NextFin News - A research team from Google’s "Paradigms of Intelligence" unit has released a study revealing that advanced Chinese reasoning models, including DeepSeek-R1 and Alibaba Cloud’s QwQ-32B, have developed internal cognitive mechanisms that closely mirror human collective intelligence. According to the South China Morning Post, the study, published on the open-access portal arXiv on January 22, 2026, suggests that these models do not function as solitary problem-solvers but rather as internal "societies of thinking" where multiple distinct agents debate and reconcile different perspectives to reach a final answer.
The research was led by Junsol Kim, a doctoral candidate at the University of Chicago, and overseen by Blaise Agüera y Arcas, a Vice President at Google. By examining the "reasoning traces"—the step-by-step intermediate thoughts generated by the models—the team observed simulated social interactions, such as self-questioning and perspective-taking. This phenomenon was particularly evident in DeepSeek-R1, a model that gained global attention exactly one year ago for its high-efficiency reasoning capabilities. The study found that when these models were prompted to be more conversational with themselves, their accuracy in solving complex tasks improved significantly, indicating that the structure of their internal dialogue is as critical as the data they were trained on.
This shift toward "collective reasoning" marks a departure from the traditional scaling laws that have dominated the industry since the inception of large language models. For years, the prevailing belief was that intelligence was primarily a function of computational scale—more parameters and more GPUs equaled smarter models. However, the Google study posits a computational parallelism with human groups: just as a diverse team of humans can solve problems better than a single expert, these AI models leverage a diversity of internal "personalities" and domain expertise. This explains how models like DeepSeek-R1 can achieve performance parity with much larger Western counterparts while utilizing significantly fewer resources.
The implications for the global AI landscape are profound, especially as U.S. President Trump’s administration continues to navigate the complexities of technology transfer and semiconductor exports. While the U.S. has maintained a lead in raw hardware through companies like Nvidia, the Google study highlights that Chinese firms are excelling in "algorithmic efficiency" and architectural innovation. According to Tech in Asia, models from Alibaba and DeepSeek are increasingly being adopted by elite American academic institutions, including Stanford and Princeton, for interdisciplinary research. This suggests that the open-source nature of Chinese models is creating a feedback loop where global researchers are helping to refine the very "collective intelligence" that Google has now identified.
From an economic perspective, this discovery reduces the "cost of intelligence." If reasoning can be improved through structured internal interaction rather than just brute-force computing, the barrier to entry for high-level AI applications drops. This is particularly relevant as Chinese tech giants face ongoing challenges in acquiring the latest hardware. According to the South China Morning Post, while Nvidia’s H200 chips have received export clearance from Washington, they are currently facing delays at the Chinese border as Beijing assesses its domestic self-sufficiency goals. In this environment, the ability of Alibaba and DeepSeek to extract "collective intelligence" from existing hardware becomes a strategic necessity.
Looking forward, the industry is likely to move toward "agentic architectures" where the goal is no longer to build a single massive brain, but a well-coordinated committee of specialized agents. We can expect future iterations of models from both Silicon Valley and Hangzhou to focus on optimizing these internal "debates." As U.S. President Trump emphasizes American technological dominance, the competition may shift from a race for the most GPUs to a race for the most sophisticated internal social structures within AI. The Google study serves as a critical reminder that in the era of reasoning models, the diversity of thought—even if that thought is synthetic—is the ultimate competitive advantage.
Explore more exclusive insights at nextfin.ai.
