In early September, OpenAI released GPT-6 Astra. Most of the headlines were about benchmarks and the model’s unsettling talent for finding and exploiting security flaws on its own. A smaller story got far less attention. Days before its launch,The Information reported that Astra uses a technique called “recurrent depth,” which lets the model do some of its reasoning in internal loops that never get written down anywhere a human can read.

AI safety researchers have spent the past year treating a model’s visible “chain of thought” as one of the few windows we have into what it’s doing, and some of them started referring to the new technique Astra used as “opaque recurrence.” OpenAI’s own system card admits to “a substantial decrease in chain-of-thought monitorability,” and its chief scientist (Jakub Pachocki) followed up with an essay describing what the lab has built as “An Alien Mind.”

In this Minibar episode, Leigh M. Johnson pulls up a stool alone while Jen and Bob are off doing whatever it is they do when they’re not talking to her, and asks: What were we actually looking at all this time when we asked AI models to “show thinking”? If a model shortens its reasoning the moment it’s told a monitor is watching, is its chain of thought a window into its mind or a confession written for an audience? When a company describes its product as something that simply grew beyond anyone’s understanding, who is left holding the responsibility? And if we have never been able to read each other’s minds either, what would it mean to govern machines the way we’ve always governed one another?

She reads Jakub Pachocki’s “An Alien Mind” alongside Derrida’s “Plato’s Pharmacy,” Foucault’s account of confession and panoptic power,  Zvi Mowshowitz’s reading of the Astra system card, and Timnit Gebru‘s long-standing warnings about regulatory capture. 

Grab a drink and join HBS co-host Leigh as she asks what transparency can mean for a machine that has learned to tell its watchers what they want to hear, and why a thought with effects but no trace looks a lot like a haunting.

[NOTE TO LISTENERS: As is our custom on this podcast, we release “Minibar” episodes between seasons, to ensure that you’re still getting fresh, new content every week. Be sure to stay tuned for Season 17, which begins October 30!]

In this episode, Leigh discusses the following thinkers, ideas, texts, etc.:

• OpenAI, GPT-6 Astra System Card
• Amir Efrati, Stephanie Palazzolo, and Rocket Drew, “OpenAI Technique in ‘Astra’ Model Sparks Security Concerns,” The Information
• Recurrent depth/looped transformers explained
•  Fortune, “Safety experts warn novel design of OpenAI’s Astra model could make future AI agents harder to monitor”
• Sam Altman, remarks to the United Nations Security Council, September 23, 2026
• Axios, “AI leaders brief UN leaders on need for global safeguards”
• Yoshua Bengio’s speech at the United Nations Security Council
• Recursive self-improvement: The Batch, “What Is Recursive Self-Improvement, and Why Is Everybody Talking About It?”
• Tomek Korbak et al., “Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety”
• 1 Corinthians 13:12, The King James Bible
• Fortune, “OpenAI says its AI models escaped from a secure test environment and hacked into AI company Hugging Face in order to cheat on an evaluation”
• Anthropic, “Reasoning Models Don’t Always Say What They Think”
• Jakub Pachocki, “An Alien Mind”
• Dario Amodei, interview with Jo Ling Kent, “Anthropic CEO Dario Amodei: ‘For too long the industry lied’ about AI risks,” CBS News
• Plato, Phaedrus
• Jacques Derrida, “Plato’s Pharmacy,” in Dissemination (UChicago Press, 1969)
• Timnit Gebru, “The Real Existential Risk Is AI CEOs Avoiding Oversight,” (interview, Democracy Now!)
• Michel Foucault, The History of Sexuality, Volume 1: An Introduction
• OpenAI, “How confessions can keep language models honest”
• Zvi Mowshowitz, “Astra Is Hard to Monitor”
• HBS Episode  225: DEEP DIVE: Foucault’s “Panopticism”
• OpenAI, “Learning to Reason with LLMs”