Multiple people responded to a recent post about how it’s technically impossible to instruct an LLM to always follow your rules by saying you can’t do that with people either.
Guess what? I don’t want my software to have all the moral failings of people. That’s a bug not a feature.
@carnage4life maybe it would have been more clear if you mentioned that this was about presidents following constitutional rules (which I agree with you, is a moral issue), and not implied it was about people around you (or me, or anyone else for that matter), following your(my) personal rules.
:-)
@carnage4life You can form bonds with people over time. Increasing trust and shared understanding. This is a bidirectional process that has benefits beyond working together.
You can do something sort of similar with a deterministic tool: over time you develop an intuition of how it behaves and it can feel like it becomes an extension of you.
You can't do either of those things with a nondeterministic tool. It will never get to know you. You will never get to know it.
(1/2)
@carnage4life You can make agreements with people, in the forms of laws or contracts. People can still decide not to follow them, but then they will face consequences. From financial pentalties to prison sentences.
You can do that with the builders of deterministic tools. People have gone to prison for building unsafe machines.
You can't do that with the builders of nondeterministic machines. Somehow they're immune to any form of criticism or consequences. As are their tools. Insanity.
(2/2)
@txo_elurmaluta @carnage4life That would not be the *exact same conditions* anymore. I would have said 'obviously' but less obviously how about thread-scheduling order on the GPU.
Mathematically and computationally the statement 'technically impossible' is false.
@nf3xn @carnage4life ok ok, I get it. That’s why you put * to exact same conditions, because it is very dificult to happen, but if it happens the result would be the same right? So, if we rip of temperature and other “random” stuff and conditioning, an LLM would output the same to the same pront, right?
Thanks!
(Asking totaly to learn here)
@nf3xn @carnage4life They are not deterministic; they are probabilistic. Given the same prompt repeatedly, they will produce different outputs. There is always a non-zero chance they will go off the rails.
Anybody who tells you differently doesn't understand them or is trying to get your money.
@carnage4life Honestly it feels like a management challenge. Herding toddlers that don't always don't do what you tell them to do.
Just don't give them the keys to the beemer.
Dare Obasanjo
Demian
discobeez
cow99
Elric
Cybarbie
Txo!?
Urzl
RCG
Irene Zhang
sollat
Roy Greenhilt
Crack the Safe