* Rudolf Adamkovič <rud...@adamkovic.org> [2025-02-24 22:28]: > Jean Louis <bugs@gnu.support> writes: > > > Do you have examples that never seen instances of problems have been > > solved? > > Any instance that was not in the training set. It can be as simple as > asking a question about a never-seen graph. > > For a more interesting example, I recently gave a model a specification > of a "fantasy" Lisp dialect, along with a sample program written in it, > and asked what the output of the program would be. The model computed > the expected result correctly. It took a while, with pages and pages of > reasoning, but it did it. In my opinion, this kind of generalization is > at the core of intelligence. > > And we shall end it here, given this is `emacs-devel'. :)
I asked for specific instance but I got a general answer. LLM cannot answer that question. Translation, text generation, image creation, none of it appears to me as really "innovated", it appears generated based on past information. Many things in the data sets are derived, intertwined, so the knowledge existed already and principles to solutions were already recorded in the model, it just needs interfences to get it out. That is like saying every day it is giving new answer, but that answer is not innovative. It is based on past knowledge intertwined. General principles are there, and are not even considered "principles" by the model, rather those are inferred from the bunch of knowledge, model knows nothing about any principles. I find DeepSeek working well to find out what is wrong with bloody Python errors, so it can faster bring me to the goal. Though all such as brought from similar information by probability. I can understand it can predict how some chemical is going to react with other, and such reports are basically cutting the work of the human. Just I do not see the innovation. Making a picture is not an innovation to me, again there is question open, what exactly was ever truly invented by any of the LLMs? -- Jean Louis --- via emacs-tangents mailing list (https://lists.gnu.org/mailman/listinfo/emacs-tangents)
