This reminds me a lot of a natural language processing tool called word2vec. It reduces words into vectors, where each dimension represents something about the

This reminds me a lot of a natural language processing tool called word2vec. It reduces words into vectors, where each dimension represents something about the meaning about the word. E.g. whether it's a place, or if it's male or female. And these vectors are not designed by humans, but learned from data. It's optimized so that words that occur in similar contexts have similar vectors. E.g. places tend to occur near words talking about places, like "there" or "go", and male names tend to occur near "he" and "his", etc. You can do really interesting things with these vectors. Like 'King'-'man'+'woman'='queen'. Recently a computer program beat the word analogy part of an IQ test using this. It's useful in computer translation. Take two languages and constrain a few identical words to have the same vector. Then all the other words also end up with similar vectors. Anyway, with something like this, the actual words and symbols don't matter, just the 100 or so dimensions needed to accurately represent the meaning of every word. Of course speaking a 100 dimensional vector is way less efficient than a single word or symbol. But perhaps using this, we could come up with an absolute minimal set of symbols necessary. And some rules for combining them to get all words. I know a lot of the comments here are complaining that this doesn't make a good natural language. But that's not the point of an auxiliary language. The main thing auxiliary languages should optimize for is how easy it is to learn. Simpler is better. Reducing the words and grammar is hugely important. Another advantage of such an approach, is that it might be more amenable to machine translation. Which is another advantage for an auxiliary language.