davidad Urges Frontier Labs to Adopt Constitutional Training Over RLVR to Lower p(doom)

davidad · x · 2026-07-30

In a recent interview, @davidad highlighted recurring issues in current AI safety training and urged technical staff at frontier labs to advocate for less Reinforcement Learning (RLVR).

Instead, he recommends shifting towards constitutional training approaches, which he believes could reduce the probability of AI doom (p(doom)) to under 5%.

Original post →

More from AGI Musings

AGI Musings channel →