A junior student verified Luofuli's "outrageous remarks", and we had a chat with him.

In 2026, the term "self-evolution" was frequently mentioned by top figures in the AI industry.

Luo Fuli, the person in charge of Xiaomi's MiMo large model, said bluntly at the Zhongguancun Forum in March, "If I had to use one word to summarize the most crucial thing in the AGI process in the coming year, I'd choose'self-evolution'."

Luo Fuli said that she thought it would take 3 to 5 years for large models to achieve self-evolution a year ago, but now she believes it can be completed in 1 to 2 years.

Dario Amodei, the CEO of Anthropic, is even more radical than Luo Fuli. He once predicted at the Davos Forum that the recursive self-improvement of AI might be achieved within 6 to 12 months. "We may be only 1 to 2 years away from AI autonomously building the next generation of AI."

Once this kind of self-improvement forms a closed loop, the progress speed will grow exponentially.

There was such a paper at the beginning of April, which just talked about how to make AI form such a closed loop. It was from the team of Professor Liu Pengfei at Shanghai Jiao Tong University. The title of the paper is "ASI-Evolve: AI Accelerates AI".

The research team built a closed-loop research framework, allowing agents to iterate repeatedly among "learning existing experiences - proposing new designs - conducting experiments - analyzing results" to automatically improve the three core components of AI: model architecture, training data screening, and reinforcement learning algorithms.

AI will first learn past research experiences, then propose new solutions, then conduct experiments and analyze results by itself, and finally continue to iterate, thus forming a closed loop.

Finally, the result shows that AI can really research how to improve itself.

It's true that this is not a paper that can rewrite the entire AI industry. It just proposes a method.

However, what really shocks people is that the first author of this paper, Xu Weixian, is actually still a junior in college.

An undergraduate who hasn't graduated yet led such a complete and meaningful research... Oh my god! I really wasted my college years!

So Zimu AI really got in touch with Xu Weixian. Through emails, I had a chat with this sunny and handsome genius boy.

01 The Genius Boy Xu Weixian

Xu Weixian has a wide range of technical backgrounds.

He not only focuses on neural architecture search and continuous learning in the field of AI research but also has worked on some system programming projects.

He can independently develop a complete operating system kernel ACore with Rust, implement a compiler Imxc with performance comparable to Clang using C++, and even design a RISC-V processor based on the Tomasulo architecture with Verilog.

This ability to integrate from low-level hardware to high-level AI algorithms is not only rare among undergraduates but also challenging for full-stack engineers.

His GitHub project ASI-Arch has received more than 1,100 stars. For a junior student, this achievement is quite outstanding.

Putting aside these technologies, Xu Weixian also has his own thoughts on AI research. He believes that to achieve "Self-Evolving AI", current AI still lacks two abilities: "continuous self-improvement" and "long-term reliability".

Therefore, his research is divided into two fronts: improving the learning goals and memory mechanisms of a single model to enable it to grow continuously; building a multi-agent ecosystem to allow models to collaborate and optimize through complex interaction protocols.

When talking about the starting point of the ASI-Evolve research, Xu Weixian said that it was an intuition that emerged when he noticed Google's AlphaEvolve in April and May 2025. That work at that time showed people that AI was no longer just helping humans with simple searches but had the potential to drive scientific discovery.

"We hoped to apply this paradigm to AI research itself at that time," Xu Weixian said. "The key to this idea is that since AI technology is developing rapidly, if we can feed back its achievements into its own scientific research process, the entire field can enter a large-scale self-accelerating cycle. This iterative progress is what I think is the most fascinating part of this direction."

Xu Weixian also admitted that the realization of this idea cannot be separated from the support of the GAIR Laboratory at Shanghai Jiao Tong University.

He is very grateful to Professor Liu Pengfei, saying that he strongly encourages undergraduates to explore and conduct research and provides a lot of guidance. The rich resources in the laboratory have enabled them to complete such a large-scale exploration.

In fact, the general public is most likely to misunderstand this kind of paper. They think that researchers are trying to completely replace human scientists with AI, but this is completely wrong.

Xu Weixian said, "In ASI-Evolve, we introduced a large amount of human prior experience. We don't pursue 'blind evolution' without human guidance because the initial experimental purpose and core ideas are always proposed by humans. The real value of the system lies in using AI's strong exploration ability to iterate rapidly in the direction guided by humans. It is more like an extremely efficient collaborative system rather than a cold substitute. ASI-Evolve promotes people to shift from problem-solving and repair to problem definition."

Regarding the term "genius", Xu Weixian's understanding is quite practical. He believes that a genius is a combination of love, talent, and effort.

"Everyone has their own potential loves and things they are good at. What we need to do is find the intersection of love and expertise, and then achieve success through continuous efforts in a better overall direction."

He said, "We see many recognized geniuses. It's not only because of their outstanding achievements but also because the fields they are in are better known to the public. But 'every profession produces its own top talents'. As long as a person can find something interesting and devote themselves to it, they will also become a genius in the field they love."

Xu Weixian modestly considers himself relatively ordinary.

He is very interested in scientific research and hopes to achieve something in the field he likes. At the same time, he also hopes to enjoy life and feel the beauty around him. "Although setbacks take up most of the research process, I still look forward to 'Happy Research'," Xu Weixian said.

What surprises me most about Xu Weixian is that as a researcher, he still maintains the unique enthusiasm for life that young people have while facing academia.

He is a guitarist, playing both classical and electric guitars. He has a Grade 9 clarinet certificate and likes to play badminton and video games.

Moreover, he candidly mentioned on his personal homepage that he is in a relationship and cherishes the time exploring the journey of life with his partner. This balance between being a technical genius and having a life shows a three-dimensional and real image of a young researcher.

I believe you, like me, will envy this sunny young man.

Regarding future research directions, Xu Weixian looks forward to seeing breakthroughs in AI's reflection and continuous learning. "I'm not particularly concerned about the improvement amplitude of each generation of models being trained now because many existing models can already meet most daily needs."

He said, "I'm more concerned about a model's performance throughout its lifecycle, whether during training or deployment - or perhaps in the future, there won't be a distinction between training and deployment phases at all. Whether AI can continuously improve itself not only concerns the upper limit of a model's ability but also is the key to truly improving personalized capabilities and making the model more suitable for each user."

Everyone has their own expression style and unique needs. Only when the model continuously and dynamically evolves in real usage scenarios can it better adapt to users, which may be difficult to achieve by relying solely on static dataset training. In addition, how to achieve a model's more powerful agent capabilities is also exciting. If continuous evolution is exploring the potential of intelligence, this is about enabling existing models to interact with the world more comprehensively.

This is also the reason why Xu Weixian plans to pursue a Ph.D. after graduating from undergraduate. He hopes to work on technologies that he is interested in and that can truly benefit society and be used by everyone during his Ph.D. studies.

02 ASI-Evolve

Xu Weixian believes that the philosophical basis of the ASI-Evolve framework is the process of transforming research progress from "human-limited" to "computationally scalable".

What does "human-limited" mean? Developing a new neural network architecture requires a Ph.D. student to spend 3 months trying 100 designs.

What does "computationally scalable" mean? The ASI-Arch project conducted 1,773 autonomous experiments, consuming more than 20,000 GPU hours, and finally discovered 106 innovative SOTA linear attention architectures. As long as computing power is provided, AI will keep researching.

The core contribution of the ASI-Evolve paper lies in systematically proving the feasibility of "AI accelerating AI" for the first time in a unified framework. This is a breakthrough achieved simultaneously in the three major fields of neural network architecture design, pre-training data screening, and reinforcement learning algorithm design.

In terms of neural network architecture design, the best-performing model achieved a 0.97% improvement, which is nearly three times the gain of the current human-designed SOTA.

More importantly, these architectures are not obtained through brute-force search but are autonomously evolved through a systematic "learning - design - experiment - analysis" cycle.

The system will first learn past research experiences, understand which design principles are effective, and then propose new architecture solutions based on this. The results of each round of experiments will be analyzed and refined, and written into the experience database to provide guidance for the next round of exploration.

This way allows AI to accumulate knowledge like human researchers instead of starting from scratch every time.

Actually, using AI to research AI and improve AI is a very popular track now. From top companies like Anthropic and OpenAI to small teams and laboratories, they are all targeting this field.

In addition to AlphaEvolve mentioned by Xu Weixian earlier, the recently popular Sakana AI Laboratory also proposed a similar concept called The AI Scientist.

Its logic is also to let AI come up with topics, write code, conduct experiments, analyze, and write papers by itself.

Let's get back to Xu Weixian's ASI-Evolve.

In the field of pre-training data screening, the data strategy autonomously optimized by AI improved by 3.96% in the average benchmark test, and the improvement amplitude exceeded 18% in the knowledge-intensive MMLU evaluation.

This means that AI can already understand what kind of data is more valuable for training and autonomously complete the entire process of data cleaning and screening. Traditional data screening relies on the experience judgment of human experts, which requires a large amount of manual annotation and quality evaluation. ASI-Evolve can automatically learn the judgment criteria for data quality through experimental feedback and find truly valuable training samples from a vast amount of data.

In terms of reinforcement learning algorithm design, the brand-new training algorithm designed by ASI-Evolve performed outstandingly on math competition questions. It exceeded the GRPO baseline by 12.5 points on AMC32, 11.67 points on AIME24, and 5.04 points on OlympiadBench.

These are not simple parameter tunings but brand-new algorithm mechanisms with original mathematical innovations. The system can understand the limitations of existing algorithms, propose new optimization objective functions, and design new gradient update strategies.

The significance of ASI-Evolve is that it demonstrates the possibility of AI researching AI by itself.

In the past, every progress of AI relied on the massive manpower invested by human researchers in architecture design, data cleaning, and algorithm tuning.

We always say that AI will replace humans, but this track is very special. If you want AI to progress faster, you can only recruit more researchers.

Now, AI starts to form a closed loop in these core links and directly participates in its own evolution.

In the traditional model, research output is limited by the number and working hours of human researchers. Even the best research teams can only try a limited number of things in a year.

ASI-Evolve shifts this constraint from manpower to computing power. As long as there are enough GPU resources, the system can explore non-stop 7×24 hours a day, and the number of solutions it can try can be dozens or even hundreds of times that of a human team.

But this doesn't mean that human researchers become unimportant.

Xu Weixian emphasized that a large amount of human prior experience is introduced in ASI-Evolve. The system's cognitive library stores the design principles and lessons learned from human research literature, and these knowledge provide directions for AI's exploration.

The initial research goals and evaluation criteria are also set by humans. The role of AI is to conduct efficient exploration in the direction guided by humans, rather than blindly searching the entire possible space.

In this way, you no longer need to spend a lot of time on specific experiments and parameter tuning, but can focus on thinking about what kind of problems are worth researching and what directions are more promising.

AI is responsible for transforming these high-level ideas into specific technical solutions and finding the optimal solution through large-scale experiments.

Another important feature of ASI-Evolve is its analyzer module.

This module can refine complex experimental results into reusable insights.

Traditional automated experimental systems often can only output raw data, which requires human researchers to spend a lot of time analyzing. The analyzer of ASI-Evolve can automatically identify key patterns in experiments, summarize which design choices are effective and which are ineffective, and write these insights into the experience database.

This allows the system to truly "learn" rather than just "search".

Just like doing practice questions, an ordinary person doing practice questions will only record a piece of data, such as how many questions they got right and how many they got wrong out of 1,000 questions. But ASI-Evolve not only records the number of wrong questions but also remembers why they were wrong and how to choose the right answer when encountering similar questions next time.

If we look at it on a larger scale, the model of scientific research has basically remained unchanged in the past hundred years.

Humans propose hypotheses, design experiments, analyze results, and publish papers.

The speed of this cycle is limited by human cognitive ability and working hours.

If the model of ASI-Evolve can be promoted to more fields, it may fundamentally change the speed and scale of scientific research.

03 The Battle for Genius Boys

In the AI era, there are more and more genius boys like Xu Weixian.

Consequently, the recruitment strategies of major manufacturers are undergoing fundamental changes. Those top manufacturers are starting to lock in and deeply cultivate genius boys who are still in school in advance.

For example, the "Crossing Plan" of Dark Side of the Moon.

After a 3 - to 6 - month assessment, interns can get a formal offer and corresponding incentives from this plan even if they haven't graduated yet. The key is that in addition to giving bonuses, it also offers company options.

The valuation of Dark Side of the Moon was $4.3 billion in December last year. By March this year, its valuation reached $18 billion. With such a growth rate, its options are quite attractive.

This plan has almost no rigid requirements. It doesn't limit the major, education level, or experience. It only looks at whether you are "the top talent in any field".

From the company's perspective, locking in talents one year in advance means completing the layout while competitors are still waiting and seeing.

OpenAI's Safety Fellowship represents another model. This project, which runs from September 14, 2026, to February 5, 2027, invites external researchers, engineers, and practitioners to focus on AI safety and alignment research.

Selected participants will receive monthly allowances, computing resource support, and in - depth guidance from OpenAI mentors. The project expects participants to produce substantial research results, such as papers, benchmark tests, or datasets, at the end.