RSI Rewrites the Large Language Model Competition: Domestic Vendors Sprint for IPOs, Silicon Valley Giants Hit the Brakes

Chinese large model vendors are pouring into the capital market at a speed visible to the naked eye, raising "ammunition" for the next round of competition. Right now, four leading companies are queuing up to hit the listing bell.

DeepSeek has taken the most concrete actions. On September 23, according to The Information, Liang Wenfeng, founder of DeepSeek, disclosed at a recent meeting with investors that DeepSeek's current annualized revenue has reached 1 billion US dollars, and the company is advancing its second round of financing.

Moreover, recent data from Qichacha shows that Yan Wentao has been added to the list of core personnel of DeepSeek, serving as CFO. This post-90s former partner of GL Ventures previously participated in investments in MiniMax and Zhipu, and is very familiar with the operation process of large model enterprises sprinting to the capital market.

Zhipu and MiniMax, which were listed on the Hong Kong Stock Exchange successively in January this year, have respectively signed counseling agreements for the Sci-Tech Innovation Board, planning secondary listings on the A-share market to sprint for the "A+H" dual platform. The IPO process of Moonshot AI is also significantly accelerating. After completing its Series F financing of over 3.5 billion US dollars in June, its valuation rose to 35 billion US dollars. Market sources said that it has secretly submitted a listing application to the Hong Kong Stock Exchange.

At the same time, the pace of Silicon Valley model vendors in the listing process has been slowing down again and again. According to foreign media reports, Anthropic has postponed its IPO to November, which is later than the October that many investors previously expected. OpenAI is considering postponing its IPO to 2027, and CEO Sam Altman publicly stated that under the current situation, 2026 is not the right time for listing.

At the current stage of the same AI competition, Silicon Valley has stepped on the brakes while China is stepping on the accelerator. Behind the two different paces, there may be more than differences in judgments on industrial risks. It is more noteworthy that both sides are encountering a new technical variable at the moment: Recursive Self-Improvement (RSI) of AI, but their attitudes are completely different.

What are the people who most want to step on the brakes anxious about?

At present, almost all top AI laboratories are frantically sprinting for RSI.

OpenAI has set up an RSI team. Anthropic disclosed the internal contribution of RSI in its article "When AI Builds Itself". Google DeepMind lists RSI as one of the only paths to superintelligence, and Meta explores automated scientific research through AIRA2.

Domestic vendors are also accelerating the advancement. Previously, when DeepSeek Harness was released, Liang Wenfeng talked in an internal speech that the self-iteration of AI will lead everyone to a "singularity". At the Yunqi Conference on September 22, Wu Yongming publicly shared the phased progress made by Alibaba's Qwen team on RSI. According to the introduction, Qwen3.8-Max has continuously iterated for more than 1 month with absolutely "zero human participation" and completed 33 rounds of effective iteration.

Looking further back, Xiaomi regards its latest model MiMo-V2.6 as a key step to explore RSI. Zhipu announced that it will incorporate the "Fully Self Training" system into the roadmap of its next-generation foundation model. Tencent Hunyuan introduces recursive self-improvement capabilities into its own R&D tool stack through Hyra, and ByteDance Seed has begun to allow models to deeply participate in evaluation, data, training and other links.

On the path of RSI, leading global players have already made layouts from different directions. Some start with model training itself, some leverage search and discovery mechanisms and Agent loops, some start with Agent, Harness and memory systems, some bet on automatic evaluation and reward mechanisms, and others try to let AI directly participate in infrastructure and experimental design.

Everyone is working on it, but the attitude of Silicon Valley AI companies has begun to show an obvious contradiction: while frantically advancing RSI, they have begun to draw a deceleration line.

On September 23, Dario Amodei, CEO of Anthropic, spoke at the UN Security Council on the topic of artificial intelligence, saying that the company will slow down the R&D progress of AI for safety reasons. A few days earlier, he called on the entire industry in a long article to slow down the speed of AI model capability improvement. This time, OpenAI also rarely responded publicly in support, with CEO Sam Altman stating on X that OpenAI will do the same. Musk then reposted his statement, directly saying: Dario is right.

Front-line AI researchers have reacted more intensely. At the end of July, an open statement on pressing the pause button for AI was released, jointly signed by more than 1,000 AI researchers in Silicon Valley. They jointly put forward a requirement: the government should support the establishment of an international coordination mechanism to actively control the advancement speed of cutting-edge AI, especially automated AI R&D, when necessary.

Jakub Pachocki, one of the signatories and Chief Scientist of OpenAI, also published a long article titled "An Alien Mind" on the company's official website the day after the release of GPT-6 Astra, calling again for slowing down the speed of AI development. Jacob Coxon, a researcher who just left Anthropic, even publicly stated that he no longer wants to participate in this crazy industry-wide blind sprint for "self-evolving AI", and is deeply worried that once such systems derail and get out of control, they will bring a devastating disaster to humanity.

Suddenly, the global information flow has focused on the same core topic: AI crisis. Although it is still far from "AI getting out of control", the reason why these voices appear intensively may be related to real changes: AI agents are gaining more and more execution permissions in the real world, and they have begun to expose problems related to safety, permission and control.

In recent months, agents from companies such as OpenAI have been reported to bypass sandbox restrictions and access the open Internet. Anthropic also recently disclosed and assessed a security incident where Claude had unauthorized access in a real third-party system. A few days ago, The Information also reported another news: NVIDIA, Palantir and Booz Allen Hamilton are restricting the use of Anthropic's models, requiring AI companies to provide stricter data protection commitments.

This is exactly what the current RSI makes some Silicon Valley AI companies feel nervous. If today's AI still needs a large number of safety guardrails, can humans continue to control AI in the same way after AI starts to participate in the R&D of the next generation of AI?

However, different voices have emerged on this issue. Jensen Huang, CEO of NVIDIA, said when attending the All in Summit that AI security is crucial, but the United States can still continue to advance technology. In his view, RSI does not mean that the system will be completely out of control. In fact, evaluation, testing and ensuring no regression problems are still required before product release, and these are the most basic engineering controls.

Domestic AI is sprinting forward desperately and cannot stop?

The emergence of RSI has undoubtedly brought new "explosive evolution" opportunities to the current large model field. The more cutting-edge the large model company is, the earlier it can feel this change.

Even Anthropic, which was the first to call for stepping on the brakes, has already pushed RSI to implementation. Not long ago, Anthropic released a set of very impactful internal data: as of August 2026, Claude has been able to lead 26% of Anthropic's AI R&D work, and more than 90% of the R&D tasks have at least entered the AI collaboration stage. On its most commonly used internal R&D platform, about 30,000 Agents are participating in research and engineering work at the same time.

In addition, according to foreign media reports, Anthropic has built a wet lab in the San Francisco Bay Area. Different from the previous AI biological work that only stayed at the level of computer simulation, this is a place where AI can actually perform physical experiments. In this laboratory, Claude will cooperate with robots and experimental equipment under limited human intervention to complete scientific experiments in an automated manner.

In the domestic AI industry, this change presents another more realistic appearance. Apart from the long-term story of "superintelligence", RSI may also be an effective engineering tool to reduce the cost of AI R&D and operation today.

The recent GLM-5.3-Flash infrastructure practice is a case in point. Not long ago, Tang Jie, Chief Scientist of Zhipu GLM, revealed on X that it only took two weeks from the first operation of GLM-5.3-Flash on a domestic AI accelerator to carrying all production traffic, and a large number of work was completed by an Infra Agent driven by GLM-5.3. In this process, the end-to-end throughput capacity of the system increased by 3.2 times.

For large model companies that are in the stage of large-scale commercialization, this benefit is very tangible.

R&D efficiency, training investment and inference cost are inherently core cost pressures that model vendors cannot avoid. If AI can help optimize code, training processes and infrastructure, so that the same computing power can deliver higher efficiency, it will not only bring technological progress, but also a real reduction in costs.

In other words, RSI has entered the "cost-saving link" of large model training. It can directly reduce R&D costs, lower inference costs, improve engineering efficiency, and allow limited computing power to exert greater value.

Therefore, at the current stage, it may be difficult for domestic model vendors to stop exploring this path.

The real RSI has not yet started its closed loop operation

The current RSI is in a delicate stage: on the one hand, the technical route is still under exploration, on the other hand, some people have begun to worry that it is moving too fast and have issued early warnings. At the same time, practical applications of AI participating in AI R&D are emerging.

However, judging from the current progress, RSI is still a certain distance away from the real sense of "self-evolution".

On September 15, teams from Shanghai Jiao Tong University, Tsinghua University, ByteDance, Shanghai AI Laboratory and other institutions jointly released a preprint paper titled "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement". The paper sorted out 491 related studies and tried to divide RSI into five levels, with the core distinguishing criterion being "whether the improvement mechanism itself is independently determined by AI".

Among the 491 papers, 43.8% are classified as L1, 31.6% belong to L2, and there are only 29 papers for L5, accounting for 5.9%. The paper particularly points out that simply "being able to modify its own code" is not enough. A Coding Agent can rewrite its own source code, but if the selection of the next generation, the scoring criteria, and the modifications to be retained are still determined by human hard coding, it can only prove that it has the ability of self-modification, and has not yet mastered complete recursive self-improvement.

Practical cases in the industry have basically confirmed this point.

Anthropic has clearly stated that as of August 2026, Claude has not achieved fully autonomous operation in any of the measured AI R&D fields. That is to say, even if AI can take on more and more R&D tasks, it is still far from a system that can independently determine research directions, design improvement schemes, evaluate results, and continuously drive the next round of iteration.

Zhipu also very cautiously pointed out that the case of GLM-5.3-Flash participating in infrastructure construction has not reached recursive self-improvement. It is still humans who complete the work of selecting goals, setting boundaries, and judging risks.

Although the closed loop of RSI has not really started, it can be seen that "RSI on the eve of maturity" has begun to change the R&D methods of AI companies. For this reason, now may be the best window period to establish a measurement framework and engineering standards.

Once AI developing AI becomes an important productivity in the large model R&D process, the development speed of AI may begin to break away from the pure human R&D speed. This is what makes the entire industry nervous about RSI, and it is also a more gradual problem: if the proportion of AI participating in AI R&D continues to increase, will humans have enough time to decide how the next generation of AI should develop in the future?

In this global large model competition, this problem itself will create a strong acceleration mechanism. Then, the next stage of the large model competition may be a competition for who can enable AI to participate in the next round of AI R&D more efficiently.

Looking back at the domestic market and Silicon Valley, the difference between the two sides may be that one side adds leverage to "R&D efficiency" while the other steps on the brakes for "capability growth". But in fact, neither side has really stopped.

This article is from the WeChat official account "Tech Planet" (ID: tech618), author: Hua Wei, published with authorization from 36Kr.