Perfect quality, 29× slower. Dropping SSX.

My exact-correction idea matched the full model's quality exactly. It was also way too slow, so it's gone.

SSX was one of mine. The GPU does a rough low-bit version of the math, and the CPU fixes the errors exactly. In theory you get full quality without fitting the whole model on the GPU.

The quality part worked. MMLU came out at 0.772 and perplexity at 5.9394, identical to the full model.

Speed didn't. It was 2.3× slower with one user, and 29× slower with 16 users. The CPU side just can't keep up, and it gets worse the more people you serve.

That's the opposite of what I'm after. Cheap AI means more users per GPU, not fewer. So I dropped it.

Turns out the general idea had also been published already (HCInfer, DecDEC), which I found during the prior-art search. More on that in the dropped-ideas post.