to emphasize among the four engines.

The stronger the capabilities, the more robust the safety constraint mechanisms must be. From the very beginning of Zhipu’s founding, we established the principle: AI must serve human well-being and serve national strategies. The company rejects outsourced safety patching and insists on writing human ethics, social norms, and national laws and regulations as bottom-line axioms into the model’s value function. We plan to invest at a hundred-billion-yuan scale to tackle “mechanical interpretability,” clarifying the neural logic behind model decisions and pushing black-box systems toward transparent and interpretable systems. At the same time, we actively participate in international AI governance to prevent AI technologies from being misused.

This sense of urgency is not paranoia. When overseas cutting-edge top models temporarily delay full public release due to risk considerations, and their executives publicly warn about the far-reaching impact of AI—then the global power landscape will be reshaped profoundly. We should be more clear-minded: the realization of superintelligence and research on super-alignment must advance in parallel. This is also the proposition we repeatedly revisit when facing disruptive technologies. History has repeatedly shown that when a technology reaches a level of power sufficient to change the course of civilization, safety is no longer an accessory—it becomes the fundamental prerequisite for the technology to persist and be allowed to be applied.

Open ecosystem: the underlying logic of intelligent universal access and safety governance

We have always believed that as artificial intelligence—the strategic technology that leads the future—its long-term development cannot be separated from an open, collaborative industrial ecosystem. The value of frontier intelligence is not only in the technological breakthroughs themselves, but also in whether it can broadly empower industries and benefit every developer. We firmly believe that real safety is not built on technical closure and barriers, but comes from widespread co-building, sharing, and oversight in the sunlight.

Precisely based on our deep recognition of technology universal accessibility, Zhipu has delivered its own strategic answer. Recently, we released what is currently our strongest open-source model, GLM-5.2. It supports truly usable million (1M) context windows, continues to stay ahead in long-horizon tasks, opens to the entire user base, and will be officially open-sourced under the most permissive MIT license—anyone can download, deploy, and use it for commercial purposes, with no subject-based restrictions. This is the company’s firm stance expressed in the form of a product.

We choose to believe in another path: frontier intelligence should not belong only to a few people, nor should it be taken back by a few rules at any time. It should be open, usable, buildable, and serve every developer.

This is not contradictory to “Touch High.” It’s actually two sides of one coin: one hand reaches upward to Touch High and challenge the limits of intelligence; the other hand paves the way downward, making the most cutting-edge capabilities as open and widely accessible as possible. The height we touch belongs to all humanity, and the road we build belongs to everyone.

Conclusion: Why now, and why it’s us

Some people may ask: after Zhipu goes public, why continue to pour core resources into “Touch High” in the most uncertain direction? Because we believe in a plain truth: true summit conquerors will build the road as they build the mountain.

The essence we worked out was first consolidated into the consensus of hundreds of scientists through the “悟道 large model” project, and then became the cornerstone for a generation of entrepreneurs to launch through Zhipu’s industrial investment and the entire ecosystem. Today, we want to build this road higher and wider—so high that it can protect ourselves and guard national security; so high that humans have the opportunity to explore more unknowns and the mysteries of the universe; and so wide that every developer and every team can climb up.

In the AGI era, things that were once out of reach for the first time have become possible. This is our generation’s greatest fortune as Chinese people, and also our heaviest responsibility.

The great wave has come, and the trend is irreversible. Zhipu wants to be the one who faces the oncoming wave and Touch High upward.

If you don’t reach the summit, it’s failure.

This time, the height we will touch is the one that belongs to all humanity.

Zhipu founder Tang Jie

July 11, 2026

Original link

Click to learn more about LyDing BlockBeats recruiting for positions

Welcome to join LyDing BlockBeats official community:

Telegram subscription group: https://t.me/theblockbeats

Telegram group chat: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia