And when you do, please don't spend *any* time optimizing the Keras 3 version. Just time your first, naive implementation. Because you'll still see the speedup. That's the point of using good tools

→ View original post on X — @fchollet