>Hey, JeetPT, does X work like Y?>>Yes, here are some completely bullshit reasons for why Y is correct.>Hey, JeetPT, how does X work?>>Yes, let me make up some completely irrelevant nonsense about how X works.How do you tard-wrangle these things? Do they only work on the most basic shit that it has 500GB of reddit retards asking the same questions about? Do they pretend to be retarded until you dig out the shekels?And why the fuck can they never just admit to not knowing shit rather than bullshitting obvious nonsense?
>why the fuck can they never just admit to not knowing shit rather than bullshitting obvious nonsense?because you're a fucking retard that doesn't understand how these things workThey literally can't do that
>>108496073Why can't they check the token probabilities to detect if it's bullshitting?
Wait for AGI first, early reports are suggesting about two weeks
>>108496122because their training data doesn't include 4chan
>>108496359>The future is by reddit, for reddit and from reddit
Prompt issue>hi jeetpt how xyz works use google, k?
>>108496122The reinforcement learning, ie the training stage, had retards vote that they would rather give bullshit wrong answers instead of saying "I dont know"
>>108496061>>108496061doesn't even matter
>>108496061>How do you tard-wrangle these things?You pay for it instead of using the free model that only works for simple things.
>>108496122how the fuck would that works? every token has 100% proba by virtue of having been picked. There is no bullshit detector because It doesn't pick a low proba token in the first place, that's the entire point of the system.>but then what happens when it hallucinates?it doesn't hallucinate, it's doing exactly what it was designed for: picking whatever correlated most with the previous token based on its training data. If you ask the spec of a car that doesnt exist and it responds with a sheet that is bullshit, it's not coming out of nowhere, it's what was the next most probable thing to say, generally what appeared most in term of car spec sheets in the training data, with flavors based on all the previous tokens.>but then AI is inherently flawed and fucking bullshit?yes, enjoy the bubble while it lasts.
>>108497761Bros, we're being replaced...
>>108496061its not that hard to just do your own research anon, it uses information from google and reddit, you can just go there to get the information you need, people have been doing this for decades before this useless psyop crudely disguised as revolutionary technology came along, and it rarely took longer than it took for the bot to generate a half coherent response (especially considering you need to fact check its bullshit by using google anyways)
>>108496122Because they are token probabilities, not truth probabilities. The model have no idea what's true, only what's an appropriate token based on the training data.
>>108496061As if this is worse than the old days>How do you do x?>This thread is duplicative. Deleting>The duplicate thread:>Nobody does x, do y instead>Here's a link on how to go x>Link to imageshack that's been down for years>I sent an explanation in your dmI'd rather tard wrangle gpt than ever go back to dealing with the quasi redditors over on stack overflow.
