Events, programs, and opportunities from MAIA. Join mailing list
Control & Scalable Oversight
← All weeks
This week focuses on how to supervise models that are much smarter than their overseers. We begin with an introduction to scalable oversight, which studies how weaker overseers can reliably evaluate the work of stronger models. Next, we look at how training against a chain-of-thought monitor can improve behavior at first but eventually teaches the model to hide its reasoning, and read the case for why chain of thought monitoring may be crucial for safety. Finally, we turn to the AI control agenda, which seeks to prevent harm in worst-case worlds where AIs turn out to be scheming.