There might be an analogy to AI here: suppose you take an embedding network (e.g. word2vec) and instantiate two instances with random weights. You then train ea

There might be an analogy to AI here: suppose you take an embedding network (e.g. word2vec) and instantiate two instances with random weights. You then train each network on the same training data up to the same level of accuracy. Given any particular pair of words (e.g. "red", "book"), the models should produce pairs of vectors that have similar similarity scores, and thus one could argue that they have the same "understanding" of the word "red" and its relationships with other words. But the actual vectors that the two model instances output for the word "red" will be totally different, and entirely incomparable, due to different random initializations. (It's overwhelmingly likely that they will converge to different local minima). Likewise, the actual internal representation of the color "red" inside your brain differs completely from the color "red" inside my brain, in terms of e.g. the set of neurons activated and the weights they use. That effect should run quite deep, and will be influenced by a myriad of life experiences - for example, maybe someone fears red because they saw too much blood as a kid, while someone else loves red because their culture views it as a symbol of luck & prosperity. It's nice that we have language as a common protocol for these deeply convoluted internal representations, even if it fails to communicate the full nuance of someone's subjective experience.