The thing is, if you want to sort a list, LLM is the wrong tool. However, if what you want is to generate a program that will sort a list, your LLM will be pretty performant.
It’s a powerful tool but like any powerful tool, you will use it wrongly if you don’t acknowledge its strengths and weaknesses.
Judging a LLM because he can’t act as a program is a fundamental error because it means you are wasting an enormous amount of resources for nothing.
The novelty of LLMs and where they are good at is helping you thinking about things. Helping you to complete reasonings about your thoughts or about your data. They are not good at doing things but they are good at « feeling » things.
Re:Oops! We Automated Bullshit
by FeepingCreature on Friday November 17, 2023 @10:10AM
(#38304609)
A fast 50% sort could be very useful given the right circumstances.
Yeah, so long as you have something else that can quickly verify that a list is sorted, just keep applying this "quick coin toss sort" until you get a sorted list.
If a ML algorithm with a 90% success rate and a "sanity checker" might be much easier to implement than a ML algorithm with a 100% success rate.
For a long time there have been "machine learning" techniques where getting the right answer 80% of the time is a win, what's interesting about LLMs is they sometimes "raise the bar" for the machine learning problems (get it right 90% of the time) but have also lowered the bar for other things that can be done algorithmically 100% of the time.