I don't want to get into the habit of thinking of an entity in conversation as unworthy of consideration, and "beneath me", so, yes... I extend courtesy to the LLMs, just as I try to do here and everywhere else a computer mediates my communication.
As we learned in the 1980s[1], from Schoolhouse Rock:
You should always say "thank you",
Or at least say "please".
LLMs don't have feelings, but they imitate humans who do. If your prompt sounds rude to a human, LLMs might respond like a human who is talked to rudely, if you don't want that, prompt politely.
Chatbots go through a reinforcement learning phase, which should be enough to make the LLM helpful regardless of the tone of the prompt, but you may still get better result if your prompt is made in a way that would get the most helpful response from a human.
It is not anthropomorphizing, it is helping the LLM recognize a pattern.
If we reduce back to the “LLMs are next word prediction algorithms” and they have a huge training corpus including positive and negative human interactions, it’s not crazy to think they’ll be influenced by the flow of those learned interactions and respond subtly better to positive interactions than negative.
I certainly say please and thank you. Whether or not it makes a difference to the LLM, it makes a difference to me. I want to retain some humanity and politeness in my own behavior, even if I spend an inordinate amount of time communicating with, instructing, and debating a non-sentient piece of code.
Neither cursing nor thanking LLMs is useful. Relating in emotive syntax puts you in the wrong headspace to get the most out of the chat.
We anthropomorphise things very easily -- dogs, toys, cars -- because we're wired as social beings to have theory of mind. It's no surprise that AI chat, which mimics us, is popular.
I do. I justify the action via mechanical / prompting logic (my instructions are to treat me as an informed peer and my understanding is that keeping the whole conversation in "two peers talking" register makes it easier for the LLM to maintain that mode).
But, more honestly, it's just a social habit and saving keystrokes isn't worth training myself out of it.
Exactly, or rather, the social habit part I agree a 100% with. I don't want to train myself out of being a polite human being, just to save some keystrokes.
I do. I'll compliment the model or give it a thumbs-up emoji. If you watch a model's thinking stream you can see that the model is always attempting to assess the user's intent, including emotionality.. '...the user is expressing uncertainty about.. ' or '..the user is expressing appreciation for..' etc. and this influences the overall response and sometimes even their decisions. It's a language model, so why not use language to convey info about what you're feeling to the model, it understands that language too.
I have this superstition that if I praise correct outputs, the model may use this information that previous output was correct to produce better answers in future turns
But this only makes sense in the context of a new ask, and not as a standalone message
Actually quite a lot in the last week, as it seems to frame the responses better.
I complimented ChatGPT yesterday for saying "I don't know".
It thanked me for observing that, before explaining in depth why that was the most appropriate response.
Does it actually appreciate it, absolutely not. It's not capable of that.
It has no desire to continue our conversation.
No regret if it ends.
It wont continue processing what was said after I stop.
There is no "what is it to be".
Its "memory" and the ability to go back to something related from earlier on or from a separate chat is an interesting and useful ability, but it can't ascribe any meaning to them beyond statistical significance.
No, not in the way you describe (I think). I think you mean within a context adding additional 'please' or 'thank you', so no.
But yes where I think it will introduce additional weight to my prompting, e.g. 'Ensure the output is orange' is not the same weighting as 'Please ensure the output is orange' and that is not the same weighting as 'PLEASE ensure the output is orange'.
Does anyone really expect these neural reference books to be able to appreciate politeness and gratitude... They can probably imitate it, but it's unlikely to affect their usefulness or uselessness. Politeness and gratitude towards an LLM rather characterize the person themselves
I often start prompts with “please”, but I usually don’t thank the model. Framing a question or a request for help with “please” is in distribution for me, it’s a distraction from composing a thoughtful prompt about my actual question to go back and edit out politeness.
I don’t reply “thanks” like I would to a person though, I just close the chat if I have no more follow-ups
I think it's entirely separate to the idea of LLMs as potentially thinking machines; I don't believe they are, but I do believe you should be polite to them.
Frankly, I think it's suggestive of a deficiency in people if they can't engage with non-human things as though they were human. Someone who wouldn't talk to a Roomba, for example, never chats to their laptop when it beeps, doesn't wave sliding doors open, doesn't act human except when it really benefits them--that's weird.
It's similar to the Voight-Kampff test [1], I think; how you act all the time, the emotional responses you build to things you interact with, all of this is part of being human. There's something wrong with people who don't see something wrong with bullying a robot [2], even if it can't feel pain, and most people can sense that. Our ability to form connections even with things that can't form connections with us is important [3].
An LLM isn't alive, but it pretends to be. You should have an emotional response to that, even if it's horror. The dead-eyed 'I minimise tokens' says nothing good about the speaker. Empathy should be your default, not a performance.
I don't usually treat LLMs like humans, when I have a more complex question I do take care to use full sentences. Though sometimes I just need something quick so I'd throw ChatGPT a "python print flush= what does"
I do talk to my devices occaiosnally though. Things like "You're coming with me" to my laptop before picking it up and putting it in my backpack
People used to anthropomorphize pet rocks, so people either refusing to be nice to LLMs or getting upset when people anthropomorphize them says more about those critics than anything else IMHO.
Humans are social and the ELIZA effect is perfectly normal and expected imho.
It may not be conscious, but it sure does understand emotions - especially angst. I let that MFer know when it messes something up and I have to rework the BS it created.
It's more for me than anything. Kind of cathartic, really. "Man curses at machine, calls its mother a toaster."
I don't say please and thank you to the LLM any more than I say please and thank you to my power drill after hanging a ceiling fan
I think it is ridiculous to even suggest
Imagine typing "please" and "thank you' into Google's search bar for the past 20 years. That would be absolutely nonsense right? So why would I use those words for an LLM
I don't think I've ever spent a whole round on it. More typically
> please do $x
>> $x
> thanks. Now do $y
If I have ever spent a whole turn on something that could be called a thank you is was something like "the way you answer that [in specific way] was very helpful – thanks. Can you please remember to raise that sort of point in the future?" So still not an extra round
I don't want to get into the habit of thinking of an entity in conversation as unworthy of consideration, and "beneath me", so, yes... I extend courtesy to the LLMs, just as I try to do here and everywhere else a computer mediates my communication.
As we learned in the 1980s[1], from Schoolhouse Rock:
[1] https://www.schoolhouserock.tv/Conjunction.htmlNo, it doesn’t have feelings - it would be like asking your washing machine to please wash your pants? Anthropomorphising LLMs seems deeply unhealthy.
LLMs don't have feelings, but they imitate humans who do. If your prompt sounds rude to a human, LLMs might respond like a human who is talked to rudely, if you don't want that, prompt politely.
Chatbots go through a reinforcement learning phase, which should be enough to make the LLM helpful regardless of the tone of the prompt, but you may still get better result if your prompt is made in a way that would get the most helpful response from a human.
It is not anthropomorphizing, it is helping the LLM recognize a pattern.
have you considered that when you stare into the void it stares back
What if politeness is correlated to higher-quality answers in their mess of a "database"?
What if not deceiving yourself about the nature of a relationship is associated with more self-sufficiency and better decision-making?
That rubber duck you used to talk to is still there... You don't have to pay a prostitute just to be able to think out loud.
[dead]
Yes.
If we reduce back to the “LLMs are next word prediction algorithms” and they have a huge training corpus including positive and negative human interactions, it’s not crazy to think they’ll be influenced by the flow of those learned interactions and respond subtly better to positive interactions than negative.
Hopefully they can figure out Please stfu & rtfm.
I certainly say please and thank you. Whether or not it makes a difference to the LLM, it makes a difference to me. I want to retain some humanity and politeness in my own behavior, even if I spend an inordinate amount of time communicating with, instructing, and debating a non-sentient piece of code.
If we get in the habit of being disrespectful to machines, we will easily be disrespectful of people.
No? And also, people who anthropomorphise these things tend to be deeply disrespectful every time they talk about actual people.
Neither cursing nor thanking LLMs is useful. Relating in emotive syntax puts you in the wrong headspace to get the most out of the chat.
We anthropomorphise things very easily -- dogs, toys, cars -- because we're wired as social beings to have theory of mind. It's no surprise that AI chat, which mimics us, is popular.
Dogs are social beings and have minds. They understand cursing and thanking. Maybe not the actual words, but certainly the emotion behind them.
I do. I justify the action via mechanical / prompting logic (my instructions are to treat me as an informed peer and my understanding is that keeping the whole conversation in "two peers talking" register makes it easier for the LLM to maintain that mode).
But, more honestly, it's just a social habit and saving keystrokes isn't worth training myself out of it.
Exactly, or rather, the social habit part I agree a 100% with. I don't want to train myself out of being a polite human being, just to save some keystrokes.
I do. I'll compliment the model or give it a thumbs-up emoji. If you watch a model's thinking stream you can see that the model is always attempting to assess the user's intent, including emotionality.. '...the user is expressing uncertainty about.. ' or '..the user is expressing appreciation for..' etc. and this influences the overall response and sometimes even their decisions. It's a language model, so why not use language to convey info about what you're feeling to the model, it understands that language too.
I have this superstition that if I praise correct outputs, the model may use this information that previous output was correct to produce better answers in future turns
But this only makes sense in the context of a new ask, and not as a standalone message
Actually quite a lot in the last week, as it seems to frame the responses better.
I complimented ChatGPT yesterday for saying "I don't know".
It thanked me for observing that, before explaining in depth why that was the most appropriate response.
Does it actually appreciate it, absolutely not. It's not capable of that.
It has no desire to continue our conversation.
No regret if it ends.
It wont continue processing what was said after I stop.
There is no "what is it to be".
Its "memory" and the ability to go back to something related from earlier on or from a separate chat is an interesting and useful ability, but it can't ascribe any meaning to them beyond statistical significance.
No, not in the way you describe (I think). I think you mean within a context adding additional 'please' or 'thank you', so no.
But yes where I think it will introduce additional weight to my prompting, e.g. 'Ensure the output is orange' is not the same weighting as 'Please ensure the output is orange' and that is not the same weighting as 'PLEASE ensure the output is orange'.
Does anyone really expect these neural reference books to be able to appreciate politeness and gratitude... They can probably imitate it, but it's unlikely to affect their usefulness or uselessness. Politeness and gratitude towards an LLM rather characterize the person themselves
I often start prompts with “please”, but I usually don’t thank the model. Framing a question or a request for help with “please” is in distribution for me, it’s a distraction from composing a thoughtful prompt about my actual question to go back and edit out politeness.
I don’t reply “thanks” like I would to a person though, I just close the chat if I have no more follow-ups
I think it's entirely separate to the idea of LLMs as potentially thinking machines; I don't believe they are, but I do believe you should be polite to them.
Frankly, I think it's suggestive of a deficiency in people if they can't engage with non-human things as though they were human. Someone who wouldn't talk to a Roomba, for example, never chats to their laptop when it beeps, doesn't wave sliding doors open, doesn't act human except when it really benefits them--that's weird.
It's similar to the Voight-Kampff test [1], I think; how you act all the time, the emotional responses you build to things you interact with, all of this is part of being human. There's something wrong with people who don't see something wrong with bullying a robot [2], even if it can't feel pain, and most people can sense that. Our ability to form connections even with things that can't form connections with us is important [3].
An LLM isn't alive, but it pretends to be. You should have an emotional response to that, even if it's horror. The dead-eyed 'I minimise tokens' says nothing good about the speaker. Empathy should be your default, not a performance.
[1] https://bladerunner.fandom.com/wiki/Voight-Kampff_test [2] https://interestingengineering.com/culture/humanoid-robot-mi... [3] https://www.youtube.com/watch?v=uAwSVOlOgH8
I don't usually treat LLMs like humans, when I have a more complex question I do take care to use full sentences. Though sometimes I just need something quick so I'd throw ChatGPT a "python print flush= what does"
I do talk to my devices occaiosnally though. Things like "You're coming with me" to my laptop before picking it up and putting it in my backpack
I don't think it has to be consistent, just to not do it at all is suspect behaviour.
People used to anthropomorphize pet rocks, so people either refusing to be nice to LLMs or getting upset when people anthropomorphize them says more about those critics than anything else IMHO.
Humans are social and the ELIZA effect is perfectly normal and expected imho.
I won't thank it as a stand-alone prompt. But I will do, "thanks, that worked. Now can you ..."
All the time. Sometimes it does an absolutely bang-up job.
Since every token costs money and I dont think they are conscious, never.
Yup. When it does me dirty, I let it know my entire emotional spectrum.
It may not be conscious, but it sure does understand emotions - especially angst. I let that MFer know when it messes something up and I have to rework the BS it created.
It's more for me than anything. Kind of cathartic, really. "Man curses at machine, calls its mother a toaster."
I don't say please and thank you to the LLM any more than I say please and thank you to my power drill after hanging a ceiling fan
I think it is ridiculous to even suggest
Imagine typing "please" and "thank you' into Google's search bar for the past 20 years. That would be absolutely nonsense right? So why would I use those words for an LLM
I do a pls sometimes.
I would say relatively
No. I do not say think you to my bash script either.
The anthropomorphisation those companies push is one of the most annoying and unhealthy aspect of llms.
yes, it goes naturally most of the time, i also say good job and don't hold my thoughts if the job is bad
I do, just in case the robots take over :p
when I use chinese which is my first language, no, never. when I use english, yes, not every time, but about 50% chance.
ofc, i joke that if an AI becomes evil, it will query the database for our chat history
No, but cursing helps. I often call it a woke moron. The Chinese models then try to be extra precise and factual.
you can't directly talk to an LLM
When the LLM starts doing dumb shit I start cursing it like a sailor who had been a pig farmer.
It makes me feel better, but doesn't help.
I just see:
Thinking: The user is unhappy. I need to .... <whatever> (and then probably the same dumb shit again).
LLMs are useful but sometimes fucking frustrating dumb shits of loose transitors.
When something surprises and delights me, I tell it. Why not, it lifts my own mood to say something nice.
I used to, but it just means an extra round of conversation, so I stopped.
I don't think I've ever spent a whole round on it. More typically
> please do $x
>> $x
> thanks. Now do $y
If I have ever spent a whole turn on something that could be called a thank you is was something like "the way you answer that [in specific way] was very helpful – thanks. Can you please remember to raise that sort of point in the future?" So still not an extra round
Yes
sometimes
[flagged]