Autonomous AI research that hand-writes transformer weights to solve tasks. Each run iteratively proposes interpretable circuits; below, every point is one circuit, numbered in the order it was tried.
Pick a task. The plot shows accuracy vs. number of parameters; points are numbered by the order tried and connected along the path the search took. Hover a point for its method & description.
Every task overlaid. Each line traces one task's search path through (parameters, accuracy) space, in the order circuits were tried.