|
My issue with LLMs is that you require a review-competent human in the loop, to fix confabulations. Yes, I’m using them from time to time for research. But I’m also aware of the topics I research and see through bs. And best LLMs out there, right now, produce bs in just 3-4 paragraphs, in nicely documented areas. A recent example is my question on how to run N vpn servers on N ips on the same eth with ip binding (in ip = out ip, instead of using a gw with the lowest metric). I had no idea but I know how networks work and the terminology. It started helping, created a namespace, set up lo, set up two interfaces for inner and outer routing and then made a couple of crucial mistakes that couldn’t be detected or fixed by someone even a little clueless (in routing setup for outgoing traffic). I didn’t even argue and just asked what that does wrt my task, and that started the classic “oh wait, sorry, here’s more bs” loop that never ended. Eventually I distilled the general idea and found an article that AI very likely learned from, cause it was the same code almost verbatim, but without mistakes. Does that count as helping? Idk, probably yes. But I know that examples like this show that you cannot not only leave an LLM unsupervised for any non-trivial question, but have to leave a competent role in the loop. I think the programming community is just blinded by LLMs succeeding in writing kilometers of untalented react/jsx/etc crap that has no complexity or competence in it apart from repeating “do like this” patterns and literally millions of examples, so noise cannot hit through that “protection”. Everything else suffers from LLMs adding inevitable noise into what they learned from a couple of sources. The problem here, as I understand it, is that only specific programmer roles and s{c,p}ammers (ironically) write the same crap again and again millions of times, other info usually exists in only a few important sources and blog posts, and only a few of those are full and have good explanations. |