Comment by philipkiely

I think the biggest net new recent technique is P/D disaggregation. And that spec dec is very different now especially post DSpark/DFlash.

But overall yes the fundamentals of LLM performance optimization have been remarkably stable over the last few years.