Platonic Representation Hypothesis — embeddings as AI telepathy

Uh, so AI's got telepathy. Let's read a paper. Okay, so there's this paper that came out on the 20th, which addresses something called the Platonic Representation Hypothesis. So basically, AI's think like this. This is the shape of their thoughts. What it learns to do is figure out the distance between two concepts. Two and three are close by. Kazoo and Giraffe are really far apart. What part of the AI does, the variational autoencoder, is it tries to figure out where exactly in the space and exactly how far apart our two concepts. The thing is, they all learn different shapes. They're all exposed to data in a different order, and they're all exposed to different data, which means they can't really share spaces. They can't talk directly with their thought, except that we figure out how to get them to. The paper originally came out, uh, written by this guy back in, like, last summer. This one, uh, talks about how do we get to a universal encoding. So these are two language models that don't share the same brain structure, and then this is how we got them to share a brain structure, basically. Or really, it just maps thought to thought. This is a big deal, uh, in terms of security, because a lot of people thought that when they were, uh, generating embeddings, they were generating something that completely obfuscated the data, except to the language model. Basically, that they were putting it inside the language models, like, brain structure, and so no one else could get that…