In an earlier reflection I asked whether we still need formal semantics when language models already seem to converge on meaning. A paper has sharpened that question for me.
The Platonic Representation Hypothesis argues that representations in deep networks are measurably converging across architectures, modalities and training objectives, towards a shared statistical model of reality.
That raises a question I want to sit with: if the models are converging, are we converging too?
A quiet delegation
The mechanism is hard to argue with. We mediate more and more of our understanding through language models: drafting, summarising, reasoning, deciding. What we read is sampled from a converging representation. What we write with their help is shaped by it.
If meaning-making routes through that layer, it inherits its geometry.
This is a quiet delegation. We hand the work of understanding to a layer that increasingly speaks with one voice, and accept its synthesis as if it had been ours.
From tool to repository
The language model is becoming more than a tool. It is becoming the repository of knowledge: the place where synthesis lives.
Libraries preserved divergence. They kept disagreeing books on the same shelf.
A converging repository does the opposite. It offers a single statistical horizon that every consultation returns to.
Why this matters for semantics
For those of us who work on semantic interoperability, this is a strange mirror. We spend our careers helping meaning converge where convergence serves people: shared definitions that let services work together. Convergence of everything, including how we think, is a different matter.
Perhaps this is another reason formal, deliberate models still matter. They record the meanings we have chosen, and keep room for the ones we have chosen differently.
What happens to us when a converging model is where we go to think?