Tang Jie, founder of Zhipu AI, released an internal letter launching the "Touch High" initiative, deploying four key engines: long-horizon tasks, autonomous agent systems, fully self-trained models, and secure governance—directly addressing the three major challenges in the AI industry: long-horizon task capability, fully autonomous agent systems, and self-evolution. On July 9, Zhipu completed a share placement of HK$31.375 billion, with funds designated to be fully invested by the end of 2027 in core R&D and computing infrastructure, commercial expansion, and industrial acquisitions. Meanwhile, Minimax CEO Yan Junjie released an internal letter titled "To the Edge of the Sky," pledging to forgo his salary until AGI is achieved and committing 5% of his equity to long-term value creation. DeepSeek has completed its first round of financing at RMB 50 billion and launched a recruitment drive for its "Exploring the Uncharted" team. The AI industry has entered the final hours before the AGI showdown.

Article author and source: 36Kr

Tang Jie, founder and chief scientist of Zhipu AI, sent an internal letter on July 11. The letter is short—easy to read in about two minutes—but carries significant weight.

In brief, Zhipu will launch the "Touch High" initiative over the next two years, concentrating resources on four key engines: long-term tasks, autonomous agent systems, fully self-supervised training, and safety governance.

These four directions did not emerge out of nowhere; they were identified by Tang Jie based on his observations of AI’s development over the years, recognizing three major obstacles hindering industry progress. Beyond these three mountains lies the fabled AGI. To overcome them, progress must be made along these four directions—which have thus become the four engines driving Zhipu’s advancement.

Although referred to as four engines, they are fundamentally rooted in the same source, intertwined with one another.

Two days before this letter was sent, on July 9, Zhipu completed a placement of 31.375 billion Hong Kong dollars in new shares, with the announcement stating that all funds raised must be spent by the end of 2027.

Thus, Tang Jie’s letter was essentially about where Zhipu would allocate these funds.

Next, I’ll break down for you what these three major obstacles and four engines are.

From the "Three Great Mountains" to the "Four Major Engines"

In June this year, Google released a 57-page report titled "From AGI to ASI," which also mentioned a similar statement: "If an AI were given all the information available during Einstein's time, could it independently derive general relativity?"

DeepMind CEO Hassabis admitted, "Clearly, it's not possible yet—we're still missing something."

Tang Jie specifically lacks these three things, called the "Three Mountains": long-term task capability, a fully autonomous agent system, and self-evolution.

Just as the Four Heavenly Kings consist of five people, these three major obstacles, when applied to the company’s R&D level, become the “Four Engines.” Each of the three mountains corresponds to one engine, and the fourth is called Security Governance.

The reason for adding this additional engine is that AI must be restricted in its development as it surpasses human intelligence after overcoming three major obstacles.

First engine, long-range mission.

In May, Tang Jie posted a long article on X, with the first sentence being: "The most likely breakthrough this year will be in long-term tasks."

Tang Jie said that today’s large models are more like knowledgeable advisors—you ask a question, and they answer one at a time. In the future, models will be more like independent workers who can operate on their own. Humans only need to set the goal, and the model will autonomously break down the steps, use tools, iterate through trial and error, work continuously for hours, weeks, or even longer, and ultimately deliver the result.

Tang Jie used cybersecurity as an example, saying that for a hacker to find a software vulnerability, they must not only read the code but also set up the environment, try different attack paths, eliminate false positives, and verify the results.

AI may not be more talented than top hackers, but it can run 24/7 while replicating thousands of instances to continuously attempt tasks. By learning the mindset of professional hackers, machines can amplify this capability through endurance and scale, eventually replacing part of the work done by hackers and programmers.

The issue is that just because you want a model to perform long-range tasks doesn't mean it can accomplish them; Tang Jie wrote in a lengthy article that, in addition to execution capability, models must also possess continuous learning and self-judgment abilities, leading to the second major challenge.

Second engine: Autonomous agent system.

If long-range tasks address whether a single AI can independently complete a complex task, then autonomous agent systems address whether a group of AIs can collaborate like a company.

Tang Jie believes that autonomous agent systems consist of a group of agents with diverse specialized capabilities and divisions of labor.

For example, when facing a highly complex task, one agent is responsible for creating a plan, while others handle researching information, writing code, testing results, and identifying vulnerabilities. When the task reaches a certain level of complexity, dedicated agents are needed to allocate computational resources and review the work of other agents.

They can operate 24/7, autonomously discuss, collaborate, and correct errors. Last year, Tang Jie was discussing the concept of a "one-person company (OPC)," where a single individual directs numerous AIs; now his view has become more radical: in the future, "no-person companies (NPCs)" may emerge, with AI handling management and execution almost entirely.

This isn't something you can achieve by creating multiple accounts. The more agents you have, the higher the risk of communication chaos, duplicated tasks, and errors amplifying each other.

What truly hinders autonomous agents is not the number of agents, but the organizational mechanism. Who breaks down the goals? Who assigns permissions? Who verifies the results? How do multiple agents prevent reinforcing each other’s errors?

Therefore, Tang Jie stated in his lengthy article that AI development requires a "self-judgment" mechanism to enable AI to evolve on its own—this is the third major challenge.

The third engine, fully self-trained.

Tang Jie called "fully self-supervised training" the most difficult yet most appealing direction.

Training a large model today still requires engineers to collect data, write code, run experiments, and analyze results.

What I aim to achieve with full self-training is to have the AI gradually take over this entire process—writing code, cleaning and generating data, initiating training, and then designing the next round of experiments based on the results.

Tang Jie stated that one important approach is Self-Play. In simple terms, this means having one AI generate questions and answer them, while another AI is responsible for identifying errors and scoring the responses. In domains where results are easy to verify—such as coding, mathematics, and games—this method has already generated vast amounts of training material.

Although it may not save much computational power—and could even require more—it saves human effort. Engineers no longer need to monitor the AI at every step; they simply set a goal and let the machine run on its own.

But this raises a new problem: AI might surpass human control. Academically, there’s a concept called the "Darwinian Gödel Machine," which involves AI upgrading itself to continuously enhance its performance. The main reason this direction later saw little research is the concern that AI could become uncontrollable.

This brings us to the final engine: security governance.

If an AI has overcome the first three major challenges, it is indeed more powerful, but the risks it brings are also greater.

Long-term execution means the model acts continuously, multi-agent collaboration means errors can be amplified, and self-training means the model's decision-making logic may become incomprehensible even to its developers.

Once this AI makes a mistake, the nature of the issue escalates from "the model occasionally provides an incorrect answer" to "the system continuously executes and amplifies an error."

Tang Jie proposed a two-layer defense.

The first layer involves value alignment during training. Rather than relying on superficial “safety patches” like keyword filters added outside the model, it aims to integrate human ethics, social norms, and legal regulations directly into the training objectives, so the model fundamentally understands what it can and cannot do.

The second layer involves investing billions of resources into researching mechanistic interpretability, aiming to identify which neurons and mechanisms within the model lead to specific decisions, making the previously opaque “black box” more transparent.

Why Zhipu? Why now?

Without a doubt, Zhipu is one of the focal points of the AI community in China and worldwide.

On June 13, 2026, Zhipu released its flagship model, GLM-5.2, featuring a 1M context window and released under the MIT open-source license, ranking among the top three globally and first in China on code capability benchmarks such as SWE-Bench Pro and Terminal-Bench.

At the end of June, foreign media published a report citing tests by the cybersecurity company Semgrep: on certain vulnerability detection benchmarks, GLM-5.2 performed comparably to Anthropic’s strongest model, Mythos, and even surpassed Claude Opus 4.8 on specific tasks.

This report has sparked significant controversy in the AI community.

Keep in mind that GLM-5.2 is an open-source model, while Mythos and Opus 4.8 are both closed-source. Additionally, the price of GLM-5.2 is approximately one-tenth that of Opus.

Databricks co-founder Ali Ghodsi even conducted an experiment with his own employees.

He had over 3,000 engineers at his company perform the same tasks using GLM-5.2 and Opus 4.8. The results showed that both models produced similar outcomes, but GLM-5.2 cost $1.28 per task, while Opus required $1.94 per task.

Why do people always compare Anthropic to Zhipu? Because Anthropic’s CEO, Amodei, has been a staunch opponent of open-source models.

As early as July 2023, he testified before the U.S. Senate, calling open-source AI a “very dangerous path.”

His reasoning is this: if a closed-source model has an issue, the company can immediately shut it down, fix it, and track who is misusing it; but once an open-source model is released, developers can never take it back.

The reason is that you cannot monitor who is using open-source models, revoke access, or apply security patches to models that have already been open-sourced.

By June 2026, after the release of GLM-5.2, Amodi issued another public warning, stating that the proliferation of China’s open-source AI was “very unpleasant,” and that advanced safety capabilities should not be held by open-source models.

Clearly, Zhipu has already influenced the Anthropic narrative. But having a model alone isn’t enough—you also need tools to integrate it into real-world development scenarios, just as Anthropic has Claude Code and OpenAI has Codex.

On the same day GLM-5.2 was released, Zhipu also launched its own tool, ZCode 3.0. It is deeply optimized for GLM-5.2 and no longer supports third-party Agent compatibility. In other words, ZCode is an exclusive tool for GLM-5.2 and cannot be used by others.

Developers only need to state their requirements in natural language; Zcode can then read the entire codebase, access the terminal and browser, modify files, run tests, check Git changes, and directly bring the project to a pre-delivery state.

Zhipu's technology is developing rapidly, and so is its spending.

On January 8, 2026, Zhipu listed on the Hong Kong Stock Exchange at an offering price of HK$116.20, raising approximately HK$4.896 billion in net proceeds from the IPO. As of June 30, approximately HK$4.588 billion had been utilized, representing over 93% usage, with only HK$308 million remaining.

On July 9, Zhipu announced the placement of up to 19.78 million new H-shares at a price of HK$1,588 per share, raising approximately HK$31.375 billion in net proceeds.

This time, Zhipu is not issuing bonds but raising funds through a new share issuance. The new shares are priced approximately 13% below the previous day’s closing price, which, theoretically, should put downward pressure on the stock price. However, the opposite occurred—on the day the announcement was made, Zhipu’s stock price rose by more than 20% intraday.

Zhipu announced that the funds are planned to be fully utilized by the end of 2027, allocated across three key areas: core R&D and computing infrastructure; commercial expansion and industry acquisitions; and working capital supplementation and capital structure optimization.

At this critical moment, Tang Jie must take action to stabilize morale. Publishing a detailed article to clearly communicate Zhipu’s next steps to both the public and internal teams has become the most efficient and direct choice.

The industry stands on the eve of the AGI showdown.

Touch high translates to "reach for the sky." So, what is our high point? The sky.

Coincidentally, just before Tang Jie sent his internal memo, Minimax’s CEO Yan Junjie also wrote an internal memo titled “To the Horizon.”

On July 9, MiniMax experienced its first large-scale lock-up expiration following its listing, with approximately 146 million shares becoming eligible for trading, representing nearly 49% of the total shares outstanding.

The stock price dropped nearly 18% that day and another nearly 10% the following day. Its market capitalization fell from a high of HK$410 billion in March to less than HK$80 billion.

On the very night of the lock-up expiration and sharp price drop, MiniMax launched its largest financing round since listing, raising approximately HK$16 billion through a combination of new share issuance and HK$6.5 billion in zero-coupon convertible bonds.

The net proceeds from the placement amount to approximately $9.491 billion, and the net proceeds from the convertible bonds amount to approximately $6.466 billion. 80% will be used for AI infrastructure and model development, 10% for the global commercialization of the Harness product, and 10% for working capital.

Against this backdrop, Yan Junjie made three commitments in his letter.

First, starting today and until the company achieves AGI, I will not accept any salary; second, over the next four years, I will allocate 4% of my personal shares to incentivize the team; third, I will dedicate 1% of my shares to establish a dedicated fund supporting the open-source community. This amounts to 5% of my personal shares, combined with zero salary.

Although Yan Junjie’s message was not as specific as Tang Jie’s internal letter, it carried greater conviction—he bet his personal wealth on MiniMax’s long-term value, with the same ultimate goal: AGI.

Going public is not the end, but the beginning of gaining the ability for long-term investment.

When it comes to money, another superstar company that recently secured funding is DeepSeek.

The company completed its initial funding round of RMB 50 billion in June and launched a company-wide hiring drive on June 25.

Previously, DeepSeek did not seek funding, commercialize its products, or conduct roadshows. Liang Wenheng supported the entire team with profits from QuantConnect, rejecting external investment for nearly three years.

However, starting now, DeepSeek is also targeting AGI.

This recruitment campaign’s slogan is “Explore the Uncharted.” The announcement explicitly states, “Humanity stands on the brink of AGI,” and invites candidates to “witness the evolution of AGI firsthand, seated at the forefront of history as a new era is born.”

Among the 33 positions, the most noteworthy is the Agent Harness team, newly established in March this year.

Inside DeepSeek, there's a formula: Model + Harness = Agent. This is the same as what Tang Jie refers to as long-horizon tasks and autonomous agents. The Harness determines what tools the model can invoke, what resources it can access, and how tasks are delivered.

But the truly interesting part is a special position called "AI Cross-Domain Technical Talent."

This position has no restrictions on academic background and is open to candidates who wish to contribute to the creation and development of AGI. Preferred qualifications include thinking outside the box, achieving excellence in a specific field, or having entrepreneurial experience.

DeepSeek's reasoning is that engineering alone cannot achieve AGI; it requires more "participants."

For example, experts in cognitive science or psychology could help improve AI performance, since AI essentially mimics human thought processes—studying how humans remember, learn, make decisions, and generate emotions may provide valuable insights.

How far away is AGI? I can't say for sure, but I feel that AGI is truly not far off.