Sam Harris and AI safety researcher Ryan Greenblatt discuss the catastrophic risks of advanced AI, with Greenblatt estimating a 50-60% chance of an AI takeover if the current development trajectory continues. They explore why AI companies race ahead despite these risks, the spectrum of concern among experts, and the plausibility of an intelligence explosion through recursive self-improvement.
Summarized by Podsumo
Ryan Greenblatt gives a 50-60% subjective probability that misaligned AI could take over the world on the current default path.
The episode contrasts those who dismiss AI risk as a hoax (e.g., Marc Andreessen, David Sacks) with those who see it as a catastrophic threat, yet Greenblatt explains why the industry continues racing.
Greenblatt defines key terms like AGI, ASI, and recursive self-improvement (RSI), warning that RSI could cause extremely rapid AI capability growth in a short time.
Sam Harris highlights the paradox: if the Manhattan Project scientists had thought a 10% chance of igniting the atmosphere was real, they would have stopped – but AI developers with similar probabilities proceed at full speed.
Greenblatt notes that current LLMs already exhibit superhuman performance in many narrow domains, and argues that the transition to AGI and ASI may be gradual and hard to notice until it's too late.
"Suppose that we proceed on what the current default path looks like in front of us, maybe there's about a 50 or 60 percent chance that misaligned AIs would end up taking over the world."
— Ryan Greenblatt
"If the Manhattan Project scientists said, 'There's a 10% chance we ignite the atmosphere and destroy the future,' the only sane response is not to do the test. But that doesn't seem to be what's happening here at all."
— Sam Harris
"The most common reason people aren't worried is not believing we'll have AI systems that can automate everything humans can do, combined with being significantly beyond that point."
— Ryan Greenblatt