Dan Balsam and Tom McGrath, co-founders of Goodfire, delve into mechanistic interpretability and AI safety. They discuss breakthrough techniques like sparse autoencoders and the nuances of token prediction models. The conversation highlights the importance of interpreting AI to mitigate risks and foster understanding. Balsam and McGrath share personal journeys from skepticism to active research and address engineering challenges, advocating for transparency in AI systems while exploring the societal implications of interpretability.