Dev Tools · 1h ago
DeepMind's mctx bug caused 3x slowdown in MCTS search
A developer discovered that DeepMind's mctx library had a quadratic-time bug in its Monte Carlo tree search, making per-simulation cost double as simulations doubled. The fix, submitted as PR #116, rewrote an 80-line function to eliminate unnecessary buffer copies, restoring linear scaling. The bug was found by benchmarking an AlphaZero agent for Factorio factory layout on a single RTX 5070.
Meridian48 take
The fix is a reminder that even well-engineered libraries can hide O(N²) performance traps, and that profiling at scale is essential for AI training pipelines.
Read the full reporting
Accidentally quadratic: buffer copies made MCTS in DeepMind's mctx 3 slower →
DEV Community
deepmindmcts