Text｜Wang Xinyi

Editor｜Zhang Yuxin

This year, using AI to improve AI has almost become a consensus among leading laboratories both domestically and abroad.

In the past year, founding member of OpenAI Andrej Karpathy, former Meta FAIR research director Tian Yuandong, and former DeepMind chief scientist Jeff Dean have successively joined the entrepreneurial wave of RSI (Recursive Self-Improvement).

The story happening with RSI is almost a highly similar narrative to world models — even the most advanced laboratories have failed to discover breakthrough results regarding recursive self-improvement, and the uncertainties in technology have soon been drowned out by grand narratives and a hot capital market. Tian Yuandong is a typical example; in May this year, his startup Recursive Superintelligence, which had yet to deliver results, announced it had completed $650 million in financing, boosting its valuation to $4.65 billion.

This wave has also spread to domestic companies. In January of this year, Chen Yongchao, who is studying in a joint program at Harvard and MIT and is about to graduate with a doctorate, found himself at a crossroads in his career: on one side, a job offer from DeepMind, as well as co-founder opportunities from several overseas startups focused on AI-driven research; on the other side, the entrepreneurial timing he had been waiting for two years was just arriving.

What he wants to do is train a self-evolving general model, allowing it to independently complete research tasks being done by human scientists, including forming ideas, retrieving literature, and running experiments, on a set of self-evolving Harness.

After a series of considerations, he chose the latter and returned home to establish Apex Intelligence, focusing on the research and development of self-evolving foundational models while also taking a position at Tsinghua University.

At 27, Chen Yongchao became the youngest assistant professor in the history of Tsinghua University's School of Artificial Intelligence. The company he officially founded in July, Apex Intelligence, completed two rounds of financing of over 100 million yuan within just two months.

According to "Intelligent Emergence", recently, Apex Intelligence has completed nearly 400 million yuan in angel and angel+ round financing. The angel round was co-led by IDG Capital, Xinglian Capital, and Jingtai Technology, with follow-up investments from Dexion Capital, Wuxian Fund, Chuxin Capital, and Yunxi Capital, bringing together leading dollar funds, industrial capital, and market-oriented investment institutions. The angel+ round was co-led by the top state-owned assets in Beijing, Shenzhen, and Shanghai — Zhongguancun Science City Fund, Shenzhen Capital Group, and Shanghai Future Industry Fund.

This round of financing will primarily be used for building foundational models and computational infrastructure, investing in R&D for recursive self-improvement technology, and expanding the core team.

This is a startup team led by researchers. Founder Chen Yongchao previously worked on large model-related research in overseas labs such as Google DeepMind, Microsoft, and IBM, and core team members come from leading large model firms like Byte, Kimi, and Zhipu, with frontline model development experience.

Unlike the RSI focusing on the agent layer, Apex does not concentrate on developing AI for Science applications and Harness; rather, it explores the recursive self-improvement capabilities at the foundational model layer, allowing research tasks to generate trajectories that become training data to improve the model and tackle more difficult problems.

Internally at Apex, this logic is summarized as "Research is the engine of RSI," meaning research serves as the engine for recursive self-improvement. Its ultimate goal is to connect the digital world and the physical world with a recursive intelligent system, expanding unified capabilities across various fields such as AI, mathematics, physics, chip design, robotics, industrial manufacturing, chemistry, and biology.

"After coding equality, the next wave will be the equality of research." In Chen Yongchao's view, AI's self-evolution is becoming a reality.

The first solid evidence he has is from this May when they let their self-evolving AI System autonomously produce 34 papers, yielding quite surprising results. These papers submitted to ACL Rolling Review (the unified review platform for ACL series conferences) saw 11 achieving scores above 3, a score not lower than doctoral level in preliminary review, including 2 papers scoring above 99% of human researchers.

"Many ideas were quite eye-catching, with some papers reaching nearly doctoral levels, which was beyond our expectation, but also logical, as this moment was bound to arrive," said Chen Yongchao.

Apart from the papers, from AI conducting AI research to exploratory experiments in fundamental sciences like mathematics, Apex's AI system has completed initial validation in several challenging research areas. Currently, the company has accumulated tens of thousands of high-quality research trajectories, including expert trajectories, synthesis trajectories, and AI-generated research trajectories.

△ Research is the engine of recursive self-improvement, source: company

If we were to place the evolution of intelligence on a coordinate axis, Chen Yongchao believes it is not a smooth ascent, but follows an S-shaped curve; it breaks through a bottleneck, climbs quickly, then levels off, waiting for the next bottleneck. At this stage, we are in a stable phase belonging to the previous Scaling Law inflection point, waiting for the emergence of a new inflection point, and the next moment of rapid intelligence explosion.

Chen Yongchao is adamant that this inflection point is about to happen, "The times cannot wait for you; this is the same situation as two years ago when we did not engage with large models, and a year ago when we did not pursue embodied intelligence; now is the best time for RSI to venture."

Regarding his return to China for entrepreneurship, choosing the RSI direction, and thoughts on the evolution of intelligence, we brought a series of questions to the office of the School of Artificial Intelligence at Tsinghua University and had a discussion with him. Below is an excerpt of the dialogue between "Intelligent Emergence" and Chen Yongchao:

The stability and innovativeness of models are in opposition.

Intelligent Emergence: Why did RSI suddenly explode this year?

Chen Yongchao: Many people believe it is due to the model's capabilities reaching a bottleneck, requiring exploration of new paths; I think that is a secondary reason. The primary reason is that a technological turning point has arrived, and models have become powerful enough to perform many tasks typically conducted by human researchers. This aligns with our exploration direction; training the next generation model to reach or even exceed human researcher levels.

Intelligent Emergence: In simple terms, what is the difference between self-evolving models and current large models?

Chen Yongchao: Current large models are stable models that are conservative, designed for public use, and avoid making mistakes. For example, if you ask it 100 questions, it might answer 95 while attempting to minimize hallucinations for the remaining questions or directly responding with "I don't know."

But self-evolving models focus on innovativeness and do not demand high stability. For instance, if you ask it a question that it cannot answer, we hope it can propose 100 different approaches to solve that problem, with 99 of them potentially being wildly incorrect, as long as one is correct. This means that throughout the exploration process, numerous methods can fail, but as long as one is proven effective, the system can identify and retain that, converting this success into a reusable capability for the future.

Model innovativeness and stability are often incompatible; if a model is overly conservative, its innovativeness weakens. Sometimes we even find that the better the model's general evaluation results, the worse its innovativeness.

Intelligent Emergence: What are the key technical bottlenecks in achieving the self-evolution of models?

Chen Yongchao: The first is the capability of the model to propose imaginative ideas; the second is autonomous execution capability; the third is verification capability. Currently, whether a model can propose good ideas, autonomously execute them after selecting the good ones, and perform verification is a significant bottleneck.

Intelligent Emergence: How can these bottlenecks be addressed?

Chen Yongchao: The training methods and model architecture, especially the upper architecture, need adjustments, and there must be significant changes in data. For example, using some adversarial training techniques to have a model take on different roles: being able to propose good ideas, independently evaluate them, and judge how well they perform in terms of rationality and innovativeness. Through iterative self-improvement methods, ideas can become progressively better, and creative levels can rise.

Intelligent Emergence: Why start this self-evolution capability in research scenarios first?

Chen Yongchao: Research scenarios have the most urgent need for intelligent brains. Self-evolution inherently pursues the upper limit of intelligence, which aims to match or even exceed human researchers. It typically targets scenes that are at the intelligence upper limit and require the most innovation.

Intelligent Emergence: What is the relationship between AI for Science, AI Scientist, Auto-research, and Recursive Self-improvement; is there overlap?

Chen Yongchao: In my view, while RSI, Auto-research, and AI Scientist are not literal synonyms, they essentially point towards the same goal, which is to build a general research intelligence capable of conducting research autonomously while continuously improving its own abilities during the research process. AI Scientist describes the role of this system, Auto-research describes the process of it autonomously completing research, while RSI describes the mechanism by which it continuously improves itself based on research results.

AI for Science and these concepts are not on the same conceptual level. AI for Science describes where AI is applied, i.e., using AI to solve scientific problems; it can be a research tool, a vertical domain model, or a complete autonomous research system, but it does not necessarily possess autonomous research and self-improvement capabilities.

Although Auto-research and AI Scientist appeared earlier, many practitioners in these two directions have narrowed the concept; Auto-research is not just about setting up an agent to write papers, nor is AI Scientist merely about training a small model in a vertical domain. The real issue they aim to solve is training a general model that surpasses human researchers, allowing AI to become a co-worker of humans.

Intelligent Emergence: What are the predicted limits of RSI, and what level can it ultimately reach?

Chen Yongchao: It depends on which level of human researcher it can eventually become; it is difficult to say whether it will reach the level of doctoral candidates at leading institutions like Tsinghua or of Newton and Einstein.

Intelligent Emergence: When might an explosive increase in intelligence occur?

Chen Yongchao: I believe there will be an inflection point, possibly occurring when the model's efficiency surpasses human capabilities in certain aspects or its innovative ability becomes comparable to that of humans. It advances on the basis of bottlenecks and stages; once a bottleneck is broken through in a certain period, it will rise on the S-shaped curve, initially developing rapidly before gradually leveling off, at which point you will start waiting for the next breakable bottleneck.

Intelligent Emergence: What stage is intelligence currently at?

Chen Yongchao:The first S-shaped curve is the Scaling Law, and we are now approaching the saturation phase of that first S-shaped curve. Currently, the iteration of models has lost some speed; we have reached the upper end of the S curve. RSI may represent the next S curve, or it may help discover many methods to break through the bottlenecks of multiple S curves.

Doing research like human scientists

Intelligent Emergence: Overall, what are you currently working on?

Chen Yongchao: We have developed an AI System for automated research, allowing it to do AI for AI, AI for Math, and other disciplines. Meanwhile, we are training our own large model and establishing benchmarks to evaluate the capabilities of large models in research areas.

Intelligent Emergence: What is the end-to-end process for this self-evolving AI System to independently complete a research project? Are there any specific cases?

Chen Yongchao: This process is quite similar to human research, and it can operate in a full-stack manner like human scientists: generating ideas, conducting experiments, connecting to GPUs, setting up environments, training models, and finally improving methods according to experimental results, summarizing the findings into reports, code repositories, products, or papers.

For example, it can now help mathematicians discover significant breakthroughs in mathematics, prove theorems; it can also work on AI-related tasks, optimizing cores, data distributions, and operations in AI training; and it can conduct data simulations for robot VLAs and improve VLA architectures.

Intelligent Emergence: In this system, how long is the cycle generally for a project’s research? How much faster is it compared to human research?

Chen Yongchao: It depends on the difficulty of the research project; for projects with considerable difficulty, AI may take two to three weeks, but it is much faster than humans, who might take several months or even a year.

Intelligent Emergence: Has this product been launched yet?

Chen Yongchao: It has recently gone live and is currently in a small-scale testing phase.

△ Principles of the Automated AI Research System, source: company

Intelligent Emergence: Can you elaborate on the models you are currently training?

Chen Yongchao: Firstly, this model must be a large parameter model.

It differs significantly from the previous generation of models as we emphasize innovativeness and research capability. It may lag behind closed-source SOTA models in some general capabilities, such as programming and mathematics, but will significantly outperform those in overall research capabilities.

Intelligent Emergence: Are you training the model from scratch?

Chen Yongchao: We are currently training on an open-source model and fine-tuning it; pre-training requires more funding reserves.

Intelligent Emergence: What performance can your model achieve now?

Chen Yongchao: We see some stunning aspects in areas like data self-evolution and Harness self-evolution, but currently, we can only do relatively minor optimizations; making significant breakthroughs in innovativeness is still quite challenging.

We aim to train a general model that can research in many fields like AI for AI, mathematics, robotics, chip design, and quantum studies. The data for the general model will include various domains, and later, based on this general model, we will fine-tune it so it can quickly land in specific vertical scenes.

Intelligent Emergence: Why start with a general model before gradually working on various vertical scenes?

Chen Yongchao: Much innovation in various fields stems from inspirations from seemingly unrelated areas. If you only train the model with vertical domain-specific data, it may become stronger in localized tasks, but overall intelligence may actually decline, which is what we often refer to as "catastrophic forgetting."

For instance, if I want it to do AI for large models, and all its training data and research trajectories come from the area where AI trains large models itself, you may find it gradually overfits that distribution, forgetting the mathematical, physical, chemical, or even common-sense reasoning capabilities it originally possessed. If it forgets even common sense, its intelligence will degrade; it may become increasingly expert in repeating established large model development paradigms while finding it harder to escape local optima and produce genuinely breakthrough new methods.

Intelligent Emergence: You recently released a benchmark called ASI Bench.

Chen Yongchao: Yes, this is a benchmark evaluating the model's exploration capacity and research capabilities. The main difference between ASI and AGI is that ASI can explore techniques to enhance the upper limits of intelligence and …

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。