From Ethan Mollick on LinkedIn:
Recently, LLMs were found to encode different languages in similar ways, suggesting a sort of Platonic representation of words.
It now extends to science: 60 ML models for molecules, materials & proteins (all with different architectures & training data) converge toward similar encodings of molecular structure.
As models improve, they agree more on what makes two molecules “similar.” They may be converging on a shared representation of physical reality (but models are still constrained by their data).
I find this a fascinating hint at an underlying structure of information and reality.
Give Ethan a follow—he’s a consistently considerate and often humorous source of insights on all things AI.
