AI doesn’t hallucinate. It’s a fancy marketing term for when AI confidently does something in error.
The tech billionaires would have a harder time getting the mass amounts of people that don’t understand interested if they didn’t use words like hallucinate.
Talking about hallucinations lets us talk about undesired output as a completely different thing than desires output, which implies it can be handled somehow.
The problem it the LLM can only ever output bullshit. Often the bullshit is decent and we call it output, and sometimes the bullshit is wrong and we call it hallucination.
But it’s the exact same thing from the LLM. You can’t make it detect it or promise not to make it.
You can’t make it detect it or promise not to make it.
This is how you know these things are fucking worthless because the people in charge of them think they can combat this by using anti hallucination clauses in the prompt as if the AI would know how to tell it was hallucinating. It already classified it as plausible output by creating it!
“On two occasions I have been asked, – “Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?” … I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question”
Its been over a hundred years since this quote and people still think computers are magic.
It’s a fancy marketing term for when AI confidently does something in error.
How can the AI be confident?
We anthropomorphize the behaviors of these technologies to analogize their outputs to other phenomena observed in humans. In many cases, the analogy helps people decide how to respond to the technology itself, and that class of error.
Describing things in terms of “hallucinations” tell users that the output shouldn’t always be trusted, regardless of how “confident” the technology seems.
Humans will anthropomorphize damn near anything. We’ll say shit like “hydrogen atoms want to be with oxygen so bad they get super excited and move around a lot when they get to bond”. I don’t think characterizing the language output of an LLM using terms that describe how people speak is a bad thing.
“Hallucination” on the other hand is not even close to describing the “incorrect” bullshit that comes out of LLMs as opposed to the “correct” bullshit. The source of using “hallucination” to describe the output of deep neural networks kind of started with these early image generators. Everything it output was a hallucination, but eventually these networks got so believable that sometimes they could output realistic, and even sometimes factually accurate, content. So the people who wanted these neural nets to be AI would start to only call the bad and unbelievable and false outputs as hallucinations. It’s not just anthropomorphizing it, but implying that it actually does something like thinking and has a state of mind.
AI doesn’t hallucinate. It’s a fancy marketing term for when AI confidently does something in error.
The tech billionaires would have a harder time getting the mass amounts of people that don’t understand interested if they didn’t use words like hallucinate.
It’s a data center, not a psychiatric patient
It’s also not intelligent but stochastic language models.
Agree, the term is misleading.
Talking about hallucinations lets us talk about undesired output as a completely different thing than desires output, which implies it can be handled somehow.
The problem it the LLM can only ever output bullshit. Often the bullshit is decent and we call it output, and sometimes the bullshit is wrong and we call it hallucination.
But it’s the exact same thing from the LLM. You can’t make it detect it or promise not to make it.
This is how you know these things are fucking worthless because the people in charge of them think they can combat this by using anti hallucination clauses in the prompt as if the AI would know how to tell it was hallucinating. It already classified it as plausible output by creating it!
They try to do security the same way, by adding “pwease dont use dangerous shell commands” to the system prompt.
Security researchers have dubbed it “Prompt Begging”
Its been over a hundred years since this quote and people still think computers are magic.
How can the AI be confident?
We anthropomorphize the behaviors of these technologies to analogize their outputs to other phenomena observed in humans. In many cases, the analogy helps people decide how to respond to the technology itself, and that class of error.
Describing things in terms of “hallucinations” tell users that the output shouldn’t always be trusted, regardless of how “confident” the technology seems.
Humans will anthropomorphize damn near anything. We’ll say shit like “hydrogen atoms want to be with oxygen so bad they get super excited and move around a lot when they get to bond”. I don’t think characterizing the language output of an LLM using terms that describe how people speak is a bad thing.
“Hallucination” on the other hand is not even close to describing the “incorrect” bullshit that comes out of LLMs as opposed to the “correct” bullshit. The source of using “hallucination” to describe the output of deep neural networks kind of started with these early image generators. Everything it output was a hallucination, but eventually these networks got so believable that sometimes they could output realistic, and even sometimes factually accurate, content. So the people who wanted these neural nets to be AI would start to only call the bad and unbelievable and false outputs as hallucinations. It’s not just anthropomorphizing it, but implying that it actually does something like thinking and has a state of mind.