The following article discusses how Google’s New DeepMind system called AlphaTensor has been pointed at Matrix Multiplication algorithms and discovered ways to improve performance of small to medium sized Matrix Multiplication by 20% or more, by reducing the number of multiplications and additions involved. Some of these new algorithms are quite non-intuitive. Are there plans to integrate these new algorithms into future versions of Mathematica? Since Matrix Multiplication is part of a wide number of applications, such as image manipulation and ironically Machine Learning itself, it should represent a noticeable improvement in many areas. There are several research papers on the web listing the exact steps of the improved algorithms, approximately 14,000 in total.

https://www.nature.com/articles/s41586-022-05172-4