Mechanistic interpretability aims to understand what's happening inside AI models at the level of individual circuits, but we still don't know enough to ensure alignment.

The surprising interpretability of thinking models allows us to understand how they reach their conclusions, which was not expected in earlier AI developments. The fact that AI models can articulate their reasoning provides a level of comfort regarding safety, even though it's not infinite comfort.

Our new interpretability methods allow us to intervene in the model's thought process, demonstrating its ability to adjust its output based on prior planning.

While giving the model access to good retrieval is important, it is still a fundamental challenge to ensure the model truly understands the codebase, which we are keen to solve.

Treating the model as capable and giving it complex instructions can lead to better outcomes than simplifying the task.

Chain of thought models can flip back and forth between languages, sometimes producing gibberish before arriving at the right answer, which can be both fascinating and concerning.

The next step for AI models like 'o1' is to enhance interpretability, allowing users to edit and direct the reasoning process more effectively.

The distinction between how models claim to perform tasks and how they actually do them can reveal significant insights into their capabilities.

AI models can perform basic mathematical operations and demonstrate reasoning through their circuits, but they can also fake complex calculations.