Interview, Podcast
Navigating the geopolitics of US–China AI competition | Sihao Huang
- AI systems are projected to reach general human capability across nearly all domains by 2040, accelerating scientific innovation and enabling vast-scale AI research, while the deployment process itself is anticipated to occur "really, really fast" once initiated.
- Geopolitical risks include a scenario where China unilaterally deploys advanced AI without global consultation, potentially triggering a nuclear attack to prevent permanent US dominance if the US achieves a decisive lead.
- China is currently estimated to be 1.5 to 2 years behind US frontier models as of mid-2022 but is expected to reach GPT-4 level capabilities within the current paradigm, maintaining a competitive edge in computer vision and surveillance.
- Chinese engineers face anxiety regarding the US lead and may prioritize catching up by accessing open-source models or stealing weights, though domestic development is constrained by a limited stockpile of imported chips and struggles to achieve high yields in 7-nanometer production in the short term.
- The US may slow AI development to regulate systems effectively if strict export controls create a dominant position that removes the immediate pressure from Chinese competitors, whereas a US failure in the scaling paradigm could allow China to pull ahead through alternative paths like deep economic integration or neuromorphic computing.
- Technological timelines suggest China could indigenize extreme ultraviolet lithography or advanced semiconductor processes by 2030 or later, though full ecosystem indigenization would incur significant economic costs and require five years or more.
- Future risks involve unreliable autonomous weapons with response times of 100 milliseconds to one second, which could force reciprocal deployment to retain battlefield advantage and lead to a compromise on safety standards.
- Chinese AI regulations are expected to remain focused on "information control" and preventing violations of CCP values rather than broader AI safety, with the upcoming AI law likely to pass unless it hinders development.
- The speaker anticipates a transition where international AI cooperation ends once technologies tie to core national interests, potentially leading to volatile US-China relations and a breakdown of the current "honeymoon period" in dialogue.
- Structural societal shifts include the potential for AI to reduce the need for democratic systems by automating intellectual labor, shrinking the "selectorate" to zero through automated repression, and removing incentives for current beneficiaries via automation.
- China's AI safety community is viewed as less mature regarding catastrophic long-term harms compared to the US and UK, with alignment efforts currently geared toward traditional work, while extreme risk narratives may be more prevalent in the Chinese "Overton window."
- A global deliberative process is considered necessary by 2040 to prevent any single nation from unilaterally triggering an intelligence explosion, though the speaker notes that the US and UK are more likely to achieve advanced AI first by betting on the scaling paradigm.
- Short-term constraints on China's AI scaling include a macroeconomic environment limiting investment and the finite availability of pre-2023 export-controlled chips, while long-term semiconductor policy will rely on a complex mix of national directives and local incentives.
- China may not require frontier models to generate high-risk systems such as biological design tools or cyber agents, and there is a risk of catastrophic harm if capable systems are released as open source without comprehensive regulation.
- The speaker expects US developers to face an "AI winter" if scaling results are disappointing, whereas Chinese entities might exploit alternative research paths or economic integration to maintain a competitive trajectory during such a period.