[amluto]

I’m somewhat skeptical, at least for LLMs per se. They are language models. They have, in many respects, superhuman abilities, but I wouldn’t really want to trust things that require accuracy to an unaided human or even an unaided superhuman.

I would expect better results from LLMs using programming languages, perhaps ones tailored to LLMs, to prepare tasks on behalf of their users.

(Also, LLMs doing anything direct are an incredibly inefficient use of computing resources. There are quite a few orders of magnitude of difference between the FLOPs needed to do basic calculations and the FLOPs needed to run inference on a large model that may be able to do those calculations if well trained.)

[jameshart]

The thing is, a lot of tasks don't require accuracy.

We force accuracy on them when we computerize them, because computers historically haven't handled ambiguity well. That demands the skill of 'programming' - interpreting fuzzy real world problems, making them precise enough to be modeled in a way that classical computing can handle, and making computer routines to help.

But the underlying problem humans are looking for a bicycle-for-the-mind to make easier often didn't start off 'precise' at all.