Long list of interventions

Abstract

A public Notion page maintained by BlueDot Impact that presents a non-exhaustive list of example interventions to help orient learners on how to make AI development go well. The interventions are grouped into four sections: preventing the training of dangerous AI systems (a global AI treaty and verification methods, compute governance and licensing for large training runs, IAEA-style international monitoring of frontier labs, algorithmic secrecy, and non-agentic ``Scientist AI’’); constraining dangerous AI capabilities (input-data filtering, alignment techniques such as constitutional AI and RLHF, interpretability, AI control, adversarial red-teaming, and model-weight security); withstanding dangerous AI actions in biosecurity (pandemic prevention, surveillance and containment, vaccines and PPE stockpiles) and cybersecurity (memory-safe refactoring, government red-teaming of critical national infrastructure); and meta-level work (public education, mobilising citizens, and training decision-makers).