Betting on the same direction as Jeff Dean, a 27-year-old assistant professor at Tsinghua University founded his startup and secured hundreds of millions of yuan in financing in just two months.
By Wang Xinyi
Edited by Zhang Yuxin
This year, using AI to self-improve AI has almost become the consensus among top labs at home and abroad.
In the past year, Andrej Karpathy, founding member of OpenAI, Tian Yuandong, former research director of Meta FAIR, and Jeff Dean, former chief scientist of DeepMind, have successively joined the upsurge of RSI (Recursive Self-Improvement) entrepreneurship.
What happened to RSI is almost a story highly similar to that of the world model — even the most advanced laboratories have not explored breakthrough achievements in recursive self-improvement. The technical uncertainty was quickly overwhelmed by grand narratives and the hot capital market. Tian Yuandong is a typical example. In May this year, his startup Recursive Superintelligence had not yet delivered results, but it had officially announced the completion of a $650 million financing, pushing the valuation to $4.65 billion.
This upsurge has also spread to China. In January this year, Chen Yongchao, who was studying in the joint training program of Harvard and MIT and was about to graduate with a doctorate, stood at the crossroads of career choices: on the one hand, there was a job offer from DeepMind, and co-founder invitations from many overseas start-ups focusing on AI automatic scientific research; on the other hand, the entrepreneurial timing he had waited for two years had just arrived.
What he wants to do is to train a general model that can self-evolve, so that it can independently complete the research done by human scientists on a set of self-evolvable Harness, including conceiving ideas, retrieving literature and running experiments.
After a series of considerations, he chose the latter, returned to China to found Apex Intelligence, focusing on the R&D of self-evolving foundation models, and joined Tsinghua University at the same time.
Chen Yongchao, 27, became the youngest assistant professor in the history of the School of Artificial Intelligence of Tsinghua University. Apex Intelligence, which he officially announced the establishment of in July, completed two rounds of 100-million-level financing in just two months.
According to The Rise of Intelligence, Apex Intelligence has recently completed nearly 400 million RMB in angel round and angel+ round financing. The angel round was jointly led by IDG Capital, Starlink Capital, and XtalPi, followed by D&X Capital, Infinite Fund, InCapital, and Azure Capital, bringing together top dollar funds, industrial capital, and market-oriented investment institutions. The angel+ round was jointly led by top state-owned assets in Beijing, Shenzhen, and Shanghai — Zhongguancun Science City Fund, Shenzhen Venture Capital Group, and Shanghai Future Industry Fund.
This round of financing will be mainly used for the construction of foundation models and computing power infrastructure, investment in R&D of recursive self-improvement technology, and expansion of the core team.
This is a research-led entrepreneurial team. Chen Yongchao, founder of Apex Intelligence, has conducted large model-related research in overseas laboratories such as Google DeepMind, Microsoft, and IBM. The core members of the team come from top large model manufacturers such as ByteDance, Kimi, and Zhipu AI, with first-hand model R&D experience.
Different from RSI that builds the Agent layer, Apex does not focus on building Harness and applications for AI for Science, but explores the recursive self-evolution capability at the foundation model layer, so that the trajectories generated by each research task can be turned into training data to improve the model and further overcome more difficult problems.
Inside Apex, this logic is summarized as "Research is the engine of RSI", which means research is the engine of recursive self-improvement. Its ultimate goal is to connect the digital world and the physical world with a set of recursive intelligent systems, and expand unified capabilities to various fields such as AI, mathematics, physics, chip design, robotics, industrial manufacturing, chemistry, and biology.
"After the equalization of Coding, the next wave will be the equalization of Research." In Chen Yongchao's view, the self-evolution of AI is becoming a reality.
The first strong evidence in his hand came from May this year. They let the company's self-evolving AI System independently produce 34 papers, and the results were quite surprising. After these papers were submitted to ACL Rolling Review (the unified review platform for ACL series conferences), 11 of them scored more than 3 points, which is a preliminary review score no lower than the doctoral level. Among them, 2 papers scored higher than 99% of human researchers.
"Many ideas are quite eye-opening, and some papers are almost at the level of doctoral students, which is beyond our expectation but also reasonable, because this moment will come sooner or later," Chen Yongchao said.
In addition to papers, from AI independently conducting AI-direction research to exploratory experiments in basic science fields such as mathematics, Apex's AI system has completed preliminary verification in many high-difficulty research fields. At present, the company has precipitated tens of thousands of high-quality research trajectories, covering expert trajectories, synthetic trajectories, and research trajectories independently generated by AI.
△ Research is the engine of recursive self-improvement, Source: Enterprise
If we put the evolution of intelligence on a coordinate axis, Chen Yongchao believes that it does not rise smoothly, but follows an S-shaped curve, breaking through a bottleneck, climbing rapidly, then flattening out, and waiting for the next bottleneck. At this stage, we are in the stable stage after the inflection point of Scaling Law, waiting for the emergence of a new inflection point and the next moment of rapid intelligence explosion.
Chen Yongchao is firmly convinced that this inflection point is about to come. "The times will not wait for you. This is the same as the situation where you did not do large models two years ago and did not do embodied intelligence one year ago. RSI is the best entrepreneurial time point now."
Regarding returning to China to start a business, choosing the RSI direction, and thinking about the evolution of intelligence, we came to the office of the School of Artificial Intelligence of Tsinghua University with a series of questions and had an exchange with him. The following is the record of the conversation between The Rise of Intelligence and Chen Yongchao, slightly edited:
The stability and innovation of the model are contradictory
The Rise of Intelligence: Why did RSI suddenly break out this year?
Chen Yongchao: Many people think that the model capability has reached a bottleneck period and new paths need to be explored, but I think this is a secondary reason. The main reason is that the technological turning point has come, and the model is strong enough to do a lot of research done by human researchers. This is also the direction we are exploring, to train the next generation of models so that they can reach or even surpass the level of human researchers.
The Rise of Intelligence: In simple terms, what is the difference between a self-evolving model and the current large models?
Chen Yongchao: The current large models are stability-oriented models for the public, conservative, and designed to avoid making mistakes. For example, if you ask it 100 questions, it can answer 95 of them, and for the remaining unanswerable questions, we try to reduce its hallucinations or let it directly answer that it does not know.
But self-evolving models are innovation-oriented models, and they do not have high requirements for stability. For example, if you ask it a question that it cannot answer, we expect it to propose 100 different methods to solve the problem. It does not matter if 99 of them are absurdly wrong, as long as one is correct. That is to say, a large number of method failures are allowed in the exploration process, but as long as one of them is proved effective, the system can identify it, retain it, and turn this success into reusable capabilities in the future.
The innovation and stability of the model are often mutually exclusive. If the model is too conservative and stability-oriented, its innovation will be weak. Sometimes we even find that the better the general evaluation result of the model, the worse its innovation.
The Rise of Intelligence: To realize the self-evolution of the model, what are the key technical bottlenecks?
Chen Yongchao: The first is the ability of the model to come up with imaginative ideas; the second is the autonomous execution capability; the third is the verification capability. At present, whether the model can come up with good ideas, screen out good ideas, execute them independently and verify them well is a particularly big bottleneck.
The Rise of Intelligence: In what ways can these bottlenecks be solved?
Chen Yongchao: The training method and model architecture of the model, especially the upper-layer architecture, need to be adjusted, and the data will also change greatly. For example, use some adversarial training methods to let a model play different roles: it can not only put forward good ideas, but also evaluate itself to judge how reasonable and innovative the ideas are. Through iterative self-improvement methods, the ideas will get better and better, and the level of creativity will get higher and higher.
The Rise of Intelligence: Why do you choose to implement the self-evolution capability of the model starting from the scientific research scenario?
Chen Yongchao: The scientific research scenario has the most urgent demand for intelligent brains. Self-evolution itself is pursuing the upper limit of intelligence, which is to benchmark or even surpass human researchers. It usually targets those scenarios that are at the upper limit of intelligence and most in need of innovation.
The Rise of Intelligence: What is the relationship between AI for Science, AI Scientist, Auto-research and Recursive Self-improvement? Is there any overlap between them?
Chen Yongchao: In my opinion, although RSI, Auto-research and AI Scientist are not synonyms in the literal sense, they essentially point to the same goal, which is to build a general research intelligence that can independently carry out research and continuously improve its own capabilities in the research process. AI Scientist describes the role that this system plays, Auto-research describes the process of it independently completing research, and RSI describes the mechanism by which it continuously improves itself based on research results.
AI for Science is not at the same conceptual level as them. AI for Science describes where AI is applied, that is, using AI to solve scientific problems; it can be a scientific research tool, a vertical domain model, or a complete independent research system, but it does not necessarily have the capabilities of independent research and self-improvement.
Although Auto-research and AI Scientist appeared earlier, many people working in these two directions have narrowed the concept. Auto-research is not just building an Agent to write papers, and AI Scientist is not just training a small model in a vertical field. The real problem they want to solve is to train a general model to surpass human researchers, so that AI can become the coworker of human researchers.
The Rise of Intelligence: What is the foreseeable upper limit of RSI? What level can it reach?
Chen Yongchao: It depends on which level of human researchers it can finally reach. It is hard to say whether it will finally reach the level of doctoral students in top universities such as Tsinghua, or the level of Newton and Einstein.
The Rise of Intelligence: When will the explosive growth of intelligence occur?
Chen Yongchao: I think there will be an inflection point, which may occur when the model is more efficient than humans in some links, or its innovation capability is similar to that of humans. It advances according to bottlenecks and stages. Once a bottleneck is broken through in a certain period of time, it will rise in an S-shaped curve, develop rapidly first, and then gradually flatten out. At that time, you will start to wait for the next bottleneck that can be broken through.
The Rise of Intelligence: What stage is intelligence in now?
Chen Yongchao: The first S-shaped curve is Scaling Law, and it is now in a stage where the first S-shaped curve is nearly saturated. At this stage, the capability improvement of model iteration is no longer so fast, and it has reached the upper end of the S-shape. RSI may be the next S-shape, or it may be a methodology that helps discover many S-shaped curves to break through bottlenecks.
Do research like human scientists
The Rise of Intelligence: Overall, what are you doing now?
Chen Yongchao: We have built an AI System for automated scientific research, letting it do AI for AI, AI for Math, and various disciplines. At the same time, we are training our own large models and building a benchmark that can evaluate the capabilities of large models in the research field.
The Rise of Intelligence: This self-evolving AI System independently completes a scientific research project. What is its end-to-end process? Are there any specific cases?
Chen Yongchao: It is very similar to human research, and it is "full-stack" like a human scientist, able to do experiments by itself: conceive ideas by itself, do experiments, connect GPUs, build environments, train models, finally improve methods based on experimental results, and summarize the results into reports, code repositories, products or papers.
For example, it can now help mathematicians make major mathematical breakthroughs and prove theorems; it can also do things in the AI field, optimize the kernel, data distribution, and operators in AI training by itself; it can also do data simulation of robot VLA and improve the VLA architecture.
The Rise of Intelligence: How long is the cycle for a research project in this system? How much faster can it be than human research?
Chen Yongchao: It depends on the difficulty of the research. For more difficult projects, AI may take two or three weeks, which is already very fast compared with humans, while humans may take several months or even a year.
The Rise of Intelligence: Has this product been launched?
Chen Yongchao: It has been launched recently and is in a small-scale beta test.
△ Principle of automated AI research system, Source: Enterprise
The Rise of Intelligence: Can you elaborate on the model you are training now?
Chen Yongchao: First of all, this model is definitely a large-parameter-scale model.
The big difference between it and the previous generation of models is that we focus on innovation and research capabilities. It may be slightly worse than the closed-source SOTA model in some general capabilities such as programming and mathematics, but it will be significantly stronger in complete research capabilities.
The Rise of Intelligence: Are you training the model from scratch?
Chen Yongchao: We are now doing intermediate training and post-training based on open-source models. Pre-training requires more capital reserves.
The Rise of Intelligence: What effects can your model achieve now?
Chen Yongchao: From our own perspective, some aspects are quite amazing, such as data self-evolution and Harness self-evolution, but currently we can only make small optimizations, and it is difficult to make great breakthroughs in innovation.
We want to train a general model first, which can do research in many fields, such as AI for AI, mathematics, robotics, chip design, and quantum. Among them, a lot of data of the general model will cover various fields, and after fine-tuning based on the general model, it can also be quickly implemented in vertical scenarios.
The Rise of Intelligence: Why do you build a general model first, and then gradually develop various vertical scenarios?
Chen Yongchao: Innovations in many fields come from inspirations from unrelated fields. If you only use data from vertical fields to train the model, it may become stronger and stronger in local tasks, but the overall intelligence will decline, which is what we often call "catastrophic forgetting".
For example, if I want it to do AI for large models, and let all its training data and research trajectories come from the field where AI itself trains large models. As a result, you will find