What an interesting mess! It occurs to me that a canonical semantic representation of all known (extracted) language concepts would be useful too. Now that we

What an interesting mess! It occurs to me that a canonical semantic representation of all known (extracted) language concepts would be useful too. Now that we have multi-language LLM's it would be an interesting challenge to create/design a canonical representation for a minimum number of base concepts, their relations and orthogonal "voice" modifiers, extracted from the latent representations of an LLM across a whole training set, over all training languages. While the best LLMs still have complex reasoning issues, their understanding of concepts and voice at the sentence level is highly intuitive and accurate. So the design process could be automated. The result would be a human language agnostic, cross-culture concept inclusive, regularized & normalized (relatively speaking) semantic language. Call it SEMANTICODE. We need to get this right, using one standard LLM lineage, before the Unicode people create a super standard that spans 150 different LLM's and 150 different latent spaces! :O Stability between updates would be guaranteed by including SEMANTICODE as a non-human language in training of future LLM's. Perhaps including a (highly) pre-normalized semantic artificial language would dramatically speed up and reduce the parameter count needed for future multi-language training?* Then LLMs could use SEMANTICODE talk to each other more reliably, efficiently, and with greater concept specificity than any of our single languages.