Okay, your output is different given the same input… So what? It’s a well known fact that these LLMs are non deterministic. Theres a guy on youtube that asks chatgpt everyday to count to 200 until it doesn’t fuck up. Your output does not prove or disprove the authenticity of the original post.
Tbh them being nondeterministic is a big part of why they’re so unreliable. Like, maybe it’ll work fine for 9/10 people, but then there will be that one person whose home directory gets wiped for whatever reason. Or maybe it’ll do math right for those nine people, but then for that one person it’ll say 1 + 1 = 11.
You’re basically gambling if you don’t verify the answers.
I’m referring to nondeterminism for the same prompt, since unless you start a session from scratch, it’s unlikely you’ll have the same history. If you give it a prompt, then depending on what you’ve told it previously, it may blow up in your face.
Determinism for the same prompt means you can’t give it context through a conversation, which vastly shrinks its utility.
That said, even that form of determinism can be unreliable: the example of arithmetic still works; you could have it completely deterministic, but if it only performs correctly on 80% of arithmetic problems, it’s still unreliable.
In fact, if you give it this prompt 50 times and it only fucks up once, that clearly indicates that this is post is misleading. It’s also likely this post was faked with a different prompt than the one shown.
Okay, your output is different given the same input… So what? It’s a well known fact that these LLMs are non deterministic. Theres a guy on youtube that asks chatgpt everyday to count to 200 until it doesn’t fuck up. Your output does not prove or disprove the authenticity of the original post.
Tbh them being nondeterministic is a big part of why they’re so unreliable. Like, maybe it’ll work fine for 9/10 people, but then there will be that one person whose home directory gets wiped for whatever reason. Or maybe it’ll do math right for those nine people, but then for that one person it’ll say
1 + 1 = 11.You’re basically gambling if you don’t verify the answers.
Not really… Determinism would only help if you could copy someone else’s prompt and history 100%, which you generally would not be able to.
Because maybe it always gets 1+1 correct, but fails 1+2.
I’m referring to nondeterminism for the same prompt, since unless you start a session from scratch, it’s unlikely you’ll have the same history. If you give it a prompt, then depending on what you’ve told it previously, it may blow up in your face.
Determinism for the same prompt means you can’t give it context through a conversation, which vastly shrinks its utility.
That said, even that form of determinism can be unreliable: the example of arithmetic still works; you could have it completely deterministic, but if it only performs correctly on 80% of arithmetic problems, it’s still unreliable.
In fact, if you give it this prompt 50 times and it only fucks up once, that clearly indicates that this is post is misleading. It’s also likely this post was faked with a different prompt than the one shown.