AlphaDev

AlphaDev is a reinforcement learning system from Google DeepMind that discovers faster low-level computer science algorithms. Introduced in a Nature paper on June 7, 2023, it treats algorithm design as a single-player assembly game where each move appends a CPU instruction, building on the AlphaZero lineage behind AlphaTensor. It discovered faster assembly sequences for small sorts, including the novel AlphaDev swap and copy moves, and hashing algorithms up to 30 percent faster in the 9-to-16-byte range.

What makes AlphaDev unusual is production impact: in January 2022 DeepMind submitted its sorting algorithms to the LLVM libc++ maintainers, and after independent vetting they were merged. Google called it the first change to that part of the sorting library in over a decade and the first to come from reinforcement learning. A hashing routine followed into the Abseil library in January 2023. Google estimates the two algorithms run trillions of times per day. As with AlphaTensor, there is no downloadable model: openWeights is false, while the discovered algorithms themselves were open-sourced through the libraries that adopted them.