Kale Good morning folks
I'm not saying that LLMs ignoring stop tokens and just going on their own weird dream-cycle exploration of time and space isn't interesting, but I am absolutely saying it's not sentience.
Basically:
* You ask a question
* The model answers it
* The model ignores (probably a setup issue) the trigger to stop talking built into the source data
* The model then goes 'well... what's next?'
* Probability says that what comes next is user input
* The model simulates the user asking a question
* The model then answers that question
* Cycle continues
* Things go off the rails eventually.
* You ask a question
* The model answers it
* The model ignores (probably a setup issue) the trigger to stop talking built into the source data
* The model then goes 'well... what's next?'
* Probability says that what comes next is user input
* The model simulates the user asking a question
* The model then answers that question
* Cycle continues
* Things go off the rails eventually.
You ever have a coworker who answers a question, poses a question to themselves, answers that question, then keeps going?
Yeah, it's exactly that. They're missing the 'and now I stop talking' cue.
Craig P@DarkestKale I have the ability to stop rambling some days! On those days, I'm no longer sentient, I guess.