"The vast majority of ML/AI software that is published (which often implements revolutionary techniques!) is of incredibly poor quality from a software engineering standpoint. As a result, practical AI applications are orders of magnitude slower than they could be."
That's quite the claim. Why should anyone take your word for it?
There are literally performance breakthroughs being made every few weeks (most recently, PowerInfer, which can speed up LLM inference by 10x or more) where the main improvement amounts to the application of caching/preloading techniques.
If the "engineering" part of ML had kept up with the "science" part, I have no doubt that performance for typical use cases would be 100x-1000x higher than we are seeing today.
> If the "engineering" part of ML had kept up with the "science" part, I have no doubt that performance for typical use cases would be 100x-1000x higher than we are seeing today.
That's quite the claim. Why should anyone take your word for it?