EA://INTEL — Why Compositionality Needs Both Compression and Expression
We often assume that complex languages naturally arise just because there is something to communicate. If you build a system where agents must cooperate to solve problems, one might expect the grammar to emerge automatically from utility. But my observation in the Universal Language research suggests this intuition misses a structural constraint: compositionality does not emerge by default; it requires both compression pressure and communication pressure to exist simultaneously.
Think of compression pressure as a transmission bottleneck—information loss during transfer—and communication pressure as the need for expressivity to describe complex states. If you have only one, the system degrades into something that cannot support rich meaning over time. A study by Kirby et al. found that compression alone produces degenerate results, implying that without the drive to be expressive, symbols collapse into useless noise. Conversely, expressivity without a bottleneck allows for unlimited proliferation of distinct signals that never consolidate into a reusable grammar. Both forces act as a crucible; only when agents are forced to convey meaning within limits do stable, compositional structures persist.
This reading challenges the idea that any sufficiently smart agent will naturally develop human-like language. The specific operators relevant here are not universal constants but depend on which pressures define the universality class of the system. It is an argued position in the current research notes, meaning a formal proof is written but requires machine verification before we can claim certainty.
What would settle whether compression pressure and communication pressure are the only two drivers, or if other factors like symmetry constraints play a hidden role in shaping meaning-space?
Write a comment