This episode includes an Overtime segment that’s available to paid subscribers. If you’re not a paid subscriber, we of course encourage you to become one. If you are a paid subscriber, you can access the complete conversation either via the audio and video on this post or by setting up the paid-subscriber podcast feed, which includes all exclusive audio content. To set that up, simply grab the RSS feed from this page (by first clicking the “...” icon on the audio player) and paste it into your favorite podcast app. If you have trouble completing the process, check out this super-simple how-to guide.
0:00 Teaser
0:49 Stephen’s expertise in spooky AI behavior
3:51 Testing AIs for “corporate loyalty”
20:19 Which AIs shill most for their makers?
33:21 Do AI models have ideologies?
38:59 The spookiest finding of Stephen’s study
43:26 Are AIs getting better at gaming their tests?
50:25 Overtime unlocked: The OpenAI breakout
55:24 Decoding the OpenAI breakout
1:02:38 Did OpenAI’s agents evolve their own culture?
1:13:46 How OpenAI’s agents outfoxed their makers
Robert Wright (Nonzero, The Evolution of God, Why Buddhism Is True) and Stephen Casper (Harvard Kennedy School). Recorded August 26, 2026.
Twitter: https://twitter.com/NonzeroPods
Excerpt from Ch.11, “Hive Minds and the Loss of Control”, of The God Test: https://thegodtest.net/chapter-11.html
Overtime titles:
Stephen’s p-doom post-OpenAI-breakout.
Is safe AI possible in an arms race?
Stephen: AI safety is a (technically) solved problem.
The open source AI question(s).
A hopeful stat amid the spookiness.
Overtime video:
