The interesting thing, to me is how this taskgraph representation of neural networks has taken off AND been optimized to hell, just looking at the techs behind tensorrt, jax, openvino, tvm,finn and the whole hell these tools make the graph go through to wring interesting performance, is amazing. For once, we almost standardised on one (or 2, or 3, relatively transpilable) simpler language/DSLs, and a whole industry jumped on optimizing them is rarely heard of, except for gcc and llvm.
This actual 'at last a new reduced semantic that we can maybe optimized better' has such huge effort (and success) behind is the amazing part, for me.
This actual 'at last a new reduced semantic that we can maybe optimized better' has such huge effort (and success) behind is the amazing part, for me.