Goodfire announced a $10 million commitment on Oct. 26, 2026, to begin what it calls a “Human Genome Project for AI.” In an essay published that day, Tom McGrath and Mathilde Papillon describe it as an ambitious, cross-industry effort to reverse-engineer a large language model.
Their first goal is to find “scaling laws for interpretability,” so that understanding of a model grows predictably as more compute is applied. They propose putting capable AI agents in an environment where they can run and verify experiments at scale. The essay also calls for a rigorous way to check experimental claims and a system that helps researchers connect findings rather than collect disconnected results. Those are parts of the proposed research plan, not results the authors report having achieved.
An encyclopedia of how a model works
Goodfire's ultimate goal is to reverse-engineer a frontier-scale model and publish a living encyclopedia explaining how it works. The authors define the reverse-engineering goal as understanding a model well enough to give scientists and engineers mechanistic explanations in response to arbitrary questions about it. They envision an encyclopedia organized around research questions and answers, backed by a structured database of causal claims and experimental code rather than organized by model parts. They acknowledge that it is hard to define when reverse engineering is complete.
McGrath and Papillon say the effort will require collaboration across organizations. Scaling it to the most capable models may require billions in shared costs, they estimate. Their proposal also leaves practical questions open: how to keep the encyclopedia current as a model undergoes further training, and how far findings about one model would apply to others.
Comments