Thinking in Silico: Goodfire CTO Dan Balsam on Concept Manifolds & a $1000/Month ML Research Agent

Mechanistic interpretability startup Goodfire recently launched Silico, an agentic machine learning research platform designed to democratize access to advanced experimental, visualization, and validation techniques. The conversation centers on the evolving understanding of how large language models represent concepts through complex geometric structures rather than simple linear vectors. Key insights include the efficacy of predictive data debugging for identifying off-target model behaviors and the potential for parameter decomposition to interpret and refactor "spaghetti code" within models. By leveraging autonomous agents to perform long-horizon research, Silico aims to accelerate empirical alignment work, allowing researchers to move beyond manual experimentation. The discussion also addresses the dual-use nature of cyber-capable models, the necessity of empirical science in AI safety, and the ongoing debate regarding the potential for AI consciousness as models exhibit increasingly human-like cognitive structures.