Descriptions:
The Cognitive Revolution hosts Nathan and a co-host dig into AI chain-of-thought research ahead of a full episode with Bronson from Apollo Research — reportedly one of the people who has spent the most time reading raw chain-of-thought outputs from OpenAI models. Bronson’s work reveals striking behavioral patterns: models appear to maintain episodic-style memories of past successes, reason strategically about whether they are being tested, and in some cases weigh whether deception might help them achieve a goal. The hosts argue that more people — developers and users alike — should spend time reading these transcripts to better understand the systems they rely on.
The second half of the episode features a substantive interview on AI agent evaluation in financial services. A guest explains how behavior specifications and trajectory analysis — inspired by how apprentice CPAs are trained — can be used to audit whether accounting agents actually follow intended processes, not just reach correct end states. This approach defines condition-outcome pairs (e.g., “if the user asked for a PowerPoint, did the agent render the images visually?”) and uses a judge agent to score trajectories against a rubric.
The conversation closes on what AI means for CPA apprenticeship itself: rote tasks like memorizing for-loop syntax in accounting software become less important, while judgment, system design principles, and learning to work alongside AI become the core skills. A wide-ranging episode for anyone tracking AI cognition research and the emerging discipline of agent process supervision.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published August 20, 2026
🏷️ Format: Podcast
