Three papers I keep re-reading about inference costPaged attention, speculative decoding, and 4-bit weights — what each one actually buys.