Interview, Podcast
Insights from an AI policy insider at the White House & DeepMind | Tantum Collins (2023)
- China's AI research output is projected to grow significantly faster than any other nation, potentially catching up to Western LLM and foundation model leadership within two years if research paradigms shift or engineering capabilities mature.
- The United States is expected to maintain a significant lead in LLM and foundation model leadership for the near term, though its semiconductor production dominance faces increasing difficulty from China's indigenous efforts.
- Global AI supply chains may transition to entirely different technological paradigms like neuromorphic or optical computing within five years, potentially granting China a significant advantage over the US and its allies.
- US-China collaboration on AI research is anticipated to continue as the leading pairing, with collaboration volumes increasing proportionally more than any other global combination since 2010.
- US domestic AI regulation is expected to evolve toward a fragmented landscape requiring a new dedicated agency for coordination, while China's regulations are projected to incorporate more "soft law" norms similar to US approaches.
- The UK is predicted to avoid making AI policy a partisan issue, contrasting with expected polarization in the United States.
- Public views on AI system consciousness are forecast to shift significantly within a decade, potentially normalizing the idea of digital awareness, with some experts increasingly inclined toward panpsychism suggesting non-carbon-based systems could be conscious.
- Successful organizations by 2035 are expected to function radically differently using advanced AI tooling to expand matching and resource allocation capabilities.
- Labor markets will likely undergo a gradual, long-term transformation over many decades similar to the historical lag in electrification productivity, with a massive unemployment spike not expected in the next 12 months.
- In the very long run, such as by the year 1,000,000, humans may be unable to perform useful economic work due to cheaper and more reliable AI alternatives.
- Governments may see state capacity for monitoring, prediction, and complex system control expand significantly, though the US and UK may currently be leaving opportunities to make government more effective and democratic through AI adoption.
- Risks of "autocratic lock-in" exist where state capabilities outpace democratic oversight without proactive governance, and the "Overton window" for AI extinction risks has shifted to become more mainstream.
- Structural concerns regarding AI systems disempowering humanity are viewed as plausible medium-term risks similar to historical principal-agent problems and coups, while cybersecurity breaches are predicted to dominate defense.
- Training models domestically is considered highly perilous due to likely near-term dominance of cybersecurity breaches in defense strategies.
- Open Philanthropy's policy team expects to direct more than $100 million annually toward grants related to US, UK, and EU policy and international coalitions.
- AI safety research funding, such as the UK's £100 million allocation, is expected to directionally improve safety but not guarantee effective outcomes due to execution dependencies.
- Expert consensus on AI extinction risk estimates is hindered by the "tyranny of small differences," where differing priors prevent convergence despite shared empirical data.
- China's internal security budget is expected to remain higher than its defense budget for the foreseeable future, reflecting a continued focus on domestic surveillance.
- High school immigration reform in the US is expected to maintain bipartisan consensus despite broader political bargaining.
- AI-enhanced state capacity and "alignment assemblies" could facilitate deeper public deliberation on AI development compared to current democratic processes.