An OpenAI board member wants AI labs to pause progress "right now."
Paul Christiano, who invented RLHF, the technique behind the ChatGPTs training, said on the Devarkash podcast on Monday that "it's not good for US AI labs to improve pre-training and reinforcement learning given their current security posture. ".
Should the labs be put on hold? "I think we should. " On improving algorithms that can be stolen quickly: "I think they should stop it now. ".
He fears that China could steal super-intelligence in the middle of the race, so labs couldn't slow down even if their models went off the rails. Host Dwarkash Patel noted one security startup entered every Anthropics Slack within 12 hours.
"But you're on the OpenAIs board," Patel said. Cristiano: "It is very important in this conversation that I will not speak as a board member of OpenAI. ".
He gives maybe a 1 in 3 chance of superintelligence within a year if nothing slows it down, and puts the risk of AI takeover in that time at "a 10% risk rather than a 1% risk." ".