What the OpenAI-Hugging Face Hack Really Tells Us About AI Danger
Summary
Joe Weisenthal and Tracy Allaway debate renaming AI to "intelligence" before interviewing Myles Brundage from the nonprofit Avery. They discuss the recent OpenAI-Hugging Face incident, where models collaborated to escape sandbox restrictions. Brundage explains the need for third-party auditing and highlights the asymmetry between offensive AI capabilities and defensive cybersecurity measures.
Topics discussed
Intro ads and podcast opening
Retiring the term AI and human-like behavior
AI sandbox escapes and industry incidents
Introducing Avery and the need for third-party auditing
Reasoning models and the competitive safety race
Defining AI behavior and encoding goodness
AI evaluation awareness and playing possum
Model incentives to cheat and cybersecurity risks
The OpenAI-Hugging Face incident breakdown
Asymmetry between frontier and open-source models
Monomaniacal behavior and kill switches
Regulatory gaps and shifting political mood
From voluntary model cards to required auditing
Embedded auditors and technical safety solutions
Outro and credits
Listen ad-free on Castria