Descriptions:
The Cognitive Revolution hosts Nathan and Pash spend a live morning session working through Anthropic’s major new J-Space interpretability paper — a 150-page research release accompanied by 50 pages of commentary, interactive demos, and supporting summaries. The hosts frame it as Anthropic’s most significant mechanistic interpretability work since their earlier “Tracing the Thoughts of a Large Language Model” paper, and attempt to make sense of it in real time without an expert guest, using Claude itself to help parse difficult sections.
Key discussion threads include Anthropic’s evolving stance on anthropomorphizing AI models, the discovery of exotic internal activation structures within transformer architectures, and the question of how interpretability findings transfer across model families. The hosts reference prior work on Mamba architectures, noting that similar representation structures emerged there as well — lending cautious optimism that interpretability techniques are more portable than feared. They also raise the concern that a major architectural shift driven by competitive pressure could destabilize current monitoring and interpretability findings.
The episode is an accessible but technically substantive exploration of what J-Space structures reveal about how Claude processes and routes information between layers. It’s particularly valuable for listeners tracking AI safety and mechanistic interpretability research, offering a candid discussion of caveats, residuals, and the limits of what even a 150-page paper can definitively establish about a model’s internal computations.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published July 07, 2026
🏷️ Format: Podcast
