The reduction in compute required to reach a given capability level, achieved through better algorithms, architectures, or data curation rather than more hardware.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Claims 6.2 and B3 both turn on whether algorithmic progress can deliver frontier capability with less compute. The scaffold has 'scaling-law' (more compute → better models) and 'compute' (raw resource), but no concept for the inverse relationship, getting the same capability for less. This gap leaves the B3 bridge and the export-controls debate (6.1, 6.2, 6.3) without the vocabulary they need.