Can you reverse engineer our neural network?
Summary
Jane Street describes a ‘capture-the-flag’ style ML puzzle where a hand-designed neural net is provided with weights and architecture, inviting reverse engineering using mechanistic interpretability tools. The piece chronicles a solver’s path—from inspecting the final layer to reducing the network to a linear program, then a SAT formulation, and finally a brute-force approach after identifying a hash-like mechanism and a bug. It also teases a new puzzle and invites readers to join Jane Street.