While I largely agree with the premise of a converged multimodal world model, I'd argue it's more like Plato's images/eikons than the ideal forms/eidelons. The representations don't seem to be lossless, and instead derive fidelity from the parameter scale.
If we had these representations converge in a real computing neural network, maybe then we could argue that they were Plato's ideals a step above the corrupted quantization of physical forms. But as they are right now, the representations are an even more reduced representations of the physical originals.
> We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality
Quite the opposite: it's a shallow shadow (no meanings, just chains) model of shadow reality captured (no translation to abstract) by language (only things for which we have names) etc. What they wish is a model of actual reality, not of that one from distorting mirrors.
The sanity is in what the one source of truth is - not like people could be a shadows of some ideal model being a reality (edit: statistical). However people actions are mostly results of using models (of situation, reality) and making them ad hoc.
This reminds me of how, if I recall correctly, in the original paper on adversarial attacks, the authors found that adversarial attacks on one neural network would generally have some success on other neural networks if they were trained for similar tasks (say, labeling images).
.. and that's so far away from Plato's concept of an ideal reality.. (two worlds away ?)
- that shared statistical model of reality is disconnected from physical and from abstract worlds (Popper's* world 1 and 3 platonic) and reflects (so far hallucinations are not alive) Popper's ‘world 2’ only - of subjective experiences - reduced to shared weight and vectors.
If we had these representations converge in a real computing neural network, maybe then we could argue that they were Plato's ideals a step above the corrupted quantization of physical forms. But as they are right now, the representations are an even more reduced representations of the physical originals.
Still very neat that it's happening though.