At the Beijing Zhiyuan Conference on June 12, Beijing Zhiyuan AI Research Institute Director Wang Zhongyuan, Xiaomi MiMo large model lead Luo Fuli, Tsinghua University Computer Science professor and MianBi Intelligence co-founder Liu Zhiyuan, Tsinghua University AI Research Institute deputy director and ShengShu Technology founder Zhu Jun, and Nanyang Technological University President An Bo convened for a summit dialogue on China's large model industry. The discussion addressed how young people can navigate anxiety amid rapid AI technological iteration. Against the backdrop of accelerating AI development — from Anthropic's latest large model Claude Fable 5 to explosive intelligent agent growth — panelists explored AI self-evolution possibilities, world model future scenarios, and strategies for adapting to an era where technology iteration speed is redefining humanity's cognitive boundaries of intelligence.
AI Creating AI Becomes a Trend
When large models were still refreshing human cognition, a more disruptive trend emerged — AI began creating AI.
"Last year, top large models only completed high-precision execution in scenarios with clear instructions, but now, top models' capabilities have extended to abstract problem-solving levels," Luo Fuli observed. Current large models can complete core scientific research processes including experimental procedure planning and execution result verification, and the core gap with top scientific researchers is narrowing.
Liu Zhiyuan stated that the core of the industrial revolution was machines replacing human physical labor, with the ultimate form being machines manufacturing machines; the core of the intelligent revolution is AI replacing repetitive human mental labor, and "AI creating AI" is the core marker of the intelligent revolution entering an advanced stage.
In Liu Zhiyuan's view, the industrial revolution took hundreds of years to achieve autonomous machine manufacturing, while large models have entered the AI autonomous iteration stage only six to seven years after their emergence, with technological iteration speed far exceeding traditional industrial revolution.
"At the current stage, the core driver of AI recursive self-evolution is still humanity," Liu Zhiyuan stated. The current technical model is human-led, with AI assisting in completing model research and iteration. Even if high-level AI autonomous iteration is achieved in the future, human subjectivity and initiative remain irreplaceable — ultimately, AI's research direction and core goals for serving society always need human definition, which is the core relationship between humans and AI.
Two days prior, Anthropic officially released Claude Fable 5, achieving significant leaps in coding ability and intelligent agent capability — a full repository migration of a 50-million-line codebase that requires one month for a human team takes Fable 5 only one day. This milestone advancement became the first focal point of this discussion.
"Fable 5 is still an intermediate product," Luo Fuli stated. In her view, the path of large model continuous scaling is far from stopping. "Fable 5 represents large model natural expansion in three dimensions: first, pre-training parameter magnitude achieves multiple-fold scaling; second, inference-time scaling and reinforcement learning computing power investment substantially increases; third, training data advances from natural internet text to a new stage of synthetic data jointly produced by humans and intelligent agents."
In the view of Tsinghua University professor and ShengShu Technology founder Zhu Jun, landing requirements differ across scenarios, and not all scenarios require extremely precise model capabilities — most conventional scenarios can land relying on models' intuitive understanding capabilities, which is also the core value brought by large models.
Zhu Jun stated that regarding the industry's heated discussion of Agent (intelligent agents) and code problem-solving consuming large amounts of Tokens (word elements), new version models substantially reduce Token consumption for equivalent tasks, which is the correct direction for industry development.
In his view, the scaling potential of video models and world models remains enormous. "Current accumulation of physical data, data-efficient utilization technology, model architecture optimization and other directions are all just beginning, with extremely large exploration and improvement space in the future."
Innovation Often Goes Against Consensus
As AI technology rapidly iterates, large numbers of young people fall into anxiety — technology updates too fast, knowledge iteration is frequent, traditional skills and professions continuously transform. How to respond?
In Luo Fuli's view, the iteration speed of large models and intelligent agents far exceeds everyone's expectations, and the capability boundaries and division-of-labor models between humans and AI are continuously changing.
"My core advice to young people has only one point: always maintain exploratory desire and curiosity. Extremely utilize cutting-edge AI tools, and in the process of continuous trial and error, cultivate judgment and scientific research aesthetics exclusive to yourself. In an era of rapid technological transformation, unique cognition, judgment, and aesthetics are young people's most core, most irreplaceable competitiveness," Luo Fuli stated.
Zhu Jun believes that AI technology changes daily, all practitioners are continuously learning and continuously iterating, and no one can remain unchanged. The more in the tide of technological transformation, the more one must solidify one's own foundation, which is the core competitiveness for responding to industry changes. "We are committed to creating an AI-native growth environment, allowing students to embrace AI and use AI well from the beginning of learning. Everyone need not be overly anxious — all people are at the same starting line, industry predecessors, practitioners, and students are all synchronously learning and synchronously updating. Actively embracing change and continuously cultivating deeply is the best growth method."
Liu Zhiyuan advised, first, dare to be first and dare to innovate. AGI and intelligent revolution are entirely new unknown territories, and true innovation often "goes against consensus," doing non-consensus things. Daring to attempt in unexplored territories and making differentiated choices can seize future opportunities.
Second, stick to original intentions and persevere. Differentiated innovative choices will inevitably be accompanied by doubt and denial. Whether one can withstand pressure and persist in deep cultivation is key to breaking through bottlenecks and producing results.
Third, break through inherent cognition and continuously self-innovate. After achieving phased results, do not rigidly adhere to past successful paths. Proactively perceive industry trends, negate inherent experience, and explore entirely new directions to continuously keep pace with technological iteration rhythm.
"Current young people's anxiety is essentially caused by an excessively utilitarian mentality. If learning and work goals are only high salaries and following trendy tracks, one will fall into passive anxiety," An Bo stated.
In An Bo's view, first, young people should choose the right track and cultivate core problems deeply, focusing on valuable and meaningful research directions, avoiding ineffective internal competition. Second, the value of academic credentials is weakening. The truly core competitiveness today is practical ability and cutting-edge cognition — even without high academic credentials, deeply cultivating frontline cutting-edge research and accumulating practical experience can achieve rapid growth. Finally, going solo can no longer keep up with industry speed. Proactively seek like-minded partners, form communication communities, and promptly communicate and discuss when encountering problems to avoid self-internal consumption.
FAQ
What did Luo Fuli advise young people at the Beijing Zhiyuan Conference on June 12?
Luo Fuli advised young people to always maintain exploratory desire and curiosity, extremely utilize cutting-edge AI tools, and in the process of continuous trial and error, cultivate judgment and scientific research aesthetics exclusive to themselves. She stated that in an era of rapid technological transformation, unique cognition, judgment, and aesthetics are young people's most core, most irreplaceable competitiveness.
What technical capabilities does Claude Fable 5 demonstrate according to the panel discussion?
According to the discussion, Claude Fable 5 achieved significant leaps in coding ability and intelligent agent capability. A full repository migration of a 50-million-line codebase that requires one month for a human team takes Fable 5 only one day. Luo Fuli described Fable 5 as representing large model natural expansion in three dimensions: pre-training parameter magnitude achieving multiple-fold scaling, inference-time scaling and reinforcement learning computing power investment substantially increasing, and training data advancing from natural internet text to synthetic data jointly produced by humans and intelligent agents.
How did Liu Zhiyuan describe the relationship between humans and AI in the self-evolution process?
Liu Zhiyuan stated that at the current stage, the core driver of AI recursive self-evolution is still humanity, with the current technical model being human-led and AI assisting in completing model research and iteration. He emphasized that even if high-level AI autonomous iteration is achieved in the future, human subjectivity and initiative remain irreplaceable — ultimately, AI's research direction and core goals for serving society always need human definition, which is the core relationship between humans and AI.
