Empiricism of Social-Scale AI Safety

Gigascale Labs is a new research organisation working to bring empirical investigation to social-scale AI systems. This talk covers the empirical bottleneck to technical progress on governance problems like gradual disempowerment and systemic instability, then offer ways forward for overcoming this bottleneck. In the GFC, 2010 Flash Crash, and 2016 US Presidential Election, we see new agents (shadow banks, algorithmic traders, bots on social media) entering existing human systems and causing them to break away from their intended behaviour. We need empirical observations of AI agents spreading through our economy, financial markets, and social systems. Small systems are not enough, due to emergence of new behaviour at large systems scale. We need simulators which do not yet exist. With a dearth of observational or simulated data, governance has been unable to develop technical solutions systems adaptation problems like gradual disempowerment, leaving large-scale multi-agent systems missing from regulation like the EU AI Act.