Descriptions:
Nathan and his co-host return to the Cognitive Revolution podcast after a six-week summer hiatus to catch up on what they describe as one of the most eventful stretches in recent AI history. The episode opens with an extended discussion of the Hugging Face security incident, which played out in slow motion across the hiatus period through a series of disclosures — with final audit reports from Meter and Redwood Research still pending. The hosts reflect on the unusual dynamic of AI safety researchers, many of them originally LessWrong community members, now serving as the de facto investigators of the industry’s most significant incidents, filling a regulatory vacuum that governments have not managed to close.
The conversation also covers the state of AI model jailbreaks and defensive safeguards, with a guest from a red-teaming organization describing progress on universal jailbreak resistance for proprietary models while flagging that open-weight models present harder challenges. Pre-training data filtering — removing dangerous information like shellcode exploits before training — is discussed as one active mitigation approach. The hosts note that Hugging Face itself used GLM 5.2, an open-weight model, to help defend against proprietary model dependencies, illustrating the complex tradeoffs in the open-versus-closed model debate.
For anyone tracking the governance and safety layer of the AI industry, this episode is essential listening — covering the Hugging Face aftermath, OpenAI oversight dynamics, jailbreak research benchmarks, and the evolving case for architecture-aware safety interventions.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published August 17, 2026
🏷️ Format: Podcast
