You are not wrong, just ignore the comments here. Current day LLMs are a "deep seek" mechanism. For every task that had little or no equivalent in the internet, they were miserably bad (from O1 to deepseek to claude). We clearly need a different paradigm for "intelligence" and a sort of "consciousness" that gets a "handle" of reality. LLMs are very inefficient because they are trying to implement these mechanisms within a "predict the next token" environment.

All advancements in "AI" in the last year or so were in smaller models becoming more powerful, more companies reaching openai levels and a bit more accuracy. I think this is the plateau for the current theoretical tech.

+: one case I bumped into today: socks5 with the latest hyper version to proxy https requests. All models were completely off-track. None was able to make any functioning implementation. This is just a few lines of code.