Chinese AI – They're Just Like Us? With Beijing-Based Concordia AI CEO Brian Tse

Episode Summary

Executive Summary: The episode presents Brian Tsay’s view that China’s AI ecosystem is broadly more similar than different from the West’s, but with stronger state coordination, more concrete safety regulation, less AGI rhetoric, and a more application- and governance-oriented national strategy. Tsay argues China is serious about AI safety, open to international dialogue, and often more transparent than Western observers assume, while warning that U.S.-China mistrust and export controls could worsen global risk and reduce opportunities for cooperation.

Main Topics: China’s AI ecosystem resembles the West more than commonly assumed (Priority: 5/5): Tsay describes China and the U.S. as pursuing similar frontier AI architectures, scaling methods, and research agendas, with both ecosystems shaped by mission-driven startups, large tech firms, and emerging safety institutes. China’s AI governance and safety regime (Priority: 5/5): The discussion details China’s pre-deployment testing, registration requirements, national standards, labeling rules, and top-level policy signaling around AI safety as a major public and national security concern. Differences in AI culture, vision, and AGI discourse (Priority: 4/5): Tsay says Chinese AI actors focus more on practical applications, self-sufficiency, and economic utility than AGI timelines or superintelligence narratives, and that official planning emphasizes AI as an economic engine rather than an existential transition. Open source, English publication, and global scientific integration (Priority: 4/5): The conversation highlights how Chinese researchers publish in English, participate in international conferences, and share open-weight models, suggesting deep integration with the global AI research community despite geopolitical tensions. U.S.-China mistrust, export controls, and chip competition (Priority: 5/5): Nathan and Tsay debate whether China’s chip choices, U.S. export controls, and fears of secret capabilities justify the race-to-AGI framing. Tsay argues export controls fuel resentment and that China is investing in domestic alternatives and energy-based scale strategies. Track-two diplomacy and international AI coordination (Priority: 4/5): Tsay advocates dialogues, shared red lines, common evaluations, and emergency protocols to reduce escalation risk. Singapore is framed as a neutral convening hub for U.S.-China and broader global AI governance. Embodied AI, robotics, and labor impacts (Priority: 3/5): A notable China-specific emphasis is on embodied AI and humanoid robotics, reflecting China’s manufacturing strength and raising both safety and labor-displacement concerns.

Key Arguments: China’s frontier AI development is structurally similar to the U.S., with both ecosystems focused on scaling foundation models, multimodality, reasoning, agents, and open research publication. China’s AI safety regime is more formalized than the U.S. at the consumer-facing deployment stage, including pre-registration, pre-deployment testing, labeling, and regulator access to model testing. Chinese AI policy is less driven by AGI timelines and more by practical integration into science, industry, government, and public services. Open-source release and English-language publication are not anomalies but part of a long-running, deeply integrated global scientific ecosystem. Chinese policymakers and academics appear willing to treat AI safety as a national priority, with continuity and coherence stronger than in the U.S. system. Export controls are viewed in China as attempts to suppress development and may undermine trust and cooperation on shared safety risks. Chinese companies may compensate for weaker chip efficiency by using more energy and larger-scale clusters, leaning on China’s energy infrastructure and domestic hardware development. International red lines, shared evaluations, and emergency protocols are the most plausible path to stability; a deterrence-style ‘mutually assured AI malfunction’ model is too vague and destabilizing. Embodied AI/robotics is emerging as a major Chinese safety concern and a likely future area of policy and technical research. There is real concern in China about job displacement and social anxiety from AI, even if public optimism remains relatively high.

Data Points: Chinese AI safety groups identified: around 31 groups - Tsay says Concordia’s report documents about 31 Chinese groups publishing multiple AI safety papers. Top cities for Chinese AI safety activity: Beijing, Shanghai, Hangzhou, Shenzhen, Hong Kong - He lists these as the main hubs for AI safety research in China. Public optimism about AI: about 80% - A survey of Chinese students found roughly 80% agreed AI would do more good than harm. Per capita GDP increase: more than 140x - Tsay cites his parents’ lifetime experience as evidence of China’s rapid development and optimism. Extreme poverty rate decline: from approximately 88% to close to zero - Used to explain why many Chinese respondents are optimistic about technology’s benefits. Registered AI systems in China: over 500 - Tsay says over 500 systems and model versions have been filed in the Chinese registration process over the last two years. Safety evaluation threshold example: 96% acceptable output rate - He gives an example of a Chinese regulatory test requiring the model to be acceptable 96% of the time. Chinese AI safety groups publishing frontier-risk work: more than 30 - By 2025, he says more than 30 research groups in China are substantially working on frontier AI safety. DeepSeek biological refusal rate change: 11% to 54% - Concordia’s benchmark reportedly found refusal on harmful biological prompts rose from DeepSeek V3 to V3.1. DeepSeek V3.1 refusal rate relative standing: around the median - Tsay says the current refusal rate is roughly median compared with other models. Chinese power grid expansion: equal to the entire U.S. grid in the last decade - Used to argue China can lean more on energy abundance to offset less efficient chips. Huawei CloudMatrix example: uses more than five times as many chips - Illustrates China’s possible scale-over-efficiency approach to compute.

Pivotal Quotes: "China and the West are not opposing forces, but similarly structured communities pursuing remarkably familiar goals." — Nathan E. B. / podcast intro: Opening framing of the interview, arguing against simplistic rivalry narratives. "If the braking system isn't under control, you can't really step on the accelerator with confidence." — Brian Tsay: Tsay explains China’s view that safety and capability development should proceed together. "I do think the discourse in China is a bit more in the middle." — Brian Tsay: On the Chinese AI safety debate being less polarized than the U.S. spectrum from pause to accelerationism.

Implications: The episode argues against simplistic U.S.-China AI race narratives and suggests the biggest opportunities lie in shared safety standards, joint evaluation, and dialogue. For industry, it implies China is neither absent-mindedly reckless nor secretly isolated, but increasingly governed, practical, and safety-aware.