5 ms·

Isn't it wild how seemingly overnight we've become comfortable speaking to AI as if it were human? Writing things like "Actually, one more thing." to ChatGPT fo

Isn't it wild how seemingly overnight we've become comfortable speaking to AI as if it were human? Writing things like "Actually, one more thing." to ChatGPT for example, are strictly unnecessary, but I write like this also even though it's more verbose than a simple "Yes" or even just rudely submitting the next thing I need help with. Fascinating.

ChrisRR 3y ago
It's hardly overnight, we've been talking to smart assistants for years.
mtgentry 3y ago
Yeah if ChatGPT helps me solve a hard problem I always tell it thanks.
- alchemist1e9 3y ago
  I’ve read multiple places that being polite and professional actually results in better output quality. I think recently there was even a paper published quantifying it.
  - Jenk 3y ago
    It is interpreted as positive/affirmative against the correctness or relevance of the previous output. Telling it 'That is correct' will achieve a very similar effect (as will saying 'No'/'That isn't relevant'/etc. would in the converse.)
  - function_seven 3y ago
    This cracks me up. For decades only INTERCAL demanded an appropriate level of politeness. It was a great joke, and we all laughed. Now it’s an actual real thing that the computer performs the task better when you ask nicely! (I am also polite and professional whenever I talk to ChatGPT)
  - bobbylarrybobby 3y ago
    I believe the rationale is that in the LLMs’ training data, higher-quality information is available in places where people are polite. The LLM is just copying that statistical fact.
ben_w 3y ago
It's trained on humans, so it makes sense that you get the best out of it by treating it in the way that gets the best out of humans.
lopatin 3y ago
I've wondered about this. If it's trained on human responses, and humans are more likely to be helpful to a polite question, then shouldn't ChatGPT also be more inclined to give helpful replies if I say please and thank you?
- geuis 3y ago
  It sounds strange, but I actually ran into a prompt problem a few weeks ago for a GPT project I'm working on. If I don't include "please" in the prompt, the response is the typical "I'm just an AI and can't do that". I had to leave a comment in the code to be aware of changing the prompt in the future, because if we aren't polite to the LLM it may not respond correctly.
  - gpderetta 3y ago
    Obviously too much INTERCAL in the training set. edit: someone else made the same joke 5 mins ago :D
  - mensetmanusman 3y ago
    Can you pre-prompt politeness? For example, ‘assume every question I am about to ask is phrased in an incredibly polite way.’
- AndrewKemendo 3y ago
  Unfortunately the data it’s trained on (internet text - commoncrawl) has an antisocial data bias, and there’s no alternative “pro-social” data corpuses that exist.[1] Every one of these has to be “taught” not to be toxic - usually by some version of exploited labor paid pennies to look at the worst stuff on the planet [2] It’s true of every larger model trained on internet corpus - Microsoft notoriously had this issue with Taybot [1] https://hci.stanford.edu/publications/2022/Park_ContentModAudit_CSCW22.pdf https://hci.stanford.edu/publications/2022/Park_ContentModAu... [2] https://time.com/6247678/openai-chatgpt-kenya-workers/ https://time.com/6247678/openai-chatgpt-kenya-workers/
  - mensetmanusman 3y ago
meiraleal 3y ago
> "Actually, one more thing." to ChatGPT for example, are strictly unnecessary I won't run a conclusive test about it but treating ChatGPT well gives me much better results.
- lib-dev 3y ago
  Someone should. I've noticed this phenomenon as well.
atleastoptimal 3y ago
People have been doing that for decades. People were treating ELIZA (from 1964) like a real human. Human treat their household appliances like other humans.
root_axis 3y ago
It's not "strictly" unnecessary since the LLM is responding with respect to its training, this type of phrasing could actually produce more desirable output if the training data favors conversational exchanges with this type of phrasing.
famouswaffles 3y ago
It's not like it has no effect. https://arxiv.org/abs/2307.11760 https://arxiv.org/abs/2307.11760

ChrisRR 3y ago

It's hardly overnight, we've been talking to smart assistants for years.

mtgentry 3y ago

Yeah if ChatGPT helps me solve a hard problem I always tell it thanks.

alchemist1e9 3y ago
I’ve read multiple places that being polite and professional actually results in better output quality. I think recently there was even a paper published quantifying it.
- Jenk 3y ago
  It is interpreted as positive/affirmative against the correctness or relevance of the previous output. Telling it 'That is correct' will achieve a very similar effect (as will saying 'No'/'That isn't relevant'/etc. would in the converse.)
- function_seven 3y ago
  This cracks me up. For decades only INTERCAL demanded an appropriate level of politeness. It was a great joke, and we all laughed. Now it’s an actual real thing that the computer performs the task better when you ask nicely! (I am also polite and professional whenever I talk to ChatGPT)
- bobbylarrybobby 3y ago
  I believe the rationale is that in the LLMs’ training data, higher-quality information is available in places where people are polite. The LLM is just copying that statistical fact.

ben_w 3y ago

It's trained on humans, so it makes sense that you get the best out of it by treating it in the way that gets the best out of humans.

lopatin 3y ago

I've wondered about this. If it's trained on human responses, and humans are more likely to be helpful to a polite question, then shouldn't ChatGPT also be more inclined to give helpful replies if I say please and thank you?

geuis 3y ago
It sounds strange, but I actually ran into a prompt problem a few weeks ago for a GPT project I'm working on. If I don't include "please" in the prompt, the response is the typical "I'm just an AI and can't do that". I had to leave a comment in the code to be aware of changing the prompt in the future, because if we aren't polite to the LLM it may not respond correctly.
- gpderetta 3y ago
  Obviously too much INTERCAL in the training set. edit: someone else made the same joke 5 mins ago :D
- mensetmanusman 3y ago
  Can you pre-prompt politeness? For example, ‘assume every question I am about to ask is phrased in an incredibly polite way.’
AndrewKemendo 3y ago
Unfortunately the data it’s trained on (internet text - commoncrawl) has an antisocial data bias, and there’s no alternative “pro-social” data corpuses that exist.[1] Every one of these has to be “taught” not to be toxic - usually by some version of exploited labor paid pennies to look at the worst stuff on the planet [2] It’s true of every larger model trained on internet corpus - Microsoft notoriously had this issue with Taybot [1] https://hci.stanford.edu/publications/2022/Park_ContentModAudit_CSCW22.pdf https://hci.stanford.edu/publications/2022/Park_ContentModAu... [2] https://time.com/6247678/openai-chatgpt-kenya-workers/ https://time.com/6247678/openai-chatgpt-kenya-workers/
- mensetmanusman 3y ago

meiraleal 3y ago

> "Actually, one more thing." to ChatGPT for example, are strictly unnecessary I won't run a conclusive test about it but treating ChatGPT well gives me much better results.

lib-dev 3y ago
Someone should. I've noticed this phenomenon as well.

atleastoptimal 3y ago

People have been doing that for decades. People were treating ELIZA (from 1964) like a real human. Human treat their household appliances like other humans.

root_axis 3y ago

It's not "strictly" unnecessary since the LLM is responding with respect to its training, this type of phrasing could actually produce more desirable output if the training data favors conversational exchanges with this type of phrasing.

famouswaffles 3y ago

It's not like it has no effect. https://arxiv.org/abs/2307.11760 https://arxiv.org/abs/2307.11760