What a 23% kernel speedup tells us about whether AI can improve itself and how to quantify the gains. In 2025, Google DeepMind’s AlphaEvolve, an AI system that discovers algorithms, found a way to make a FlashAttention kernel run 23% faster.