Descriptions:
Goodfire CTO and co-founder Dan Balsam joins the Cognitive Revolution podcast to discuss the commercial launch of Silico, the company’s long-horizon agentic machine learning research platform. Originally built as an internal tool, Silico is now available at $1,000 per month per seat for enterprise customers — a dramatic reduction compared to Goodfire’s high-touch research engagements that can reach seven figures. Goodfire also plans to introduce special pricing and grants for safety and alignment researchers.
The conversation digs into several strands of Goodfire’s published research, including predictive data debugging — which uses interpretability techniques such as sparse autoencoders to identify which concepts a network update is likely to affect before training, allowing teams to catch anomalies early. Balsam also explains the company’s work on concept manifolds, the complex geometric structures that large language models use to represent advanced knowledge such as the periodic table and the tree of life, extending the classical linear representation hypothesis into richer nonlinear territory. He discusses how these geometric insights are being used to improve model steering and enable fine-grained control over what individual languages or skills a model retains.
The episode also covers Balsam’s views on AI safety more broadly: his reasons for signing the recent Pacing the Frontier letter, concerns about AI’s potential role in biosecurity risks, the importance of open-weight models for preventing dangerous concentrations of power, and Goodfire’s internal conversations about AI welfare and consciousness.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published August 08, 2026
🏷️ Format: Podcast
