If only about 1000 FTEs are working towards safe AI, while over 120,000 work towards frontier R&D (and over 1 million support the data and training pipeline), what are the highest-leverage interventions that correct this imbalance? Or phrased differently, with safety outnumbered 100-to-1, where would marginal resources and policy interventions generate disproportionately high leverage?
(mandatory percentage-of-compute allocated to safety per training run? making AI developers or shareholders liable for harms?criminal liability for AI developers?)
What, when, and to what degree will physical infrastructure limits slow AI development?
(e.g. power gen, grid, cooling, semiconductor fabrication, network)
To what degree does “creativity” or “research taste” require human “genius”, or is it more-so a question of being exposed to the right information and drawing the right connections? If the former then the extent to which current architectures will advance the scientific frontier is to some degree limited; if the latter then things will look very different.
What are the best metaphors and analogies to use to quickly provide people with an intuitive understanding of key AI concepts?
What role can legal regulators play in the evaluation and oversight of AI systems?
How can the law respond to the rise of synthetic content and ease with which audio-visual material is generated or edited by AI?
How can democratic institutions best adapt to the diffusion of advanced AI systems? Do we need new institutions? What would they look like?
What political/regulatory/practical bottlenecks exist in middle-power jurisdictions to the building of new data centres?
