Interview
The Geopolitics of AGI | Helen Toner (Director of CSET & past OpenAI board member)
- The US and China are projected to remain in a state of ongoing competition rather than a race with a definitive finish, with the US currently holding a capability lead estimated at six to twelve months, though this gap may fluctuate between three and nine months depending on development cycles.
- China faces significant short-term challenges in manufacturing required chips and developing competitive supercomputing capabilities within the next couple of years, but could potentially gain a long-term advantage if US policies deter high-skilled immigrants and if the US rejects high-skilled scientists and engineers.
- The US may eventually need to shift from selling chips to "renting" compute to maintain security, provided it can monitor and cut off access, a strategy that risks alienating allies and pushing them toward alternative providers like China.
- Future progress in AI is expected to follow a "drip drip drip" pattern of model releases and a steady upward climb in sophistication rather than sudden massive leaps, with the power needed to recreate frontier systems falling rapidly while the power required to build them continues to rise.
- The US government may struggle to achieve a full or permanent solution to AI risks, instead navigating a dynamic state of "muddling through" with partial measures while balancing tensions between power concentration and diffusion.
- Capital intensity in AI is predicted to naturally concentrate power in well-funded companies and governments unless regulatory or tax interventions are implemented, with the US potentially needing to use antitrust or taxation to diffuse this trend.
- The military faces institutional barriers, procurement difficulties, and a lack of "trilingual leaders" that will likely slow the integration of AI, while the sector must also defend against systems being backdoored or possessing "secret loyalties."
- The UAE and similar Gulf states may acquire world-class supercomputing power through data center deals, raising concerns about empowering autocratic regimes, though US attitudes toward China may soften driven by commercial engagement interests.
- Non-proliferation of frontier AI is expected to be difficult to sustain as the cost and complexity of building such systems drop rapidly after the first creation, necessitating a reliance on transparency and disclosure rather than binary outcomes.
- The US nonprofit structure at OpenAI faces potential loss of control if reorganization fails to maintain legal obligations to the nonprofit mission, and the US may need to develop an "adaptation buffer" against bio-threats and cyber attacks before such capabilities become widely accessible.