Original title: 《Late Night Exclusive｜Tang Jie’s internal letter says it will launch the Touch High plan (“reach high”), with “not reaching the summit means failure”》
Original source: Late Night team

Late Night exclusively learned that today, Zhipu’s founder Tang Jie released an internal letter laying out Zhipu’s understanding of the next phase of competition for AGI. In the letter, Tang Jie said Zhipu will continue to adhere to the so-called “counterintuitive” route and launch the “Touch High plan,” meaning it will continue to focus on AGI research rather than short-term commercialization and monetization.

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

- Long Horizon Task

- Autonomous Agent System

- Fully Self Training

- Extreme safety governance

Among them, extreme safety governance is specially emphasized. Zhipu plans to投入 invest a resource pool on the scale of tens of billions of RMB to tackle mechanistic interpretability. This means clarifying the neural logic behind model decisions and pushing black-box systems toward transparent and explainable systems.

On January 8 this year, Zhipu listed on the Hong Kong Stock Exchange with an issuance price of HK$116.2 per share as the first LLM stock. Since then, over the next six months, Zhipu’s share price surged to a high of HK$2,980, up more than 24 times from the issuance price. At one point, its market cap briefly exceeded HK$13 trillion.

On July 8, more than 25 million shares held by 11 cornerstone investors were set to unlock. The shares representing a market value exceeding HK$40 billion entered the market. The market had originally expected selling pressure, but it didn’t materialize—Zhipu’s stock price rose instead of falling. The next day, Zhipu announced the placement of new shares at HK$1,588 per share—at a discount of about 13%—raising about HK$31.4 billion. This is the highest single-record placement by an AI company in Hong Kong stocks this year.

According to Zhipu’s official statement, the funds raised from the placement will be mainly used for the development of foundational models, construction of compute infrastructure, commercialization expansion, and global ecosystem planning. In the open letter, Tang Jie said that its judgment about the “intelligence upper bound leap” is “also the cognition we most want to convey to everyone.”

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

The following is the full text of the internal letter:

《The Giant Wave Has Arrived》

— To every person at Zhipu and partners who care about the future of artificial intelligence

Who we are: “Essence, counterintuitive, focus”

Zhipu is not a company that chases every hype cycle. It grew out of a laboratory, carrying the methodology from that lab’s twenty years. This methodology can be summarized in three words: essence, counterintuitive, focus—think deep enough, and you dare to choose boldly; choose boldly enough, and you must stay long enough.

Looking back at our path, almost every critical choice we made once seemed “counterintuitive.” In 2006, we kept our heads down and endured as an academic search system on desktop machines, because we wanted to figure out that it was worth answering for a decade: “digging into the mechanisms of disciplinary evolution.” In 2021 to 2022, when “making machines think like humans” was regarded by most people as a crazy plan akin to going to the moon, we reallocated resources and bet on hundreds of billions of parameters, creating GLM-130B—exactly one and a half years before ChatGPT ignited 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 and firmly returned comprehensively to foundational model research, throwing ourselves fully into the next-generation model.

While others ring the bell, we reset to zero. This is not a pose; it is a belief—since the endpoint is AGI, short-term interests or industry windfalls are only scenery along the way to the endgame.

What has supported us all the way to today is a kind of extreme focus and pure idealism. We took ten years to move the academic search system from a desktop machine to tens of millions of users. On the path of large models, we have also spent nearly ten years and will continue to deepen and keep going. Today’s Zhipu is a group of people who are willing to keep asking about essence, dare to be counterintuitive, 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 is one thing we learned over the past twenty years, it is this: real business opportunities have never been in the tweaking of products and business models, but in the leap of the intelligence upper bound. This is our most fundamental judgment about the AI transformation of today, and also the cognition we most want to pass to everyone.

This transformation, at its core, is not a product innovation or business model innovation. It is a technical revolution that raises the “intelligence upper bound” itself. Whoever can push that bound up by even an inch first will be able to redefine the capability boundaries of thousands of industries. All new-generation AI enterprises focused on first principles are competing for precisely that inch of breakthrough.

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

We have a simple yet demanding definition of AGI: AGI is not the intelligence of some single genius, but the sum of human-level intelligence across all humanity. It should have the ability to create original knowledge at a “relativity” level—this is the only standard we use to measure whether we have truly reached the peak. On the road to this endpoint, there are several mountains that must be crossed—and they are also where today’s technological wave is most fierce:

First mountain: Long Horizon Task capability

The most exciting breakthrough today is teaching models to complete an extremely long task—not immediate Q&A, but planning and execution spanning weeks, months, or even years. For example, a model can tirelessly find vulnerabilities in software; fundamentally, it is learning the way a top security expert thinks, and then amplifying it through the machine’s endurance.

Second mountain: Fully Autonomous Agent Systems

Above long-horizon tasks, a group of agents that can autonomously drive, collaborate, and operate 7×24 hours will become a new form of productive force. We have mentioned “OPC, one-person company,” but the pace of technology has moved faster than expected—we are heading toward an “NPC, fully automated company.” Three problems once thought to require paradigm shifts to solve—Memory, Continual Learning, and Self-Judge—are now being gradually resolved under the dual drive of technology and applications: long context and retrieval-augmented generation (RAG) are getting closer to the embryo of memory; increasing iteration frequency is itself approaching continual learning; and frontier models have begun to show early signs of self-judgment.

Third mountain: Self-Evolving

This is the most difficult—and most enticing—mountain. Training AI by AI has already taken shape: models write code themselves, clean and synthesize 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—rapid iteration directly widens the generational gap in cognition. When overseas top enterprises begin building compute clusters at the level of hundreds of thousands to even two hundred thousand chips, their true use-case may well be to let models train themselves.

What happens after we cross these three mountains?

AI will begin to learn what “I” is, and what self-cognition is. Further on, it will touch human emotions. Even farther is consciousness itself. From perception to cognition, from cognition to generality, from generality to superintelligence (ASI)—this road is already laid out. The giant wave has arrived, and it is irreversible.

This is not just our own opinion. In Google DeepMind’s report “From AGI to ASI,” they make a cold, clear conclusion: even if the capabilities of a single model forever stay at the human level, as long as compute continues growing, superintelligence may be squeezed out “brute-force.” They reason that if the world can run AGI instances growing at a tenfold speed per year, in five years there will be 13k of them. These agents that share the same underlying “brain,” increase thinking efficiency by a hundredfold, and replicate experience at zero cost—collectively at the group level are equivalent to ASI. In other words, the move from AGI to ASI requires not only algorithmic breakthroughs, but also the aggregation of massive compute resources.

This kind of irreversible trend will penetrate the entire technology stack from top to bottom. When AGI arrives, today’s applications may need to be restructured into AI-native experiences, and may even no longer require those applications. Operating systems may be rewritten: when you open your computer in the future, what you see will be an “LLM OS,” with all functions generated on demand. Deeper still is the challenge to the 80-year-old von Neumann architecture. Finance, law, e-commerce, the internet—no industry will be left out. Many friends come to me saying they want to transform enterprises and catch up with AI, but only a handful truly grasp that “this irreversible transformation has already started.”

The direction we are pouring everything into: “Touch High”

After recognizing the trend, what remains is choice. And Zhipu’s choice, as always, is “counterintuitive”—when the industry is generally speeding up commercialization, we decide to break upward.

We name this strategy the “Touch High plan.” At the historical node where artificial intelligence is crossing 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 strategically invest—not aiming for short-term application monetization, but directly targeting the next high ground of AGI.

This investment will focus on four core engines:

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

Second, autonomous agent systems. Move from “intelligent assistants” to “digital employees.” Build an agent society containing thousands of agents with different professional “personalities” and “skills.” Let them debate autonomously, collaborate, review code, and schedule resources to realize “digital production power” at the level of “autonomous driving.”

Third, fully self-training (Fully Self Training). As high-quality human data is about to run out, convert compute into fuel for evolution. Build high-quality synthetic data factories. Achieve the creation of knowledge “out of nothing” through AI-versus-AI adversarial play (Self-Play). And within secure sandboxes, endow the system with the ability to restructure its own code so that the pace of evolution can break free from the physical constraints of human engineers.

Fourth, extreme safety governance. Among the four engines, this is the one I most want to stress.

The stronger the capabilities, the more solid the safety constraint mechanisms must be. From the very beginning of its founding, Zhipu set out its principles: AI must serve human well-being and serve national strategies. The company rejects “plug-in” safety patch approaches, and insists on writing human ethics, social norms, and national laws and regulations as foundational axioms into the model’s value function. It plans to invest resources on the scale of tens of billions to tackle “mechanistic interpretability,” clarifying the neural logic behind model decisions and pushing black-box systems toward transparent and explainable systems. At the same time, it actively participates in international AI governance to prevent AI technologies from being misused.

This sense of urgency is not overthinking. When the overseas top frontier models are temporarily held back from full public release due to risk considerations, and their executives publicly warn about the profound impact of AI, it is even more important to stay clear-headed: research on achieving superintelligence and achieving super-alignment must be advanced in parallel. This is also a proposition we repeatedly revisit when facing disruptive technologies—history has repeatedly shown that when a technology reaches a magnitude capable of changing the course of civilization, safety is no longer an accessory, but the fundamental prerequisite for the technology to endure and be allowed to be applied.

Open ecosystem: the underlying logic of intelligent empowerment and safety governance

We have always believed that as artificial intelligence is a strategic technology leading the future, its long-term development cannot do without an open, collaborative industrial ecosystem. The value of frontier intelligence is not only in technological breakthroughs, but also in whether it can widely empower thousands of industries and benefit every developer. We firmly believe that real safety is not built on technical closure and barriers—it comes from broad co-building, sharing, and oversight in the sunlight.

Precisely based on our deep recognition of technological universal empowerment, Zhipu has delivered its own strategic answer. Recently, we released GLM-5.2, the open-source model with the strongest capabilities to date. It supports a truly usable million (1M) context, continues to maintain leadership on long-horizon tasks, opens it to the full range of users, 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, and should not be taken back at any time under the control of a few rules. It should be open, usable, and buildable—and serve every developer.

This is not in contradiction with “Touch High.” In fact, it is two sides of the same coin: one hand touches high above, challenging the limits of intelligence; the other hand paves the way below, so the most advanced capabilities can be as open and universally accessible as possible. The height we touch belongs to all humanity, and the path we cultivate belongs to everyone.

Conclusion: Why now, and why us

Someone may ask: after Zhipu went public, why keep pouring core resources and “Touch High” toward the most uncertain direction? Because we believe in a simple principle: real summit-makers will turn the mountain into a road.

The essence we have thought through clearly—once consolidated into a consensus of hundreds of scientists through the “悟道” (“enlightenment-of-the-way”) large model project—has also become a foundation for a generation of entrepreneurs to start their journey through Zhipu’s industrial investments and the entire ecosystem. Today, we want to build this road higher and wider—high enough to protect ourselves and defend national security; high enough to give humankind a chance to explore more unknowns and the mysteries of the universe; and also wide enough for every developer and every team to be able to climb up.

In the AGI era, things that were once遥不可及 (“out of reach”) for decades have, for the first time, a possible path to realization. This is the greatest fortune of our generation of Chinese people, and also the heaviest responsibility.

The giant wave has arrived, and the trend is irreversible. Zhipu wants to be the one who meets the breaking waves and touches high upward.

Not reaching the summit means failure.

This time, what we will touch is that height belonging to all humanity.

Zhipu Founder Tang Jie

July 11, 2026

Original link

Click to learn about the positions at Lydong BlockBeats

Welcome to join Lydong BlockBeats’s official community:

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

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

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