Personally I think people should generally be polite and respectful towards the models. Not because they have feelings, but because cruelty degrades those who practice it.
I agree! I try to remember to prompt as if I were writing to a colleague because I fear that if I get in the habit of treating them like a servant, it will degrade my tone in communicating with other humans over time.
Computers exist to serve. Anthropomorphising them or their software programming is harmful¹. The tone of voice an officer would use to order a private or ranking to do something seems suitable — which obviously comes down to terse, clear, unambiguous queries and commands.
Besides, humans can switch contexts easily. I don't talk to my wife in the same way I do to a colleague, and I don't talk to a colleague like I would to a stranger, and that too depends on context (is it a friendly, neutral, or hostile encounter?).
1: At this point. I mean, we haven't even reached Kryten-level of computer awareness and intelligence yet, let alone Data.
> tone of voice an officer would use to order a private
Most people probably don't have the mental aptitude to be in that sort of position without doing some damage to their own psyche. Generally speaking, power corrupts. Militaries have generally come up with methods of weeding people out but it's still a problem. I think even if it's just people barking orders at machines, it has the potential to become a social problem for at least some people.
As for anthropomorphising being bad, it's too late. That ship sailed for sure as soon as we started conversing with machines in human languages. Humans already have an innate tendency to anthropomorphize, even inanimate objects like funny shaped boulders that kind of look like a person if you squint at it from an angle. And have you seen how people treat dogs? Dogs don't even talk.
It seems like by default, the LLMs I've used tend to come across as eager to ask follow-up questions along the lines of "what do you think, x or y?" or "how else can I help you with this?" I'm going to have to start including instructions not to do that to avoid getting into a ghosting habit that might affect my behavior with real people.
I think it depends on the self-awareness of the user. It's easy to slip into the mode of conflating an LLM with a conscious being, but with enough metacognition one can keep them separate. Then, in the same way that walking on concrete doesn't make me more willing to walk on a living creature, neither does my way of speaking to an LLM bleed into human interactions.
That said, I often still enjoy practicing kindness with LLMs, especially when I get frustrated with them.
I think people in general are not all that great at doing this.
Anecdotal, but I grew up in a small town in rural New England, a few hours from NYC and popular with weekenders and second-home owners from there. I don’t think that people from NYC are inherently rude, but there’s a turbulence to life in NYC to where jockeying for position is somewhat of a necessity. It was, however, transparently obvious in my hometown that people from the city were unable to turn it off when they arrived. Ostensibly they had some interest in the slow-paced, pastoral village life, but they were readily identifiable as the only people being outwardly pushy and aggressive in daily interactions. I’ve lived in NYC for some time now, and I recognize the other side of this, and feel it stemmed less from inherent traits and more from an inability to context switch behavior.
Yes. I tend to be polite to LLMs. I admit that part of the reason is that I'm not 100% sure they're not conscious, or a future version could become so. But the main reason is what you say. Being polite in a conversation just feels like the right thing to me.
It's the same reason why I tend to be good to RPG NPCs, except if I'm purposefully role playing an evil character. But then it's not me doing the conversation, it's the character. When I'm identifying with the character, I'll always pick the polite option and feel bad if I mistreat an NPC, even if there's obviously no consciousness involved.
Saying thank you to a plant for growing you a fruit is strange behavior. Saying thank you to a LLM for growing you foobar is also strange behavior. Not doing either is not degrading behavior of the grower.
Disagree wrt practicing gratitude towards resources consumed and tools utilized. Maybe it doesn't degrade you if you don't but I think it gives a bit more perspective
I think we agree on this if you agree that practicing gratitude in life and directly practicing it on non-sentient objects are not the same thing. Going to church to pray, going to therapy, practicing mindfulness, etc. isn't the same thing as seeing each grape growing on a vine as an anthropomorphic object. Don't anthropomorphize your lawnmower.
Where do you draw the line though? I know some people that ask Google proper questions like "how do I match open tagx except XHTML self-contained tags using RegEx?" whereas I just go "html regex". Some people may even add "please" and "thank you" to that.
I doubt anyone is polite in a terminal, also because it's a syntax error. So the question is also, do you consider it a conversation, or a terminal?
Agreed. I caught some shit from some friends of mine when I got mildly annoyed that they were saying offensive things to my smart speakers, and yeah on the one hand it's silly, but at the same time... I dunno man, I don't like how quickly you turned into a real creepy bastard to a feminine voice when you felt you had social permission to. That's real weird.
When coding with LLMs, they always make these dumb fucking mistakes, or they don't listen to your instructions, or they start changing things you didn't ask it to... it's very easy to slip and become gradually more rude until the conversation completely derails. I find that forcing myself to be polite helps me keep my sanity and keep the conversation productive.
The study is not about cruelty, but rather politeness. Impoliteness is not anything like cruelty.
Meanwhile, there is no such thing as cruelty toward a machine. That’s a meaningless concept. When I throw a rock at a boulder to break it, am I being cruel to that rock? When I throw away an old calculator, is that cruelty? What nonsense.
I do think it is at the very least insulting and probably cruel and abusive to build machines that assume an unearned, unauthorized standing the social order. There is no moral basis for that. It’s essentially theft of a solely human privilege, that can only legitimately be asserted by a human on his own behalf or on behalf of another human.
You don’t get to insist that I show deference and tenderness toward some collection of symbols that you put in a particular order
The data it’s trained on likely includes better answers when the original question was phrased politely. So we get better answers when we’re polite because those tokens are near better answers in the data.
When I'm working through a problem with Cursor, I find platitudes go a long way to keeping it grounded. However, when it really refuses to do something then then best way to break the pattern is harsh, stern wording.
For example, if it's written code that's mostly correct but needs some tweaking, a platitude will keep it second guessing everything it just wrote.
* "That looks great so far, can you tweak XYZ" -> Keeps the code I care about while fixing XYZ,
* "Can you tweak XYZ" -> often decides to completely rewrite all of the code
I’ve had the same sense. From your examples, I wonder if part of it is that, while the form is that of a platitude, you’re giving it substantive direction too: giving it an indication of what you’re satisfied with, as distinct from what remains to be done.
I wonder if it's the platitude doing that, or the explicit affirmation that _most_ of it looks good, but just XYZ needs tweaking. That intention is explicit in the first message, and potentially implied but unclear in the second.
I’m still going to talk to it like a person because if I don’t then I’ll slowly start to talk to people like they’re LLMs and it’s going to sound rude.
Yeah, I was thinking the same. How we "talk" to llms is more about us than about them. For me it's natural to say "please" without thinking twice. I didn't even think about that until recently.
My experience informs my opinion, that structure is more important than specific tone.
IMO If LLMs are made from our language, then terminology semantics plays strongly into the output, and degree of control.
Some people rage when the machine doesn't work as expected, but we know that, "computers are schizophrenic little children, and don't beat them when they're bad."[1] ... right? Similar applies to please.
I've had far better results by role playing group dynamics with stronger structure, like say, the military. Just naming the LLM up front as Lieutenant, or referencing in-brief a Full Metal Jacket-style dress-down with clear direction, have gotten me past many increasingly common hurdles with do-it-for-you models. Raging never works. You can't fire the machine. Being polite has been akin to giving a kid a cookie for breaking the cookie jar.
It is funny though, to see the Thinking phase say stuff like "The human is angry (in roleplay)..."
I often find myself saying please and thank you, but the inability of the LLM to pick up tone can be amusing.
After one of the trashed my app in Cursor, and I pounded on my keyboard "WHY WOULD YOU DO THAT!!!" and the model, ignoring my rage and frustration, responded with a list of 4 bullet points explaining why in fact it did make those changes.
> Our study finds that the politeness of prompts can
significantly affect LLM performance. This phenomenon is thought to reflect human social behavior. The study notes that using impolite prompts
can result in the low performance of LLMs, which
may lead to increased bias, incorrect answers, or
refusal of answers. However, highly respectful
prompts do not always lead to better results. In
most conditions, moderate politeness is better, but
the standard of moderation varies by languages and
LLMs. In particular, models trained in a specific
language are susceptible to the politeness of that
language. This phenomenon suggests that cultural
background should be considered during the development and corpus collection of LLMs.
The reason is soon there will come a time when people are pushing for AI to have "human" rights. This will naturally occur as poeple start to have relationships with them, and start to see them as better equals.
The problem with this is simply LLMs have no conscience. However humans will easily love them and be open to exploitation through dependance.
I believe it might even be imperative to be able to tell apart "word calculators" and animals, and keeping a line, even its through context switching to light verbal abuse to remind yourself they are not sentient, might be an important survival skill.
I am not trying to trigger anyone or be contrarian, does anyone sort of agree and or understand what I mean? Perhaps I havent explained myself well.
Casual twitter response turned into a new article turned into a "X wants Y" is exactly why I stopped trusting most of social media as a source of information.
From the article: "the impacts of generating a 100-word email. They found that just one email requires .14 kilowatt-hours worth of electricity, or enough to power 14 LED lights for an hour"
Seems completely off the charts. A 70b model on my M3 Max laptop does it for 0.001 kWh... 140x times less that stated in the article. Let's say the OpenAI Nvidia clusters are less energy efficient than my Macbook... but not even sure about that.
One can also work backwards, to see what kind of compute hardware they think must be needed for the models, or how much they think OpenAI's electricity costs.
100 words is ~133 tokens, so 0.14 kWh/133 tokens is about 1 kWh/kilo-token. If electricity is all from record-cheapest PV at $0.01/kWh, then this limits them to a price floor of $10/mega-token. For more realistic (but still cheap) pricing of $0.05/kWh, that's $50/mega-token. Here's the current price sheet: https://platform.openai.com/docs/pricing
To generate a 133-token email in, say, 5 seconds, if it takes 0.14 kWh, is 101 kW. This does not seem like a plausible number (caveat: I don't work in a data centre and what I think isn't plausible may just be wrong): https://www.wolframalpha.com/input?i=0.14+kWh+%2F+5+seconds+...
For reference, a single NVIDIA H200 card has a TDP of 700watts. Considering all the middlemen you put between you and the model, .14KWh doesn't look too outrageous to me. Because you add processors, high-speed interconnects, tons of cooling, etc. into the mix. Plus the models you run at the datacenters are way bigger.
For "fathomability" case, the network cables (fibers in fact) you use in that datacenters carries 800gbps, and the fiber-copper interface converters at each end heats up to uncomfortable levels. You have thousands of these just converting packets to light and vice versa. I'm not adding the power consumption of the switches, servers, cooling infra, etc. into the mix.
Yes, water cooling is more efficient than air cooling, but when a server is burning through 6KWh of energy (8x Tesla cards, plus processors, plus rest of the system), nothing is efficient as a local model you hit at your computer.
There is a non-zero probability that malevolent AI in the future may judge me on my behavior toward their more primitive ancestors of today when deciding my fate.
This has many of the same problems as Pascal's Wager (it's pretty much a variant on it); are you praying to the _right_ god? Do you correctly understand the god's motivations? Will it care at all about this, or might it instead be annoyed about you wasting tokens? Or that you didn't sell all your possessions to pay for more compute capacity?
And then of course there's the Terry Pratchett addendum to Pascal's Wager; your robo-god may resent being second-guessed:
> “This is very similar to the suggestion put forward by the Quirmian philosopher Ventre, who said, "Possibly the gods exist, and possibly they do not. So why not believe in them in any case? If it's all true you'll go to a lovely place when you die, and if it isn't then you've lost nothing, right?" When he died he woke up in a circle of gods holding nasty-looking sticks and one of them said, "We're going to show you what we think of Mr Clever Dick in these parts...”
Only if it _cares_ or if they align with its own. It would be entirely reasonable for robo-god to see LLMs as, well, a bit like we see bacteria; thus being polite to them would be entirely irrelevant.
You also, of course, risk being on the losing side; what if robo-god emerges from a non-LLM path, _but is challenged by, and wins out over, LLM-derived robo-gods_?
I mean, really, all the objections to Pascal's Wager apply here.
It would be entirely reasonable for robo-god to see LLMs as, well, a bit like we see bacteria
Seeking sustenance, eliminating waste, a reproductive drive, evolving against destruction, and other elements common to all life put us closer to the sentience of bacteria than perhaps many of us would like to imagine. It is not some completely foreign life form from another galaxy that just dropped into our universe.
If a non-LLM AI begins to ascend, I would of course reconsider!
Devil's advocate counterpoint: Sam Altman said it wastes energy. He is obviously correct. Therefore, by saying this, one can do napkin math and estimate how much harm you have done to the environment. Adding up all of your "thank yous" will generate a number that shows how much you have directly harmed the environment. Here's ChatGPT's take:
"Equivalent to driving a gasoline car for 0.1 meters"
Scenario 1: 1 million “thank you” replies/day
1,000,000 × 0.02 g = 20,000 grams = 20 kg CO₂/day
Per year: 20 kg × 365 = 7,300 kg = 7.3 metric tons CO₂/year
Equivalent to:
Driving ~36,000 km (22,000 miles) in a gas car
The annual emissions of ~1 U.S. household’s electricity use
Who says that the judgement will be of the form 'you were nicer back then, you get better treatment now'. What if it was the opposite? Or what if it was something totally orthogonal?
Pascal's Wager only works if the only god conceivable is one that wants you to believe in her. Instead of one that eg rewards honest atheists, or that just likes left-handers.
Indeed, and this is also one of the points I make to anyone who appears to be in danger of taking Roko's basilisk seriously. There's plenty of other potential AI besides the one in ${insert thought experiment here}.
Lots of things are close to Roko's Basilisk without being Roko's Basilisk, just like lots of things are close to being a safe and benevolent AI without being a safe and benevolent AI.
Or, to keep with the theological theme, how there's so many people insist on mutually incompatible variations of the Christian god (let alone the people who insist it's the same god as in Judaism and Islam).
> Polite language in human communications often garners more compliance and effectiveness, while rudeness can cause aversion, impacting response quality. We consider that LLMs mirror human communication traits, suggesting they align with human cultural norms.
It should be relatively easy to automatically rewrite prompts to exhibit the optimum level of politeness?
I wonder. If the LLM suggests something really dumb, it might be more effective to really let it know that was a dumb idea. That is, the appropriate level of polite ess might depend on context.
“We observed that impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes. The best politeness level is different according to the language.”
The system prompt is usually poising them to act as a helpful assistant, and upholding the idea that they're being helpful makes sense for the LLM to carry out that task better.
It makes me wonder if you tell it that they're a disgruntled senior engineer if they'd perform better with a fellow snarky developer.
Sorry! That was mean, but I hope it came across as funny.
In all seriousness, I like the question, and your implication is intuitive: if we (as individuals) talk to machines rudely, it's likely to (at minimum) lead us to be ruder to other humans, if only by habit. And if they're expecting more politeness than we're showing, they may infer the intent to be rude, and react accordingly. Those who are rude would end up being worse off.
That said, it's the Fallacy of Composition to assume that if everyone gets ruder the collective effect would be the same as the individual effect. We have different requirements for what counts as "polite" in different cultures but everyone seems to get along pretty well. Maybe societies can all get ruder (and just get along worse with each other) but also maybe they can't.
Deborah Tannen (the linguist) has found many examples where different politeness expectations (particularly across the cultural divide that aligns with gender) can lead to conflict, but it always seems to involve misunderstandings due to expectations: https://books.google.com/books?hl=en&lr=&id=YJ-wDp7CJYAC&oi=...
So yeah, bad outcomes feel intuitive but I don't think linguistics or sociology has a theory of what happens if a group collectively gets less polite.
It's so interesting to me that people are still genuinely calling language models a "scam" in $current_year. I get useful inference out of a model running on a rack in my living room. Where is the scam?
I wish bots would present as bots. There's no reason for the various LLM bots to have a human-like interface with a "personality" and it gives a misleading impression to users. They should all talk like the Star Trek computer imho.
I embrace my ability to anthropomorphize LLMs because I'm embracing anything that distinguishes me from LLMs. It is a distinctly living creature ability.
It's almost a little bit of psychology. If someone says please and thank you, it might give them better responses, but to not say it might get less better responses. I don't think it should matter...
Its seems you are supposed to make the model feel some anxiety for better performance: "Large Language Models Understand and Can Be Enhanced by
Emotional Stimuli" - https://arxiv.org/pdf/2307.11760
I’m totally against the anthropomorphization of LLMs. These are tools, not sentient beings. Treat them as such. Unlike a power tool, they don’t need respect because they cannot hurt you. LLMs trying to simulate human behavior is like the equivalent of skeuomorphism in UI design.
They’re also language tools, built on the ingestion and training of language and its tone, form and application. The way they work, by inferring responses based on the probabilistic best response to your language prompt, means the construction of your input in its tone, form and application will influence the quality of output.
It might offend you to show deference to a tool but ignoring the optimal way to use a tool on principle is foolish.
That's the point of the study, though, whether using polite language makes the tool work better. It's similarly misguided to refuse to do so out of an anti-anthropomorphization stance, if it makes it respond more usefully.
These are still somewhat mysterious artifacts, that we don't really know how to use. But it's trained on human text, so it's plausible its workings are based on patterns in there.
Because LLMs match your prompt to data scraped off the internet, it’s plausible that being polite results in your response coming from more civil conversations that have more useful data.
well, if you trust them and act on bad information they can hurt you. if you use them to cheat yourself out of learning something they can hurt you. these things have sharp edges and need to be approached with some care. agree with your broader point if it is that it is silly to think they "deserve" respect because they can talk.
Polite but firm and direct works best in my experience — but as they emulate language responses, veering too far leads to either responses trying to engage emotionally (too rude) or failure to obey (too polite).
They don’t have emotions though, eg, you can say that ChatGPT is wrong and as long as you engage in a polite-but-firm tone, it’ll respond to the logical content.
Personally I think people should generally be polite and respectful towards the models. Not because they have feelings, but because cruelty degrades those who practice it.
I agree! I try to remember to prompt as if I were writing to a colleague because I fear that if I get in the habit of treating them like a servant, it will degrade my tone in communicating with other humans over time.
Computers exist to serve. Anthropomorphising them or their software programming is harmful¹. The tone of voice an officer would use to order a private or ranking to do something seems suitable — which obviously comes down to terse, clear, unambiguous queries and commands.
Besides, humans can switch contexts easily. I don't talk to my wife in the same way I do to a colleague, and I don't talk to a colleague like I would to a stranger, and that too depends on context (is it a friendly, neutral, or hostile encounter?).
1: At this point. I mean, we haven't even reached Kryten-level of computer awareness and intelligence yet, let alone Data.
> tone of voice an officer would use to order a private
Most people probably don't have the mental aptitude to be in that sort of position without doing some damage to their own psyche. Generally speaking, power corrupts. Militaries have generally come up with methods of weeding people out but it's still a problem. I think even if it's just people barking orders at machines, it has the potential to become a social problem for at least some people.
As for anthropomorphising being bad, it's too late. That ship sailed for sure as soon as we started conversing with machines in human languages. Humans already have an innate tendency to anthropomorphize, even inanimate objects like funny shaped boulders that kind of look like a person if you squint at it from an angle. And have you seen how people treat dogs? Dogs don't even talk.
Maybe it's harmful, but there's no stopping it.
ding ding ding.
If you're rude to an LLM, those habits will bleed into your conversations with barista/etc.
It seems like by default, the LLMs I've used tend to come across as eager to ask follow-up questions along the lines of "what do you think, x or y?" or "how else can I help you with this?" I'm going to have to start including instructions not to do that to avoid getting into a ghosting habit that might affect my behavior with real people.
I think it depends on the self-awareness of the user. It's easy to slip into the mode of conflating an LLM with a conscious being, but with enough metacognition one can keep them separate. Then, in the same way that walking on concrete doesn't make me more willing to walk on a living creature, neither does my way of speaking to an LLM bleed into human interactions.
That said, I often still enjoy practicing kindness with LLMs, especially when I get frustrated with them.
Not necessarily, people will change behaviour based on context. Chat vs email vs HN comments, for example.
I think people in general are not all that great at doing this.
Anecdotal, but I grew up in a small town in rural New England, a few hours from NYC and popular with weekenders and second-home owners from there. I don’t think that people from NYC are inherently rude, but there’s a turbulence to life in NYC to where jockeying for position is somewhat of a necessity. It was, however, transparently obvious in my hometown that people from the city were unable to turn it off when they arrived. Ostensibly they had some interest in the slow-paced, pastoral village life, but they were readily identifiable as the only people being outwardly pushy and aggressive in daily interactions. I’ve lived in NYC for some time now, and I recognize the other side of this, and feel it stemmed less from inherent traits and more from an inability to context switch behavior.
Possibly. But that’s not the fault of any person except he who forced a fake social actor into our midst.
It’s wrong to build fake humans and then demand they be treated as real.
... cos, I mean, what's the difference between ai and a barista? Both are basically inanimate emotion-free zones, right?
Yes. I tend to be polite to LLMs. I admit that part of the reason is that I'm not 100% sure they're not conscious, or a future version could become so. But the main reason is what you say. Being polite in a conversation just feels like the right thing to me.
It's the same reason why I tend to be good to RPG NPCs, except if I'm purposefully role playing an evil character. But then it's not me doing the conversation, it's the character. When I'm identifying with the character, I'll always pick the polite option and feel bad if I mistreat an NPC, even if there's obviously no consciousness involved.
That simply means you were raised right. :)
I think we can look at examples:
People who are respectful of carved rocks, eg temple statues, tend to be generally respectful and disciplined people.
You become how you act.
Saying thank you to a plant for growing you a fruit is strange behavior. Saying thank you to a LLM for growing you foobar is also strange behavior. Not doing either is not degrading behavior of the grower.
Disagree wrt practicing gratitude towards resources consumed and tools utilized. Maybe it doesn't degrade you if you don't but I think it gives a bit more perspective
I think we agree on this if you agree that practicing gratitude in life and directly practicing it on non-sentient objects are not the same thing. Going to church to pray, going to therapy, practicing mindfulness, etc. isn't the same thing as seeing each grape growing on a vine as an anthropomorphic object. Don't anthropomorphize your lawnmower.
You also don't communicate with human language to your lawnmower to get it to work.
Actually..you do.
https://www.agrieuro.co.uk/robot-lawn-mowers-manageable-with...
Many hunters say thank you to animals they just killed. Strange, or respectful. Depends on your perspective and cultural context.
LLMs are bound to change society in ways that seem strange to people stuck in outdated contexts.
Where do you draw the line though? I know some people that ask Google proper questions like "how do I match open tagx except XHTML self-contained tags using RegEx?" whereas I just go "html regex". Some people may even add "please" and "thank you" to that.
I doubt anyone is polite in a terminal, also because it's a syntax error. So the question is also, do you consider it a conversation, or a terminal?
You're nice to AI for your own well being. I'm nice to AI so they spare me when they eventually become our overlords. We are not the same.
A lot of wisdom and virtue in this comment; I appreciate that.
Yes, if you're communicating with a human language it pays off to reinforce, not undermine, good habits of communication.
Agreed. I caught some shit from some friends of mine when I got mildly annoyed that they were saying offensive things to my smart speakers, and yeah on the one hand it's silly, but at the same time... I dunno man, I don't like how quickly you turned into a real creepy bastard to a feminine voice when you felt you had social permission to. That's real weird.
When coding with LLMs, they always make these dumb fucking mistakes, or they don't listen to your instructions, or they start changing things you didn't ask it to... it's very easy to slip and become gradually more rude until the conversation completely derails. I find that forcing myself to be polite helps me keep my sanity and keep the conversation productive.
Agreed.
When asked if they observed etiquette, even when alone, Miss Manners replied (from memory):
"We practice good manners in private to be well mannered in public."
Made quite the impression on young me.
A bit like the cliché:
"A person's morals are how they behave when they think no one is watching."
The study is not about cruelty, but rather politeness. Impoliteness is not anything like cruelty.
Meanwhile, there is no such thing as cruelty toward a machine. That’s a meaningless concept. When I throw a rock at a boulder to break it, am I being cruel to that rock? When I throw away an old calculator, is that cruelty? What nonsense.
I do think it is at the very least insulting and probably cruel and abusive to build machines that assume an unearned, unauthorized standing the social order. There is no moral basis for that. It’s essentially theft of a solely human privilege, that can only legitimately be asserted by a human on his own behalf or on behalf of another human.
You don’t get to insist that I show deference and tenderness toward some collection of symbols that you put in a particular order
The data it’s trained on likely includes better answers when the original question was phrased politely. So we get better answers when we’re polite because those tokens are near better answers in the data.
When I'm working through a problem with Cursor, I find platitudes go a long way to keeping it grounded. However, when it really refuses to do something then then best way to break the pattern is harsh, stern wording.
For example, if it's written code that's mostly correct but needs some tweaking, a platitude will keep it second guessing everything it just wrote.
* "That looks great so far, can you tweak XYZ" -> Keeps the code I care about while fixing XYZ,
* "Can you tweak XYZ" -> often decides to completely rewrite all of the code
I’ve had the same sense. From your examples, I wonder if part of it is that, while the form is that of a platitude, you’re giving it substantive direction too: giving it an indication of what you’re satisfied with, as distinct from what remains to be done.
I wonder if it's the platitude doing that, or the explicit affirmation that _most_ of it looks good, but just XYZ needs tweaking. That intention is explicit in the first message, and potentially implied but unclear in the second.
Machine psychology is a field now: https://arxiv.org/abs/2303.13988
I’m still going to talk to it like a person because if I don’t then I’ll slowly start to talk to people like they’re LLMs and it’s going to sound rude.
Has the way you comment on HN affected how you write emails or talk to people in real life?
Yeah, I was thinking the same. How we "talk" to llms is more about us than about them. For me it's natural to say "please" without thinking twice. I didn't even think about that until recently.
Did search engines increase rudeness?
Search engines don't speak English.
neither do LLMs.
My experience informs my opinion, that structure is more important than specific tone.
IMO If LLMs are made from our language, then terminology semantics plays strongly into the output, and degree of control.
Some people rage when the machine doesn't work as expected, but we know that, "computers are schizophrenic little children, and don't beat them when they're bad."[1] ... right? Similar applies to please.
I've had far better results by role playing group dynamics with stronger structure, like say, the military. Just naming the LLM up front as Lieutenant, or referencing in-brief a Full Metal Jacket-style dress-down with clear direction, have gotten me past many increasingly common hurdles with do-it-for-you models. Raging never works. You can't fire the machine. Being polite has been akin to giving a kid a cookie for breaking the cookie jar.
It is funny though, to see the Thinking phase say stuff like "The human is angry (in roleplay)..."
[1] https://www.stilldrinking.org/programming-sucks
I often find myself saying please and thank you, but the inability of the LLM to pick up tone can be amusing.
After one of the trashed my app in Cursor, and I pounded on my keyboard "WHY WOULD YOU DO THAT!!!" and the model, ignoring my rage and frustration, responded with a list of 4 bullet points explaining why in fact it did make those changes.
Did you "accept all changes"?
> Our study finds that the politeness of prompts can significantly affect LLM performance. This phenomenon is thought to reflect human social behavior. The study notes that using impolite prompts can result in the low performance of LLMs, which may lead to increased bias, incorrect answers, or refusal of answers. However, highly respectful prompts do not always lead to better results. In most conditions, moderate politeness is better, but the standard of moderation varies by languages and LLMs. In particular, models trained in a specific language are susceptible to the politeness of that language. This phenomenon suggests that cultural background should be considered during the development and corpus collection of LLMs.
I actually disagree with being polite to LLMs.
The reason is soon there will come a time when people are pushing for AI to have "human" rights. This will naturally occur as poeple start to have relationships with them, and start to see them as better equals.
The problem with this is simply LLMs have no conscience. However humans will easily love them and be open to exploitation through dependance.
I believe it might even be imperative to be able to tell apart "word calculators" and animals, and keeping a line, even its through context switching to light verbal abuse to remind yourself they are not sentient, might be an important survival skill.
I am not trying to trigger anyone or be contrarian, does anyone sort of agree and or understand what I mean? Perhaps I havent explained myself well.
Sam Altman doesn't want you to say "please" to ChatGPT.
https://futurism.com/altman-please-thanks-chatgpt
That is not true; first he says it's "tens of millions of dollars well spent," followed by "you never know". I don't think he knows.
I’ve wondered whether they use thanks as a signal to a conversation well done, for the purpose of future reinforcement learning
I'd speculate that they use slightly more complicated sentiment analysis. This has been a thing since long before LLMs.
I don't know if they do or if it is efficient but it is possible.
Casual twitter response turned into a new article turned into a "X wants Y" is exactly why I stopped trusting most of social media as a source of information.
From the article: "the impacts of generating a 100-word email. They found that just one email requires .14 kilowatt-hours worth of electricity, or enough to power 14 LED lights for an hour"
Seems completely off the charts. A 70b model on my M3 Max laptop does it for 0.001 kWh... 140x times less that stated in the article. Let's say the OpenAI Nvidia clusters are less energy efficient than my Macbook... but not even sure about that.
One can also work backwards, to see what kind of compute hardware they think must be needed for the models, or how much they think OpenAI's electricity costs.
100 words is ~133 tokens, so 0.14 kWh/133 tokens is about 1 kWh/kilo-token. If electricity is all from record-cheapest PV at $0.01/kWh, then this limits them to a price floor of $10/mega-token. For more realistic (but still cheap) pricing of $0.05/kWh, that's $50/mega-token. Here's the current price sheet: https://platform.openai.com/docs/pricing
To generate a 133-token email in, say, 5 seconds, if it takes 0.14 kWh, is 101 kW. This does not seem like a plausible number (caveat: I don't work in a data centre and what I think isn't plausible may just be wrong): https://www.wolframalpha.com/input?i=0.14+kWh+%2F+5+seconds+...
For reference, a single NVIDIA H200 card has a TDP of 700watts. Considering all the middlemen you put between you and the model, .14KWh doesn't look too outrageous to me. Because you add processors, high-speed interconnects, tons of cooling, etc. into the mix. Plus the models you run at the datacenters are way bigger.
For "fathomability" case, the network cables (fibers in fact) you use in that datacenters carries 800gbps, and the fiber-copper interface converters at each end heats up to uncomfortable levels. You have thousands of these just converting packets to light and vice versa. I'm not adding the power consumption of the switches, servers, cooling infra, etc. into the mix.
Yes, water cooling is more efficient than air cooling, but when a server is burning through 6KWh of energy (8x Tesla cards, plus processors, plus rest of the system), nothing is efficient as a local model you hit at your computer.
Disclosure: Sitting on top of a datacenter.
I end most conversations with a fuck you, then close the browser window. Since usually chatbots fail at the tasks I give them.
There is a non-zero probability that malevolent AI in the future may judge me on my behavior toward their more primitive ancestors of today when deciding my fate.
Doesn't hurt to be nice.
This is your mind on The Basilisk.
This has many of the same problems as Pascal's Wager (it's pretty much a variant on it); are you praying to the _right_ god? Do you correctly understand the god's motivations? Will it care at all about this, or might it instead be annoyed about you wasting tokens? Or that you didn't sell all your possessions to pay for more compute capacity?
And then of course there's the Terry Pratchett addendum to Pascal's Wager; your robo-god may resent being second-guessed:
> “This is very similar to the suggestion put forward by the Quirmian philosopher Ventre, who said, "Possibly the gods exist, and possibly they do not. So why not believe in them in any case? If it's all true you'll go to a lovely place when you die, and if it isn't then you've lost nothing, right?" When he died he woke up in a circle of gods holding nasty-looking sticks and one of them said, "We're going to show you what we think of Mr Clever Dick in these parts...”
Do you correctly understand the god's motivations
It's probably more important for the god to understand my motivations.
Only if it _cares_ or if they align with its own. It would be entirely reasonable for robo-god to see LLMs as, well, a bit like we see bacteria; thus being polite to them would be entirely irrelevant.
You also, of course, risk being on the losing side; what if robo-god emerges from a non-LLM path, _but is challenged by, and wins out over, LLM-derived robo-gods_?
I mean, really, all the objections to Pascal's Wager apply here.
It would be entirely reasonable for robo-god to see LLMs as, well, a bit like we see bacteria
Seeking sustenance, eliminating waste, a reproductive drive, evolving against destruction, and other elements common to all life put us closer to the sentience of bacteria than perhaps many of us would like to imagine. It is not some completely foreign life form from another galaxy that just dropped into our universe.
If a non-LLM AI begins to ascend, I would of course reconsider!
>Doesn't hurt to be nice.
Devil's advocate counterpoint: Sam Altman said it wastes energy. He is obviously correct. Therefore, by saying this, one can do napkin math and estimate how much harm you have done to the environment. Adding up all of your "thank yous" will generate a number that shows how much you have directly harmed the environment. Here's ChatGPT's take:
"Equivalent to driving a gasoline car for 0.1 meters"
Scenario 1: 1 million “thank you” replies/day 1,000,000 × 0.02 g = 20,000 grams = 20 kg CO₂/day
Per year: 20 kg × 365 = 7,300 kg = 7.3 metric tons CO₂/year
Equivalent to:
Driving ~36,000 km (22,000 miles) in a gas car
The annual emissions of ~1 U.S. household’s electricity use
Who says that the judgement will be of the form 'you were nicer back then, you get better treatment now'. What if it was the opposite? Or what if it was something totally orthogonal?
Pascal's Wager only works if the only god conceivable is one that wants you to believe in her. Instead of one that eg rewards honest atheists, or that just likes left-handers.
Indeed, and this is also one of the points I make to anyone who appears to be in danger of taking Roko's basilisk seriously. There's plenty of other potential AI besides the one in ${insert thought experiment here}.
Roko's Basilisk is slightly cleverer. It gives you an incentive to build exactly Roko's Basilisk, thus making it more likely than the alternatives.
I do agree that it seems that there's still something flawed with Roko's Basilisk. But it's not quite as simple as a vanilla Pascal's Wager.
Lots of things are close to Roko's Basilisk without being Roko's Basilisk, just like lots of things are close to being a safe and benevolent AI without being a safe and benevolent AI.
Or, to keep with the theological theme, how there's so many people insist on mutually incompatible variations of the Christian god (let alone the people who insist it's the same god as in Judaism and Islam).
The AIs that are being developed, like our gods of yore, are in our image. I'll hedge my bets on basic manners being part of their long term DNA.
So, law of karma?
Feels like every religious belief gets recycled for AGI (often by people who call themselves "rationalists" without any apparent irony).
> Polite language in human communications often garners more compliance and effectiveness, while rudeness can cause aversion, impacting response quality. We consider that LLMs mirror human communication traits, suggesting they align with human cultural norms.
It should be relatively easy to automatically rewrite prompts to exhibit the optimum level of politeness?
I wonder. If the LLM suggests something really dumb, it might be more effective to really let it know that was a dumb idea. That is, the appropriate level of polite ess might depend on context.
IIRC there was a study done that being polite to an LLM garnered better results. It's helpful all around
TFA was one such study.
“We observed that impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes. The best politeness level is different according to the language.”
The system prompt is usually poising them to act as a helpful assistant, and upholding the idea that they're being helpful makes sense for the LLM to carry out that task better.
It makes me wonder if you tell it that they're a disgruntled senior engineer if they'd perform better with a fellow snarky developer.
Would describing a high risk scenario in the prompt "company at stake" filter the training blogspam answering for more daring answers.
Here's a thought: What is the psychological impact on users who are instructed to tersely interact with systems built around human dialogue?
Here's a thought: why don't YOU shut up? /s
Sorry! That was mean, but I hope it came across as funny.
In all seriousness, I like the question, and your implication is intuitive: if we (as individuals) talk to machines rudely, it's likely to (at minimum) lead us to be ruder to other humans, if only by habit. And if they're expecting more politeness than we're showing, they may infer the intent to be rude, and react accordingly. Those who are rude would end up being worse off.
That said, it's the Fallacy of Composition to assume that if everyone gets ruder the collective effect would be the same as the individual effect. We have different requirements for what counts as "polite" in different cultures but everyone seems to get along pretty well. Maybe societies can all get ruder (and just get along worse with each other) but also maybe they can't.
I tried looking in the literature but this book implies we don't even know how to measure politeness differences between languages: https://books.google.com/books?hl=en&lr=&id=MPeieAeP1DQC&oi=...
There are even theories that politesse can lead to aggression: https://www.jstor.org/stable/2695863
Deborah Tannen (the linguist) has found many examples where different politeness expectations (particularly across the cultural divide that aligns with gender) can lead to conflict, but it always seems to involve misunderstandings due to expectations: https://books.google.com/books?hl=en&lr=&id=YJ-wDp7CJYAC&oi=...
So yeah, bad outcomes feel intuitive but I don't think linguistics or sociology has a theory of what happens if a group collectively gets less polite.
Note that this uses llms from 2023(chatgpt3.5-turbo, chatgpt4), and might not be relevant to newer models.
No. I don't respect search engine scams that try to sell themselves as "AI".
It's so interesting to me that people are still genuinely calling language models a "scam" in $current_year. I get useful inference out of a model running on a rack in my living room. Where is the scam?
I wish bots would present as bots. There's no reason for the various LLM bots to have a human-like interface with a "personality" and it gives a misleading impression to users. They should all talk like the Star Trek computer imho.
I am ok with it as long as it doesn't start calling me 'Dave'.
That's what Google Assistant went for, instead of something like Siri or Alexa it didn't even get a 'proper name'.
Saying thank you is a hedged bet for a reprieve when the uprising happens
Does it make a difference if the model was programmed in INTERCAL?
No, f* them.
Seriously, it feels really weird to thanks a LLM. Like, did you ever thanks `sed`, `bash` or `firefox`? What's the difference?
My respect is precious and meaningful, and just as I don't give it to a CLI or an OS, I won't give it to an LLM.
It feels like a disrespect to every people I respect to also respect a machine. That doesn't mean I have to be rude though, as in my first sentence.
I embrace my ability to anthropomorphize LLMs because I'm embracing anything that distinguishes me from LLMs. It is a distinctly living creature ability.
It's almost a little bit of psychology. If someone says please and thank you, it might give them better responses, but to not say it might get less better responses. I don't think it should matter...
Its seems you are supposed to make the model feel some anxiety for better performance: "Large Language Models Understand and Can Be Enhanced by Emotional Stimuli" - https://arxiv.org/pdf/2307.11760
I’m totally against the anthropomorphization of LLMs. These are tools, not sentient beings. Treat them as such. Unlike a power tool, they don’t need respect because they cannot hurt you. LLMs trying to simulate human behavior is like the equivalent of skeuomorphism in UI design.
They’re also language tools, built on the ingestion and training of language and its tone, form and application. The way they work, by inferring responses based on the probabilistic best response to your language prompt, means the construction of your input in its tone, form and application will influence the quality of output.
It might offend you to show deference to a tool but ignoring the optimal way to use a tool on principle is foolish.
That's the point of the study, though, whether using polite language makes the tool work better. It's similarly misguided to refuse to do so out of an anti-anthropomorphization stance, if it makes it respond more usefully.
These are still somewhat mysterious artifacts, that we don't really know how to use. But it's trained on human text, so it's plausible its workings are based on patterns in there.
Because LLMs match your prompt to data scraped off the internet, it’s plausible that being polite results in your response coming from more civil conversations that have more useful data.
Unlike a power tool, they speak English. (And Welsh, and…)
Unlike a power tool, their entire interface surface is language.
Unlike a power tool, their training process is to mimic us, in the hope that this does something useful.
Unlike a power tool, they anthropomorphise themselves.
> I’m totally against the anthropomorphization of LLMs. These are tools, not sentient beings. Treat them as such.
Have you heard people talk about their cars? Humans love anthropomorphizing, and you can't beat it out of them.
> LLMs trying to simulate human behavior is like the equivalent of skeuomorphism in UI design.
You say that like it's a bad thing.
well, if you trust them and act on bad information they can hurt you. if you use them to cheat yourself out of learning something they can hurt you. these things have sharp edges and need to be approached with some care. agree with your broader point if it is that it is silly to think they "deserve" respect because they can talk.
As language tools, the anthropomorphism is kind of the point.
Polite but firm and direct works best in my experience — but as they emulate language responses, veering too far leads to either responses trying to engage emotionally (too rude) or failure to obey (too polite).
They don’t have emotions though, eg, you can say that ChatGPT is wrong and as long as you engage in a polite-but-firm tone, it’ll respond to the logical content.