Inside Goodfire: Building safer AI systems with interpretability

​How can we build safer AI systems with interpretability? ​Join Goodfire and BlueDot Impact for a look at how Goodfire is working to make AI models something you can understand, debug and deliberately design. ​Most of the AI industry treats models as opaque systems — training them at scale and evaluating them by their outputs. Goodfire is taking a different approach. By studying how models organise knowledge internally, they're building tools that let developers inspect and steer what AI models learn. ​Founded in mid-2024, Goodfire has brought together researchers from interpretability teams at Google DeepMind and OpenAI, raised $207M from Menlo Ventures, Lightspeed, and Anthropic, and shipped Ember — a hosted API that lets developers work with a model's internal features directly. ​We'll be having a conversation with Dan Balsam, co-founder & CTO of Goodfire. ​He leads Goodfire's engineering and technical direction and is also a graduate of BlueDot's AI Alignment course. You can learn more about Goodfire's roles: https://www.goodfire.ai/careers Apply to the free AGI Strategy course: https://bluedot.org/courses/agi-strategy