upvote
No, you can't know that the output of a program is unreliable just from the fact that it outputs one words at a time. I already told you that you can perfectly compile a normal program, like a calculator, into the weights of an autoregressive transformer (this comes from works like RASP, ALTA, tracr, etc). And with this I don't mean it in the sense of "approximating the output of a calculator with 99.999% accuracy", I mean it in the sense of "it deterministically gives exactly the same output as a calculator 100% of the time for all possible inputs".
reply
> No, you can't know that the output of a program is unreliable just from the fact that it outputs one words at a time

Yes I can, and it shows everytime the "smart" LLMs suggest us to take a walk to the carwash or suggests 1.9 < 1.11 etc...

reply