RYS maps concepts beyond language

The core claim is strong: if the probes are sound, these models are converging on a language-agnostic internal space for semantics, not separately “thinking” in each tongue. The bigger caveat is that hidden-state clustering is evidence about representation, not proof of human-like cognition.

• –Cross-language, same-topic sentences cluster more tightly than same-language, different-topic sentences in the middle layers
• –The effect shows up across dense and MoE models, which makes it harder to dismiss as a quirk of one architecture or training run
• –Code, math, and natural-language descriptions converging in nearby regions is the more interesting result for developers: the model seems to normalize surface forms into shared concept space
• –If this holds up, it supports interpretability work that treats transformers as layered encoders/decoders around a semantic core, rather than language-specific reasoning circuits
• –It also sharpens the Sapir-Whorf comparison: the argument here is not about humans, but about LLMs treating language as I/O while meaning lives in geometry