btw, having ideas without words, is precisely something else orthogonal and related to current LLM technology. you can have a much bigger budget for thinking, if you don't translate it to english (or whatever) thus enabling a much greater amount of thinking, which is key to it resolving answers to questions. the frontier models don't do this, because of muh safety reasons. it's kinda stupid because the tokens could be logged and examined for auditing afterwards. all i know is, bumping my thinking token limit up to 20k tokens has dramatically reduced the amount of retarded things that claude does, so if i could have it not translate the thoughts into english, i mean, how precise wouldl it get? the factor is nearly orders of magnitude, but it's "dangerous"