Comment by myffical

Some DeepMind researchers used mechanistic interpretability techniques to find concepts in AlphaZero and teach them to human chess Grandmasters: https://www.pnas.org/doi/10.1073/pnas.2406675122