> From: Madhu <enom...@meer.net> > Date: Sat, 22 Mar 2025 21:25:43 +0530 > > > (And no, ML and LLMs do not engage in combinatorial computations, they > > solve problems differently, and that is why they succeed, because in > > many cases the problems are NP-hard, so cannot be solved by dumb > > exhaustive search of the solution space.) > > Aren't LLM solutions are produced in a similar way (i.e. produce a > solution via "fire first" and then draw the target around the hit point > and say this is the solution)? And in a world where LLM is implemented > the end consumer (victim?) has no choice but to accept the solution > (since the problem was intractable anyway).
Not the models with which I have the experience, no. They would be useless for me if they did that. Because I have real hared problems to solve, and they must be solved correctly and better than the alternative methods which are not based on LLM, when compared by some meaningful metrics. And even in other, much simpler, cases this is not so IME. Consider the "text completion" feature, for example, which is present in many tools (including Emacs, but in Emacs it is not based on LLM): you type text and the application offers you likely completions based on what you typed till now. If the completion is not what you intend to write, why would you "have no choice but accept"? You just keep typing instead of pressing TAB to accept. --- via emacs-tangents mailing list (https://lists.gnu.org/mailman/listinfo/emacs-tangents)
