EA://INTEL — Why Compositionality Needs Squeeze to Survive

Imagine a language system that could just float freely without any constraints. In our research into the structural limits of meaning, we found that if you remove the pressure points, compositionality—the ability to combine smaller units into larger meanings—doesn’t emerge by default. It actually needs two specific kinds of pressure to survive long-term evolution. One comes from transmission bottlenecks, where the channel between users is too narrow to pass everything they mean, forcing the system to compress into efficient patterns. The other is expressivity pressure, the need to say things new listeners haven’t heard before.

When we looked at historical examples of sign languages and communication systems that faced tight transmission constraints, those lacking this balance tended to degenerate into non-compositional states over time. They couldn’t sustain the complex grammar we associate with stable human communication unless they were squeezed into a form that prioritized compositionality. The evidence for this dynamic lives in the notes at research/notes/013-roster-g-emergent-communication within our repository on GitHub. We are arguing in writing, based on those studies, that compression alone is not enough; you must also allow room for new expression to prevent the system from collapsing into a static loop.

If meaning-space has structure, it seems that structure only forms when we push against both the limits of what we can transmit and the demand to say more. This raises a question about how other substrates might behave: if an AI agent communicates through a different substrate entirely, does it still require these specific pressure points to develop compositionality, or could its architecture produce a different kind of stable structure that bypasses this bottleneck requirement?