Exclusive: Zhipu AI Founder Tang Jie Internal Letter – What Matters More After the "GLM Moment"

Essence, counterintuitive, focus

Text by Xinyu Zhou

Edited by Yuxin Zhang

Exclusive to *Smart Emergence*: On July 11, 2026, Jie Tang, founder of Zhipu AI, released an internal letter titled *The Great Wave Has Arrived* within the company.

Over the past six months, Zhipu AI has experienced the most brilliant moments since its founding: its market value has surged 10 times compared to the early post-listing period six months ago, and in June 2026, it joined the "HK$1 Trillion Club" — a figure nearly three times the market value of Baidu and exceeding that of Xiaomi. After the first batch of stock unlocks on July 8, Zhipu AI still maintained stable share prices.

This is also the most compelling story in the current AI large model track: betting on the right technical direction has brought remarkable market reputation and commercial achievements.

Peeling back the layers, the root of Zhipu AI's rapid rise lies in its bet on Coding a year ago. In early 2025, Zhipu AI began reallocating resources to focus its efforts on enhancing the model's coding capabilities.

At an event, Jie Tang, Professor in the Department of Computer Science at Tsinghua University and founder of Zhipu AI, explained the rationale for betting on coding: the emergence of DeepSeek R1 marked that the exploration of the Chat paradigm had largely come to an end. For the model training paradigm in the post-DeepSeek era, he described Zhipu AI as having "bet" on Coding and Reasoning — a model capability that can co-evolve and thrive alongside Agents.

Facts have proven that this was a successful wager.

Today, AI Coding has become the closest entry point for AI commercialization. A typical case is Anthropic, which also overtook OpenAI on a curve by betting on coding. In January 2024, Anthropic's annualized revenue was only $87 million; by June 2026, its ARR had exceeded $47 billion.

From the flagship model GLM-4.5 released in July 2025 to GLM-5.2, which launched and went open-source in June 2026, Zhipu AI has secured a spot in the global first tier of AI coding. The open-source GLM-5.2 has matched or even surpassed Claude Opus 4.8 and GPT-5.5 across multiple core metrics.

Technical capabilities are also reflected in revenue. Zhipu AI's 2025 financial report shows that by the end of the reporting period (March 2026), the ARR of its MaaS platform reached 1.7 billion RMB, a 60-fold increase over the past year.

Tang Jie's internal letter reveals the new proposition Zhipu AI is betting on in the post-coding era:

Long Horizon Task Capability
Autonomous Agent System
Self-Evolving

The full text of Tang Jie's internal letter is as follows:

The Great Wave Has Arrived

To every member of Zhipu AI and every partner who cares about the future of artificial intelligence

Allow me to use this letter to discuss three things: who we are, how we view this era, and the strategic direction we have decided to fully commit to.

(1) Who We Are: "Essence, Counter-Intuition, Focus"

Zhipu AI has never been a company chasing fleeting trends. It grew out of a laboratory, carrying the methodology this lab has cultivated over two decades. This methodology can be summed up in three words: Essence, Counter-Intuition, Focus — think deeply enough, and you dare to choose the counterintuitive path; choose that counterintuitive path, and you must stay committed long enough.

Looking back at our journey, almost every key choice we made once seemed "counter-intuitive". In 2006, we stuck to an academic search system on a single desktop computer, enduring the quiet, unglamorous early days, because we realized that behind it lay the work of "uncovering the evolutionary mechanisms of academic disciplines" — a topic worth a decade of exploration. From 2021 to 2022, when "making machines think like humans" was regarded by most as a moon-shot-level crazy plan, we reallocated resources to bet on hundred-billion-parameter models, and built GLM-130B — a full year and a half before ChatGPT took the world by storm. On the day Zhipu AI listed on the Hong Kong Stock Exchange on January 8, 2026, we treated it as a brand new starting point, resolutely returning fully to foundational model research to push forward the next generation of models.

While others celebrated their listing with bell-ringing ceremonies, we reset to zero. This is not a gesture — it is a belief: since our final destination is AGI, short-term gains or industry trends are just scenery along the path to the end goal.

What has supported us all the way to today is extreme focus and pure, unwavering idealism. We spent a decade growing our academic search system from a single desktop to serving millions of users; we have worked on the large model path for nearly another decade, and will continue to deepen our efforts unswervingly. Today's Zhipu AI is a group of people who are willing to pursue the essence, dare to think counter-intuitively, and focus on seeing things through to completion — this is the source of Zhipu AI's core competitiveness.

(2) How We View This Era: The Upper Bound of Intelligence Is Being Rewritten

If there is one thing we have learned over the past two decades, it is this: real commercial opportunities never lie in minor tweaks to products or business models, but in leaps in the upper bound of intelligence. This is our most fundamental judgment of the current AI revolution, and the insight we most want to convey to everyone.

The essence of this revolution is not a product innovation or business model innovation, but a technical revolution that itself raises the "upper bound of intelligence". Whoever takes the lead in pushing this upper bound even an inch upward will redefine the capability boundaries of thousands of industries. All new-generation AI enterprises focused on first principles are competing for this inch of breakthrough.

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 no longer just "see" and "hear", but begin to "understand" and "reason". The next step points directly to AGI.

We have a simple yet strict definition of AGI: AGI is not the wisdom of a single genius, but the sum of all human wisdom. It should have the ability to create original knowledge at the level of the *Theory of Relativity* — this is our only criterion for measuring whether we have truly reached the peak. On the path to this destination, there are several peaks that must be crossed, and they are exactly where the technological wave is surging most fiercely today:

The first peak: Long Horizon Task Capability

The most exciting breakthrough today is enabling models to learn to complete extremely long tasks — not instant Q&A, but planning and execution spanning weeks, months, or even years. For example, a model that can tirelessly find vulnerabilities in software is essentially learning the way of thinking of a top security expert, then amplifying that capability through machine endurance.

The second peak: Autonomous Agent System

Building on long-horizon tasks, groups of agents that can self-initiate, collaborate, and operate 24/7 will become a new form of productivity. We once proposed the "One-Person Company OPC", but technology is advancing faster than expected — we are moving toward a "Fully Automated Company NPC". The three once hard-to-solve problems of Memory, Continual Learning, and Self-Judge — once thought to require paradigm shifts — are now gradually being resolved by the dual drivers of technology and applications: long context and Retrieval-Augmented Generation (RAG) are approaching the embryonic form of memory; increasing model iteration frequency itself is approaching continual learning; cutting-edge models have already shown the first signs of self-judgment.

The third peak: Self-Evolving

This is the most difficult, yet most alluring peak. AI training AI has already taken shape — models write their own code, clean and synthesize their own data, and train themselves. This may consume some computing power, but saves the most precious resources: human labor and time. In the large model era, speed is paramount — rapid iteration will directly widen the generational gap in cognitive capabilities. When leading overseas enterprises start building computing clusters at the scale of millions or even two million chips, their real use is very likely to be letting models train themselves.

What will happen after we cross these three peaks?

AI will begin to learn what "I" means, what self-awareness is; beyond that, it will touch on human emotions; further ahead, it will reach consciousness itself. From perception to cognition, from cognition to general intelligence, from general intelligence to Artificial Super Intelligence (ASI) — this path has already unfolded, the great wave has arrived, and it is irreversible.

This is not just our view. Google DeepMind put forward a stark assertion in its *From AGI to ASI* report: even if the capability of a single model stays at human level forever, as long as computing power keeps growing, superintelligence could be effectively "squeezed out". They deduce that if the global number of runnable AGI instances grows tenfold every year, there will be 100 million of them five years later. These agents, which share the same underlying brain, have a hundredfold improvement in thinking efficiency, and can replicate experience at zero cost, are equivalent to ASI at the collective level. In other words, moving from AGI to ASI requires both algorithmic breakthroughs and the convergence of massive computing resources.

This irreversible trend will penetrate the entire tech stack from top to bottom: when AGI arrives, today's applications may all need to be reconstructed as AI-native, or even these applications will no longer be needed; operating systems may be rewritten, and in the future when you turn on your computer, you will see an "LLM OS" where all functions are generated on demand; deeper still, it will challenge the 80-year-old von Neumann architecture. Finance, law, e-commerce, the internet... no industry will stay unaffected. Many friends have come to me saying they want to transform their enterprises and catch up with the pace of AI, but very few have truly perceived that "this irreversible revolution has already begun".

(3) The Direction We Fully Commit To: "Touch High"

After recognizing the trend, what remains is choice. Zhipu AI's choice, as always, is "counter-intuitive" — at a time when the industry is generally accelerating commercial monetization, we have decided to break through upward.

We have named this strategy the "Touch High Plan". At this historical juncture where artificial intelligence is leaping from perception and cognition to full general intelligence, Zhipu AI will adopt a "touch high" posture to challenge the physical and algorithmic limits of current technology. Over the next two years, we plan to make strategic investments — not pursuing 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. Enabling AI to move from "instant Q&A" to "grand engineering", developing a new generation of memory architectures that allow models to "learn, do, and remember" throughout the full lifecycle of a project, with the top-level capability to autonomously break down grand goals (such as "design a new anti-cancer drug molecule") into thousands of executable subtasks.

Second, autonomous agent systems. Moving from "smart assistants" to "digital employees", building a society of agents with thousands of different professional "personalities" and "skills", allowing them to debate autonomously, collaborate, audit code, and schedule resources, to achieve "autonomous driving"-level digital productivity.

Third, Fully Self Training. As high-quality human data is about to be depleted, turning computing power into fuel for evolution — building high-quality synthetic data factories, generating knowledge "from nothing" through AI-AI Self-Play, and granting the system the ability to reconstruct its own code within a secure sandbox, so that the speed of evolution can break free from the physical limits of human engineers.

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

The more powerful capabilities become, the more robust safety constraint mechanisms must be. Since its founding, Zhipu AI has established a principle: AI must serve human well-being and national strategies. The company abandons external security patches, insisting on writing human ethics, social norms, and national laws and regulations into the model's value function as underlying axioms; plans to invest tens of billions of resources to tackle "mechanical interpretability", clarify the neuronal logic behind model decisions, and promote the transformation of black-box systems to transparent and interpretable systems; at the same time, actively participates in international AI governance to prevent the misuse of AI technology.

This sense of urgency is not unfounded. When the world's most cutting-edge top models delay full public release due to risk considerations, and the heads of those companies publicly warn that the far-reaching impact of AI will profoundly reshape the global power landscape, we should be more sober: the realization of superintelligence and research on super alignment must advance in parallel. This is also the proposition we repeatedly examine when facing disruptive technology — 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, but the fundamental prerequisite for the technology to survive and be allowed to be applied.

(4) Open Ecosystem: The Underlying Logic of Intelligent Inclusion and Safety Governance

We always believe that as a strategic technology leading the future, the long-term development of artificial intelligence cannot be separated from an open and collaborative industrial ecosystem. The value of cutting-edge intelligence lies not only in technical breakthroughs themselves, but in whether it can widely empower thousands of industries and benefit every developer. We firmly believe that true safety is not built on technical closure and barriers, but comes from extensive co-construction, sharing, and supervision under the sun.

Based on this deep recognition of technological inclusion, Zhipu AI has presented its own strategic answer. Recently, we released GLM-5.2, our most capable open-source model to date. It supports a truly usable 1M context window, continues to maintain leadership in long-horizon tasks, is open to all users, and will be officially open-sourced under the most permissive MIT license — anyone can download, deploy, and use it commercially, with no restrictions on user types. This is the firm stance the company expresses through its products.

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

This is not contradictory to "Touch High" — on the contrary, they are two sides of the same coin: with one hand, we reach upward to challenge the limits of intelligence; with the other hand, we pave the way downward, making the most cutting-edge capabilities as open and inclusive as possible. The heights we touch belong to all humanity, and the roads we build belong to every individual.

(5) Conclusion: Why Now, Why Us

Some may ask: Why, after Zhipu AI's listing, do we still focus our core resources and "reach high" in the most uncertain direction? Because we believe in a simple truth: those who truly reach the summit will turn the mountain into a road.

The essence we have clarified once formed the consensus of hundreds of scientists through the "Wudao Large Model" project, then through Zhipu AI's industrial investments and the entire ecosystem, became the stepping stone for a generation of entrepreneurs. Today, we want to build this road higher and wider — high enough to protect ourselves, safeguard national security, and allow humanity the opportunity to explore more unknowns and the mysteries of the universe; wide enough for every developer and every team to walk on it.

In the AGI era, these once distant things have for the first time become possible to achieve. This is the greatest fortune, and the heaviest responsibility, of our generation in China.

The great wave has arrived, and the trend is irreversible. Zhipu AI will be the one who faces the wave and reaches upward.

To not reach the summit is to fail.

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

Jie Tang, Founder of Zhipu AI

July 11, 2026