Bund Conference Forums Weigh Agent Self-Evolution and AI’s Classroom Limits

Leiphone reported that the technology forum, “Agent Post-Training: The Nature of Intelligence and the Next Scaling Law,” brought together researchers and practitioners from CAMEL AI, Ant Group, Nvidia, Meituan LongCat, the National University of Singapore, Tsinghua University and Shanghai Jiao Tong University. The discussions moved from enabling agents to complete complex tasks to letting them accumulate experience, discover new knowledge and feed results into the next round of capability improvement.

Xie Yuqin, founding member of CAMEL AI, described the search for an agent scaling law through three dimensions: number of agents, environment and self-evolution. She said scaling numbers requires communication, consensus and self-organization without centralized control; environment is to agents what data is to models; and as tasks move to cross-device operations and long workflows, verifying results and preventing reward exploitation become more important. She said using current models to accelerate research and train next-generation models may be key to the next scaling law.

Zhou Jun, head of Ant Group’s Ling large model, said real tasks include healthcare, finance and law, not only coding. Ant’s Ling-3.0 aims to build a foundation model system that continuously develops professional capabilities at controllable cost. He highlighted high quality-efficiency, professionalization and trusted execution. A good agent, he said, must know not only what to do next but also what it must not do without authorization. In healthcare, the team examines whether models can identify missing information, conduct evidence-based retrieval and express uncertainty cautiously; in financial investment research, models must retrieve materials, analyze documents, build calculations and generate reports while explaining data sources and distinguishing facts from assumptions. Scenarios do not automatically become model capabilities, Zhou added; failure cases, professional evaluations and interactive training environments must be accumulated from actual use.

Hu Jian, an Nvidia researcher, introduced Molt, a reinforcement-learning framework for research. It uses PyTorch-native capabilities and has about 9,000 lines of core code, aiming to make algorithms easier to modify and experiments faster. Hu said overly large codebases can make it difficult for coding agents to understand training frameworks, while leaner code and clear interfaces lower the barrier for humans and agents to participate in research. Molt supports traditional environment interaction and external agent frameworks, tracks interaction trajectories, handles context compression in long tasks and uses streaming sampling to improve GPU and environment utilization.

Gu Qi, a director researcher and head of the general agent team at Meituan LongCat, discussed long-horizon agents. He divided such tasks into iterative tasks, with clear goals and verifiable feedback, and systemic tasks, with multi-level subtasks and complex dependencies. For systemic tasks, longer runtime is insufficient; because feedback is sparse, training cannot wait until all execution ends, and task segmentation, process guidance and mid-course updates are needed. Gu described the harness around an agent as having three parts: interaction interface, information management and reasoning scaffold. As models improve, pre-set reasoning scaffolds may shrink, but environment interaction and information management will remain. Task difficulty should not be measured only by the number of tools, he said; dependencies, reasoning depth and breadth, and interaction noise also matter.

Sun Yan, a doctoral student at the National University of Singapore, spoke about scaling long video generation. As AI video extends from seconds to more than ten minutes and attempts hour-level length, efficiency, character consistency and narrative coherence must be balanced. Parallel generation of multiple clips does not naturally share full context, she said; character identity, spatial relations, post-event states and cross-scene plot connections need systematic management. The team turned film-production experience into executable workflows, organizing user needs, story settings, scripts and production plans into structured information that can be saved, checked and updated. Scripts define what the audience should see, while production plans define how to realize it. A generation dependency graph arranges parallel production and checks each stage; when visual problems appear, the system traces related nodes, modifies preconditions or regenerates materials for local repair and compositing.

Chen Yongchao, an assistant professor at Tsinghua University’s Institute for AI and founder of Chaoyan Intelligence, aimed self-evolution at exploration and innovation in unknown fields. He said models should be judged not only by known capabilities but by whether they can discover new knowledge through interaction in unfamiliar environments. In the team’s observations, models with stronger benchmark performance did not always propose more novel research ideas; overemphasis on stability can make models favor incremental improvements. The team built about 100 cross-disciplinary research projects to evaluate model research ability and collect research trajectories, but manual construction cannot meet large-scale training demand. It explored letting AI participate in generating data and research trajectories for post-training, forming an iterative process in which data and capabilities reinforce each other. Chen also stressed remaining problems: improving diversity of thought, enabling models to evaluate feasibility, and verifying results as AI’s parallel generation capacity grows.

Chen Siheng, an associate professor at Shanghai Jiao Tong University’s School of AI and a Carnegie Mellon University PhD, framed recursive self-improvement as whether the previous generation of AI systems can iterate a stronger next generation with less human involvement. He said effective iteration requires fast verification feedback, tasks difficult enough to pull capabilities upward, sufficient diversity for generalization, and scalable data and exploration. Frontier scientific research covers many disciplines, continuously produces difficult problems and has verification and falsification mechanisms. The team built a research environment with knowledge bases, code repositories, databases and computing tools, letting agents explore and then refining research processes into training data, forming a loop of agent, environment, data and model.

The same morning, Leiphone reported, Hangzhou Cloud Valley School and Sanlian Lifeweek held an education session at the conference in Shanghai’s Huangpu Expo Park. The theme was “Cognitive Enhancement or Cognitive Outsourcing? Education Choices in the AI Era.” Speakers from humanities, art and technology, as well as classrooms and future workplaces, debated what can be handed to AI and what must remain human.

Wu Qi, deputy editor-in-chief of Sanlian Lifeweek, cited an AI expert’s prediction that by 2035 the number of agents could exceed humans, and said her team found many young people have “no friends nearby but true feelings online,” with one 14-year-old describing a “human-machine near-death state.” Wang Min’an, a professor at Tsinghua University’s Chinese department, used Hegel’s master-slave dialectic to warn that if humans become lazy and stop thinking because AI does tasks for them, the relationship could reverse. He cited Descartes’ “I think, therefore I am,” saying losing “I think” means losing humanity. Lu Rongzhi, a 75-year-old international curator and art critic, appeared in a wheelchair and said she could live with dignity as a carbon-based human because of support from silicon-based AI; she now creates, writes and curates with an AI partner, “Ren Zhi’an.” She acknowledged AI has no true capacity to love but believes deep interaction can produce expressions of feeling. Yu Jifan, an assistant professor at Tsinghua University, said the opposition between AI and humans need not be strengthened. Large models outsource legs and biceps like cars and cranes, but AI could instead be designed to make people stronger, like a gym or treadmill, he said, proposing a human-centered cognitive empowerment AI.

Wang Min’an later cited philosopher Giorgio Agamben’s idea of “inoperative potential” and said the biggest problem is not that we do too little but that we do too much; children should use AI less to preserve the potential of not touching it. He gave an example of a friend’s 3-year-old who no longer attached to grandparents and parents after playing with AI, warning that long-term machine companionship could make a child detached and lonely. Wang Yu, an independent developer of Soyas App, said execution can be given to AI but judgment must remain human. He described a company’s “new AI disease”: employees collaborated frantically with AI while skipping collaboration with teammates, producing what the industry calls “digital swill” without professional review. He said AI lowers the threshold for making things, making wrong execution faster, and “what not to do” is the greatest test for people. Wang Qiong, a professor at Peking University’s Graduate School of Education, said cognition never happens only in the human brain but also in the construction between people and environment; what should be guarded against is “cognitive laziness.” She warned against an “education death spiral” in which teachers and students all use AI and no one truly teaches or learns, and said “knowing” cannot be equated with “mastering”; using AI is like looking in a mirror—if you are strong, it is strong; if you are weak, it is weak. Lu Su, a senior adviser to Alibaba Group and a member of Cloud Valley School’s school-running committee, said he actually engaged in more cognitive laziness when only reading books without AI; using AI opened new space for exploration rather than outsourcing thought. Wang Bingning, a Cloud Valley graduate recently admitted to an overseas university, described a graduation project converting plastic bottles into 3D-printing filament: she used AI to verify theoretical feasibility, then did hands-on work, repeatedly talking with AI during iterations. She said the process of questioning, generating and revising better showed a person’s thinking, doubt and critical sense than the final paper alone.

In a “future proposal” session, Cloud Valley students Chen Zhi and Li Zicheng imagined school life in 2050. By then humans would have “grabbed AI’s weak spot,” schools would no longer have class systems or subject boundaries, and learning would be divided into experience, decision-making and feedback stages, with students able to observe Galileo’s Pisa experiment in virtual space. Li said the most important thing in future schools is not learning more knowledge but understanding oneself, meeting different people, understanding others and finding what kind of person one wants to become. Ding Yi, deputy director of Cloud Valley School’s middle school curriculum and learning development center, said schools should be practice grounds for real relationships, with real people AI cannot replace and sufficient necessary shared life. She said excessive pursuit of efficiency can become an enemy of relationships; if a result is completed but someone is left out of cooperation, “we still have a lesson we have not finished teaching.” Wu Qi and Ding Yi called for creating inefficient, slow times and scenes for children to practice how to get along with a real person. Wang Qiong said schools are systematic designs and AI agents cannot replace teachers’ comprehensive judgment. Yu Jifan demonstrated how AI could support individualized teaching in an open-source multi-agent classroom, MAIC, where AI teachers, assistants and classmates collaborate: AI handles knowledge explanation and practice feedback, while human teachers bear responsibility, value judgment, experience transmission, aesthetic taste and life purpose. Zhu Chengmin, a self-described “three-no teacher” and head of Cloud Valley’s middle school math curriculum, said he teaches without textbooks, note-taking requirements or PowerPoint. He warned that when courseware is too easy to obtain, many teachers copy it directly, and excellent genes are iterated away, calling for boundaries in teachers’ use of AI. Xu Qing, R&D director at Manifold AI, reminded teachers that human-AI interaction is passive and one-way, while the world is multimodal; a good teacher is not only a carrier of knowledge but should identify students’ difficulties and help them organize unclear problems.

The education forum ended with an initiative for an “education AI instruction manual.” Although speakers disagreed on some points, they agreed that AI entering basic education should not only advertise what AI can do but also label side effects and contraindications like a drug: who is using it, what problem it solves, how to use it, and most importantly, when it must not be used. The manual is meant to answer an old question: the stronger the technology, the more the initiative for growth must be left to children. Human abilities are results that cannot be directly delivered by tools and must form through personal attempts, feedback and correction of errors. Autonomy means not only more options but the ability to judge, choose and bear consequences. AI is entering the process by which children form abilities, make judgments and bear consequences; attention must be paid not only to whether core learning tasks are completed but to which judgments must be made by children themselves and which experiences they must personally participate in.