4 ms·
Hear me out: LLMs should store their responses so that when someone asks them "did you write this?", even with modifications of the response they can confirm or
by _8j50 3y ago
Hear me out: LLMs should store their responses so that when someone asks them "did you write this?", even with modifications of the response they can confirm or deny. Since they are costly to operate, it will cut down on their abuse.
- xtreme 3y agoAssuming this is a serious suggestion and not a joke that I have missed, there are so many cloud LLMs that this kind of search is not feasible. And local LLMs that you can run on your own hardware makes detection completely impossible. It's better to acknowledge that people will use LLMs, and judge the papers based on their content and not their writer.
- _8j50 3y agoI was serious, I meant the cloud LLMs, the ones you run on your own are out of scope
- than3 3y agoThis will never work, fundamentally any computer is unable to solve certain types of problems. They run into several different core computer science problems when you try to have them do this. There is the halting problem and the decidability problems. You can approximate certain things using a perceptron/neuron floating point configuration, but they are only approximates; and do nothing towards addressing either of those problems.
- _8j50 3y agoEven with an approximation? "That sounds like something I might say", people are capable of doing this, why can't LLMs?
- than3 3y agoSpooky action at a distance? Kidding aside, no one knows what exactly we do differently that allows for complex thought, context switching, or conflict resolution that works around those aforementioned problems.