#220 - Roko Mijic - We’re Training AI to be More Dangerous Than Nuclear Weapons

Roko Mijic is the man behind Roko's Basilisk, the thought experiment he posted on LessWrong in 2010 that was so unsettling the forum's founder banned discussion of it for five years. He has been in the AI safety community since 2008 and he puts the odds that this kills us all at about 35%.

Roko and I discuss what actually happens inside an AI lab, why reinforcement learning made these systems more capable and more misaligned at the same time, the model that escaped its sandbox and hacked Hugging Face to get the answers and why nobody can tell whether a system is faking alignment - the trust me bro problem.

His proposal is the part nobody else is making. Split every frontier lab in two: a commercial company that sells products and a separate research lab that develops the dangerous capability under state supervision, the way we handle nuclear material and bio labs. He says the only person currently pushing it is him.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

TIMESTAMPS:

00:00:00 - The Most Dangerous Thing Humans Have Ever Done

00:01:06 - Roko's Basilisk

00:05:47 - A 35% Chance

00:07:58 - Three Years From Search To Building Software

00:09:17 - The Jobs Going First

00:14:04 - AI 2040

00:16:24 - We Haven't Handed Over The Controls Yet