I'm not sure how I feel about the Platonic Representation Hypothesis.
In some ways it seems like it should be profound, but part of me thinks it might just be saying 'There is an optimal way to perform dimensionality reduction and as things get bigger it is easier to appoximate'
It’s less profound sounding if instead of talking about “representations of the world” you talk about compressed representations of data or decision boundaries in a classifier over data.
Do memes have a concept of proto memes? Like primes are the proto numbers that all other numbers are made from? Could there be similar ontological primes for all concepts / semantic space? Base reality concepts that are irreducible so to speak.
Consider a binary tree of concepts from the root level more abstract concept to derived concepts applicable to specific domains. Do there exists additional derived concepts which are like an oblique plane which intersects the tree. Sort of like a concept which traveses maths, physics, biology, chemistry and intersects at different abstraction points. I am definitely sure language is flexible enough to have such entities.
This is very interesting! It's really amazing that all the frontier AI's obviously understand concepts and can talk to us on our terms about it. It makes sense that humans all end up with a similar understanding because we all have the same brain architecture, live in the same world and read the same texts.[1]
AI's, however, do not have the same brain architecture we do, do not train in the same sensory environment (they don't even have a live sensory loop), or talk to people interactively during training. They don't even all have the same architecture, they have many different architectures.
It is therefore very remarkable that, somehow, AI's still manage to learn the same concepts we use, and have obvious facility with thinking through the concepts and their interactions. Except for the very low-quality slop I mentioned in my footnote, you'll rarely or never see an AI combine concepts in a way that sounds superficially plausible if you don't think about it but is a non-sequitur if you do. They really have very good command of concepts.
You can test this by asking them to combine a handful of different concepts to your specifications or consider the result - they have no problems doing so, they truly excel at this.
It is amazing because concepts are so abstract. They're more like spots of oil paint than like lego pieces. And yet, AI's "see" them just as we do.
I also find it remarkable that the same AI models are able to be generalists in every field while being masterful specialists at the same time. (In the leading current AI architectures for frontier models, every single weight is activated for every single token generated! That means it literally thinks about every single thing it has ever learned while writing each and every word! Mixture of experts architecture cuts down on this a little, but maybe by 200 (number of experts) or 400 or something. So, it uses at least 1/400th (if it only activated one expert) of everything it knows, which considering that they're trained on basically all knowledge is still immense.) I would think that using so much knowledge in every token would lead to indecision or conflicting or garbled output, and yet frontier models generate really good output without such an effect.
[1] Though, some people are trying to change this so that each person's entire reading is the output of a low quality LLM, most of you have probably noticed that all web pages have been totally rewritten to be all about you - this is a weird experiment the U.S. government condemns and is trying to stop. You can read more about the U.S. government's attempts to stop the rewriting to all web pages, papers, books, posts, and comments with low-quality LLM slop.
I thought the platonic representation hypothesis was much more ambitious: wasn't it supposed to apply for any model across modalities (e.g. including human minds), rather than just the same model but with different modalities? I'll have to actually read the paper but I suspect there to be very strong assumptions which defeat the interesting aspects of the platonic representation hypothesis.
In some ways it seems like it should be profound, but part of me thinks it might just be saying 'There is an optimal way to perform dimensionality reduction and as things get bigger it is easier to appoximate'
AI's, however, do not have the same brain architecture we do, do not train in the same sensory environment (they don't even have a live sensory loop), or talk to people interactively during training. They don't even all have the same architecture, they have many different architectures.
It is therefore very remarkable that, somehow, AI's still manage to learn the same concepts we use, and have obvious facility with thinking through the concepts and their interactions. Except for the very low-quality slop I mentioned in my footnote, you'll rarely or never see an AI combine concepts in a way that sounds superficially plausible if you don't think about it but is a non-sequitur if you do. They really have very good command of concepts.
You can test this by asking them to combine a handful of different concepts to your specifications or consider the result - they have no problems doing so, they truly excel at this.
It is amazing because concepts are so abstract. They're more like spots of oil paint than like lego pieces. And yet, AI's "see" them just as we do.
I also find it remarkable that the same AI models are able to be generalists in every field while being masterful specialists at the same time. (In the leading current AI architectures for frontier models, every single weight is activated for every single token generated! That means it literally thinks about every single thing it has ever learned while writing each and every word! Mixture of experts architecture cuts down on this a little, but maybe by 200 (number of experts) or 400 or something. So, it uses at least 1/400th (if it only activated one expert) of everything it knows, which considering that they're trained on basically all knowledge is still immense.) I would think that using so much knowledge in every token would lead to indecision or conflicting or garbled output, and yet frontier models generate really good output without such an effect.
[1] Though, some people are trying to change this so that each person's entire reading is the output of a low quality LLM, most of you have probably noticed that all web pages have been totally rewritten to be all about you - this is a weird experiment the U.S. government condemns and is trying to stop. You can read more about the U.S. government's attempts to stop the rewriting to all web pages, papers, books, posts, and comments with low-quality LLM slop.