what is the known-good thing? The whole point is that LLMs were not optimized at all, they got better results than older ML algorithms just because they are abl

what is the known-good thing? The whole point is that LLMs were not optimized at all, they got better results than older ML algorithms just because they are able to use all of the GPU, where older algorithms are designed for 10yo GPUs and can't make use of modern GPUs. But now you do in fact have to optimize, to the point that transformers look a lot more complicated than "attention is all you need."