Arcadia Impact shares 4-month blueprint for building an AI safety research team

Arcadia Impact, an established non-profit based in London, successfully assembled an 8-person AI alignment research team over four months, funded primarily by the UK AISI’s Alignment Project. The team collaborates closely with AISI’s alignment team and is pursuing three core projects: generating comprehensive behavioral documents for models, evaluating alignment training techniques, and validating debate protocols for scalable oversight. They are also building pipelines for automated alignment research. Led by Andrew Draganov (PhD in ML, postdoc background) and Erin Robertson, the team leveraged existing relationships from programs like LASR and ASET to accelerate hiring.

Key lessons from the experience focus on hiring strategies. Draganov advises that the highest-signal components were work tests and references, which outperformed interviews for assessing research ability. He also notes that having an established organizational infrastructure (visa sponsorship, office space) was a major advantage, but fiscal sponsors can fill that gap for newcomers. The team’s success suggests that being a well-known name is not a prerequisite to lead a research team; systematic evaluation and network building matter more. The post serves as a practical template for aspiring AI safety entrepreneurs, emphasizing that funding (e.g., from AISI or Coefficient Giving) and a structured hiring process can enable rapid team formation.