The Chinese company that is most like Anthropic wants to move away from the “three major mountains”

Zhipu Zhizhang founder and chief scientist Tang Jie sent an internal letter on July 11. The letter isn’t long—you can finish it in about two minutes—but its weight is not light.

To put it simply, over the next two years, Zhipu will launch a “Touch High (reach for higher)” plan, concentrating resources into four engines: long-range tasks, autonomous agent systems, fully self-training, and safety governance.

These four directions don’t come out of thin air. Tang Jie looked at AI’s progress over the years and identified three mountains that are obstructing the industry’s advance. After you climb over those mountains is the legendary AGI. But to cross those three mountains, you have to move in these four directions. In this way, the four directions become four engines that drive Zhipu forward.

Although they’re called four engines, in essence they share the same roots—you have me in you and I have you in me.

And just two days before this letter was sent, on July 9, Zhipu had just allotted 31.38B HKD worth of new shares to raise funds. The announcement stated that all the money raised will be spent by the end of 2027.

So in effect, Tang Jie’s letter is really about explaining where Zhipu plans to spend this money.

Next, I’ll analyze for you what these three mountains and four engines actually are.

From “Three Mountains” to “Four Great Engines”

Google released a 57-page report in June this year titled “From AGI to ASI,” which also mentions similar ideas. “If you give an AI all the information from the era of an AI Einstein, can it independently derive the general theory of relativity?”

DeepMind CEO Hassabis acknowledged, “Obviously, not today—we’re missing something.”

Tang Jie named the missing “something” as “three mountains.” They are long-range task capability, fully autonomous agent systems, and self-evolution.

Like the Four Heavenly Kings having five people, these three mountains translate into four engines at the company’s R&D level: three mountains correspond to one engine each, and the fourth engine is called safety governance.

The reason an extra engine is added is that when AI surpasses human-level intelligence after crossing those three mountains, it must be constrained in its development.

First engine: long-range tasks.

In May this year, Tang Jie posted a long piece on X. The first sentence was: “The direction most likely to break through this year is long-range tasks.”

Tang Jie said that today’s large models are more like knowledgeable consultants—you ask a question, it answers. But future models will be more like employees who can work independently. Humans only need to set goals; the model will break down steps by itself, call tools, run repeated trials and errors, work continuously for hours, weeks, or even longer, and ultimately deliver results.

Tang Jie used cybersecurity as an example. He said if you ask a hacker to find a software vulnerability, it isn’t just about reading code—it also has to set up the environment, try different attack paths, rule out false positives, and verify the results.

AI may not be more talented than top-tier hackers, but it can run 24 hours a day, and it can replicate thousands upon thousands of instances to keep trying. As long as it learns the professional hacker’s way of thinking, the machine’s stamina and scale could amplify this capability and ultimately replace some work done by hackers and programmers.

The problem is: you can’t simply tell the model to do long-range tasks and expect it to complete them. In his long post, Tang Jie wrote that besides execution capability, the model also needs continuous learning and self-judgment ability—this leads to the second mountain.

Second engine: autonomous agent systems.

If long-range tasks answer whether “can a single AI independently finish a complex job,” then autonomous agent systems answer “can a group of AIs coordinate and work together like a company.”

Tang Jie believes autonomous agent systems are composed of multiple agents with different professional capabilities and divisions of labor.

For example, facing a very complex task, you need one agent to make the plan, another to look up information, write code, test results, find vulnerabilities. When the task grows to a certain scale, you need dedicated agents to allocate compute power and check the work of other intelligent agents.

They can run 24/7, autonomously discuss, collaborate, and correct errors. Last year Tang Jie was also talking about “one-person companies (OPC),” meaning one person directing a large number of AIs. Now his judgment is even more aggressive: in the future, there may be “no-person companies (NPC),” where AI handles most of everything from management to execution.

This doesn’t mean you can achieve it by just running multiple accounts. The more agents you have, the higher the risks of communication chaos, task duplication, and error feedback loops multiplying.

What truly holds back autonomous agent systems is not the number of agents, but organizational mechanisms. Who decomposes the goals? Who assigns permissions? Who checks the results? How do you prevent different agents from reinforcing each other’s mistakes?

Therefore, in his long post, Tang Jie said that AI development needs a mechanism of “self-judgment,” so that AI can self-evolve. This is the third mountain.

Third engine: fully self-training.

Tang Jie calls fully self-training the most difficult—and most enticing—direction.

When training a large model today, you still need engineers to prepare everything: collect data, write code, run experiments, and analyze results.

Fully self-training aims to have AI gradually take over this whole process—write code itself, clean and generate data, launch training, and then design the next round of experiments based on the results.

Tang Jie said one important method is Self-Play. In simple terms, the AI both creates questions and answers them, and another AI is responsible for picking out mistakes and scoring. In domains where results can be easily validated—such as code, math, and games—this approach can already generate large quantities of training materials.

While it may not save much compute power, and could even consume more compute, it saves human effort: engineers don’t need to watch every stage of the AI; they only need to set a goal, then let the machine run on its own.

But this easily creates a new problem: AI might exceed human control. There is an academic concept called the “Darwin-Gödel machine,” in which AI upgrades itself so that model performance keeps improving. This direction later attracted little research mainly because people feared the AI would be out of control.

That brings us to the final engine: safety governance.

If an AI crosses the previous three mountains, it indeed becomes more powerful—but the risks it brings are also greater.

Long-range execution means the model keeps taking actions, and multi-agent collaboration means errors will be amplified. Self-training means the model’s decision logic may be incomprehensible even to developers.

Once this AI makes a mistake, the nature of the issue upgrades from “the model occasionally gives a wrong answer” to “the system keeps executing and amplifying a mistake.”

Tang Jie proposed two layers of protection.

The first layer is value alignment during training. It’s not satisfied with adding “safety patches” like keyword filtering on top of the model. Instead, it aims to incorporate human ethics, social norms, and laws and regulations into the training objectives, so that the model understands from the bottom layer what can and cannot be done.

The second layer is investing resources at the “tens of billions” scale to research mechanical interpretability, trying to figure out which neurons and mechanisms inside the model lead to a certain judgment—making the hard-to-understand “black box” more transparent.

Why Zhipu, and why now

There’s no question that Zhipu is one of the focal points in AI circles across all of China, and even globally.

On June 13, 2026, Zhipu released its flagship model GLM-5.2. With a 1M context window, and an MIT open-source license, it ranked in the global top three and number one among domestic models on code capability benchmarks such as SWE-Bench Pro and Terminal-Bench.

By late June, overseas media ran a report citing tests from network security company Semgrep. In certain vulnerability detection benchmarks, GLM-5.2’s performance was on par with Anthropic’s strongest model, Mythos, and in specific tasks even exceeded Claude Opus 4.8.

This report triggered massive controversy in AI circles.

You have to know GLM-5.2 is an open-source model, while Mythos and Opus 4.8 are closed-source models. Moreover, GLM-5.2’s price is about one-tenth of Opus.

Databricks co-founder Ali Ghodsi even conducted an experiment using his own employees for this.

He had his company’s 3,000+ engineers do the same tasks using GLM-5.2 and Opus 4.8. The result found the outputs are similar, but GLM-5.2 costs $1.28 per task, while Opus costs $1.94.

Why does everyone like comparing Zhipu with Anthropic? Because Anthropic’s CEO Amodei has long been a staunch opponent of open-source models.

As early as July 2023, he went to testify before the U.S. Senate, saying open-source AI is a “very dangerous path.”

His logic was like this: if a closed-source model has problems, the company can immediately shut it down, fix it, and track who is abusing it. But once an open-source model is released, you can’t take it back from developers.

The reason is that you can’t monitor who is using the open-source model, you can’t revoke access, and you can’t add safety patches to a model that has already been open-sourced.

By June 2026, after GLM-5.2 was released, Amodei publicly warned again, saying the spread (release) of open-source AI in China is “something I don’t like very much,” and that advanced safety capabilities shouldn’t be in the hands of open-source models.

Clearly, Zhipu has already affected Anthropic’s narrative. But having only a model isn’t enough—you also need tools to plug it into real development scenarios. Like Anthropic has Claude Code, and OpenAI has Codex.

On the very day GLM-5.2 was released, Zhipu also rolled out its tool, ZCode 3.0. It deeply adapts to GLM-5.2 and no longer maintains third-party agent adaptation. That means ZCode is a tool exclusive to GLM-5.2—others can’t use it.

Developers just need to describe requirements in natural language. Zcode can read the entire codebase, call the terminal and browser, modify files, run tests, check Git changes, and then directly move the project forward to the pre-delivery state.

Zhipu’s pace of technical development is fast, and its burn rate is also fast.

On January 8, 2026, Zhipu listed on the Hong Kong Stock Exchange, issuing at HKD 116.2 per share, raising net IPO proceeds of about HKD 1M. By June 30, it had used about HKD 4.9B, with a utilization rate of over 93%, leaving only HKD 308 million.

On July 9, Zhipu announced a placing at HKD 1,588 per share for up to 19.78 million new H shares, raising net proceeds of about HKD 4.59B.

Zhipu did not issue bonds this time; it raised funds by issuing additional new shares. The subscription price for the new shares was about 13% cheaper than the previous day’s closing price, so in theory this kind of action should put pressure on the stock price. But the result was the opposite: on the day the news was announced, Zhipu’s share price surged by over 20% at one point during intraday trading.

In its announcement, Zhipu said it plans to use all this money by the end of 2027. It will be invested across three major areas: core R&D and compute infrastructure; commercial expansion and industrial M&A; supplementing operating funds and optimizing capital structure.

So right at this moment, Tang Jie has to do something to steady confidence. Posting a long post so that both the outside world and the company internally clearly understand what Zhipu plans to do next is the most efficient and direct choice.

The industry enters the night before the AGI showdown

Touch high literally means “reach for higher.” So what is our “higher”? It’s the sky.

And what a coincidence—just before Tang Jie sent his internal letter, MiniMax CEO Yan Junjie also wrote an internal letter titled “Toward the End of the Sky.”

On July 9, MiniMax saw the first large-scale release of restricted shares after listing. About 146 million shares were released, accounting for nearly 49% of total shares outstanding.

That day, the stock price plunged by nearly 18%, and the next day it fell by another nearly 10%. Market cap dropped from a March high of HKD 410 billion all the way down to less than HKD 80 billion.

That same night after the unlock-and-plunge, MiniMax launched its largest refinancing since listing. It placed new shares plus HKD 6.5 billion zero-coupon convertible bonds, with total fundraise of about HKD 16 billion.

Among them, the net placing amount was about HKD 31.38B, and the net convertible bond amount about HKD 9.49B. 80% would be used for AI infrastructure and model R&D, 10% for global commercialization of the Harness product, and 10% for operating funds.

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

First, from immediately to when the company achieves AGI, he will not take any salary. Second, in the next four years, he will set aside equity incentives equal to 4% of the total share capital from his personal holdings for the team. Third, he will set up a dedicated fund supported by 1% of shares to back open-source communities. With 5% of his personal shares plus zero salary.

Although Yan Junjie’s letter wasn’t as specific as Tang Jie’s internal letter, it was even more forceful: he bet his personal fortune on MiniMax’s long-term value, with the endpoint also being AGI.

Listing isn’t the finish line—it’s the start of having the ability to make long-term investments.

Speaking of money, there’s another super-star company that recently got funding: it’s DeepSeek.

In June, this company completed its first round of financing totaling RMB 50 billion, and on June 25 it began expanding headcount for the whole team.

Previously, DeepSeek had no funding, no commercialization, and no roadshows. Liang Wenfeng funded the entire team with profits from a market-neutral quant strategy; the company was founded nearly three years ago and refused external investment.

But from now on, DeepSeek is also aiming at AGI.

This time its recruitment slogan is “Exploring the Untrodden.” The announcement directly states, “Humanity is on the eve of AGI,” inviting applicants to “witness the development process of AGI firsthand, sit at the front of the era, and witness the birth of a new epoch.”

Among the 33 positions, the one most worth paying attention to is the Agent Harness team newly established in March this year.

DeepSeek has an internal formula: Model + Harness = Agent. This is the same as what Tang Jie said about long-range tasks and autonomous agents. Harness determines which tools the model can call, which resources it can access, and how it delivers tasks.

But what’s truly interesting is a special role called “AI cross-disciplinary technical talent.”

This role has no restrictions on professional background. It targets “candidates who hope to participate in creating and building AGI.” The bonus points include “not taking the usual path,” “achieving excellence in a certain field,” and “having entrepreneurial experience.”

DeepSeek’s logic is that engineering alone is not enough to reach AGI—it needs more “participants.”

For example, if you have talent in cognitive science or psychology, since AI essentially imitates human thinking processes, by studying how humans remember, learn, judge, and generate emotion, it might help AI improve performance.

How far away is AGI? I can’t say for sure, but I feel like AGI really isn’t far.