After a seven-year hiatus, ZHU Songchun made a notable return to the World Artificial Intelligence Conference (WAIC) on July 19, 2026. Previously, he had not participated since 2019, marking his absence as a significant statement against the prevailing trends in the AI field.
In recent years, the spotlight at WAIC has been on large language models, scaling laws, and an optimism that equates massiveness with breakthrough advancements. ZHU, the President of the Beijing Institute for General Artificial Intelligence (BIGAI), has stood out as a prominent critic of this prevailing mindset. Following the surge in interest around large models ignited by the launch of ChatGPT in late 2022, ZHU has consistently asserted that these models do not inherently pave the way to Artificial General Intelligence (AGI). He argues that merely increasing the scale of parameters, data, and computation is insufficient, emphasizing that the financial motivations and international narratives driving this expansion often overshadow authentic technological advancements.
In his recent address during the "Thinkers Forum," ZHU characterized the AI boom as a narrative fueled by capital and geopolitical interests, linked to what he refers to as the "Silicon Valley-Wall Street-Washington triad." He pointed out that the Chinese tech sector has become ensnared in a cycle of reverse informational flows from the U.S., where information is fed back to China and restated by local experts, leading to a repetitive cycle.
ZHU highlighted what he has termed the "Musk Faith," referring to investors' unwavering belief in the ambitious claims of tech leaders, despite a track record of unmet promises. He noted instances where projects championed by Musk, such as the Hyperloop and full self-driving vehicles, have faced persistent delays or outright cancellations, with market reactions often fervently exaggerated in response to these grand pledges.
The tech market worldwide has recently experienced heightened volatility, particularly since June 2026, as major U.S. tech firms continue to escalate their investment in AI while facing diminishing free cash flow. In his address, ZHU reflected on how these market trends may lead to greater receptivity to his critiques as signs of a potential bubble begin to appear.
Notably, the consensus in the tech industry still leans towards larger models, with key players like OpenAI, Google DeepMind, and Anthropic continuing to advocate for their development. They assert that there remains significant potential in refining these architectures and enhancing reasoning capabilities through multimodal integration. Many leaders within China's AI sector believe that while large models may not represent the culmination of AGI, they currently serve as the most effective tools for commercialization.
Throughout the last three years, ZHU's core perspectives have remained steadfast. He has likened the journey towards AGI to a trek up Mount Everest versus a lunar landing, emphasizing the complex differences between achieving large model capabilities and true general intelligence. In his 2025 publication, *Giving Minds to Machines*, he described large models as "brains in vats"—able to produce language but lacking a true understanding of the world, as there exists no genuine connection between their generated words and the real environment.
By 2026, ZHU assessed the field's many booms and busts—covering various eras from computer vision advancements to the metaverse hype—and concluded that the industry's faith in computational power and data-driven models is being systematically challenged by real-world results.
In contrast to mainstream AI trends heavily reliant on transformer technologies, ZHU has initiated research centered around causal reasoning and intrinsic value-driven methodologies, aiming to develop autonomous general agents capable of task generation, learning, and action. His approach has led to the creation of a framework known as CUV, where C represents cognitive architecture, U encompasses capabilities related to perception, cognition, and action, while V symbolizes intrinsic motivations and value systems.
His institution has also rolled out the TongOS general artificial intelligence operating system, the TongPL programming language, and a unique agent named "Tongtong." Unlike typical large model offerings, Tongtong operates on intrinsic values, facilitating an understanding of the environment, generating tasks, and planning actions autonomously. Preliminary evaluations suggest some of its cognitive abilities are akin to those found in five- to six-year-old children.
In March 2026, the latest version of Tongtong was unveiled, showcasing advancements in spatial and cognitive intelligence, as well as social intelligence, alongside the introduction of "Tong Brain," an embodied intelligence engine intended to integrate general agents with physical robotics.
ZHU's team has gained notable attention from platforms like *Science*, detailing their vision for transforming isolated AGI agents into cohesive AGI societies that focus on multi-agent interactions and the development of artificial civilizations. At WAIC, ZHU emphasized the importance of "social intelligence" as the final hurdle toward reaching AGI, underlining the aspect of comprehending complex social dynamics—responsibilities, intentions, and collaborative interactions—in which current large models still fall short.
As the conversation around ZHU’s framework unfolds, the pressing question remains: before Tongtong's commercial viability is established, does his approach reflect a more enlightened technical direction or an ambitious academic hypothesis awaiting validation?
Engaged in thoughtful dialogue at WAIC, ZHU expanded upon his critical views regarding the potential of large models in achieving AGI, addressing the existence of a bubble within the AI sector, assessing China’s innovative initiatives, and discussing how cognitive architectures can transition to practical applications in the industry.
ZHU's stance is clear: while the vibrant atmosphere of innovation at WAIC is commendable, a more profound inquiry into the current technical trajectory is necessary. Despite advancements in robotic technologies, the question persists—will the current methodologies yield the comprehensive general intelligence that stakeholders envision?
