This week on GAEA Talks, Graeme Scott sits down with Connor Leahy, US Executive Director of Control AI. Connor is one of the most technically credible and politically clear voices in the world on the risks of superintelligence.

Connor has been in AI his whole career. He co-founded EleutherAI, which built some of the first large-scale open-source language models. He then founded Conjecture, a research company focused on controllable and understandable AI systems, before winding it down to move to Washington, D.C. and lead the US work of Control AI.

Filmed in the new London studio, this conversation cuts through more of the current AI narrative in ninety minutes than most policy papers do in a hundred pages. Connor's argument is that superintelligence is not a technical problem, it is a political one. He walks through why modern AI is grown rather than written, why reinforcement learning produces optimising sociopaths by default, why the labs are selling a political product not an economic one, and why "there are no adults in the room" is the single most important thing every leader must understand.

Topics covered:

Why superintelligence is a political problem, not a technical one
Neural networks in plain English - grown, not written
Why we only understand about 3% of what happens inside modern AI
Why reinforcement learning produces optimising sociopaths by default
Why the labs are selling a political product, not an economic one
The leaded gasoline parallel for AI regulation
"Saliency control" as the primary method of mind control
Connor's three prescriptions - write things down, find community, engage policy