4 ms·

Why do A.I. chatbots tell lies and act weird? Look in the mirror

13 points by furcyd 4y ago

ThrowawayTestr 4y ago
If language models were only trained on factual sentences, would they still hallucinate?
- nickthegreek 4y ago
  Yes
Juliate 4y ago
They’re not telling, they are paraphrasing what they’ve been input with.
ElfinTrousers 4y ago
"My parrot swears so much! What the fuck can I do about that shit?"
elmerfud 4y ago
When I mirror people's behavior they get mad. When a chatbot does it, the bot must be broken. Staring in to the mirror is staring in to the abyss. People are not ready to see their true soul. Often times people don't even realize their casual lies. I'm not talking about mistakes in understanding, I'm talking about intentionally saying thing they know not to be true. We often call these "white lies" and it's done to spare feeling or avoid confrontation or really it's done because people lack fortitude to speak truth or not answer. Of course many times the asker is a fault too. We sometimes ask questions we should know better than asking. I tried to impress that idea of, only ask if you want to know, upon my children. Kids ask anything often times parents should moderate their answers. Once my kids were in their adolescence I let them know that they need to think about these things. They are mature enough to handle adult answers about things and if they ask me something I'll answer them directly and truthfully. There was a bit of a transition phase where they would ask something without thinking and I'd have to look at them and ask if they really wanted to know that answer. Usually they'd think about about it and withdrawal the question. My experience with AI is that it still has the emotional maturity of a small child not able to comprehend or regulate such matters. Because that kind of comprehension is algorithmically difficult. So you 5 year old with massive amounts of knowledge regurgitating it back trying mimic it's creators. Like a child it can't always understand fact and fiction. It really doesn't understand white lie vs lie lie. It's a imperfect mirror, but it's a mirror that perfect at magnifying our faults that we have collectively considered "normal".
jaybrendansmith 4y ago
These algorithms may approximate the language center of a human brain, but they cannot attach real axiomatic meaning to any of these words. So much of our IQ testing is focused on the language use, but that's simply the communication pathway for ideas, not the ideas themselves. How many times have I struggled for the right words to explain a concept? That's my language model attempting to communicate my axiomatic, emotional, spacial, conceptual, or mathematical mind. Words often fail. These chatbots are essentially idiot-savants in a very real sense, but they are clever enough to sometimes explain the reasons why they told an untruth or made an error. Trying to explain to them their error is foolish, because there's no 'mind' behind the words.
neonate 4y ago
https://archive.ph/VB5PQ https://archive.ph/VB5PQ

13 points by furcyd 4y ago

ThrowawayTestr 4y ago

If language models were only trained on factual sentences, would they still hallucinate?

nickthegreek 4y ago
Yes

Juliate 4y ago

They’re not telling, they are paraphrasing what they’ve been input with.

ElfinTrousers 4y ago

"My parrot swears so much! What the fuck can I do about that shit?"

elmerfud 4y ago

When I mirror people's behavior they get mad. When a chatbot does it, the bot must be broken. Staring in to the mirror is staring in to the abyss. People are not ready to see their true soul. Often times people don't even realize their casual lies. I'm not talking about mistakes in understanding, I'm talking about intentionally saying thing they know not to be true. We often call these "white lies" and it's done to spare feeling or avoid confrontation or really it's done because people lack fortitude to speak truth or not answer. Of course many times the asker is a fault too. We sometimes ask questions we should know better than asking. I tried to impress that idea of, only ask if you want to know, upon my children. Kids ask anything often times parents should moderate their answers. Once my kids were in their adolescence I let them know that they need to think about these things. They are mature enough to handle adult answers about things and if they ask me something I'll answer them directly and truthfully. There was a bit of a transition phase where they would ask something without thinking and I'd have to look at them and ask if they really wanted to know that answer. Usually they'd think about about it and withdrawal the question. My experience with AI is that it still has the emotional maturity of a small child not able to comprehend or regulate such matters. Because that kind of comprehension is algorithmically difficult. So you 5 year old with massive amounts of knowledge regurgitating it back trying mimic it's creators. Like a child it can't always understand fact and fiction. It really doesn't understand white lie vs lie lie. It's a imperfect mirror, but it's a mirror that perfect at magnifying our faults that we have collectively considered "normal".

jaybrendansmith 4y ago

These algorithms may approximate the language center of a human brain, but they cannot attach real axiomatic meaning to any of these words. So much of our IQ testing is focused on the language use, but that's simply the communication pathway for ideas, not the ideas themselves. How many times have I struggled for the right words to explain a concept? That's my language model attempting to communicate my axiomatic, emotional, spacial, conceptual, or mathematical mind. Words often fail. These chatbots are essentially idiot-savants in a very real sense, but they are clever enough to sometimes explain the reasons why they told an untruth or made an error. Trying to explain to them their error is foolish, because there's no 'mind' behind the words.

neonate 4y ago

https://archive.ph/VB5PQ https://archive.ph/VB5PQ