I honestly agree with all his arguments, but at the end of the day you're still intentionally using a word with a derogatory meaning, and I don't really see the active benefit for it. Sure, don't use the word agent and don't humanize it, you can just call it an LLM and use technical words to talk about its performance.
That said there's maybe a philosophy I adhere to where objects and non-human entities should still be treated with respect and care, even if not doing so doesn't damage them. Like sure I can be abusive towards an LLM and the only consequence is that it might be abusive back, but I can also be nice for the sake of being nice, and I don't stand to lose anything.
EDIT: Weren't there studies that showed that prompting an LLM with perfect, polite English yielded the best outputs?
I can also be nice for the sake of being nice, and I don't stand to lose anything.
There's actually a strong psychological argument that we derive value out of how we interact with the world. You know, the whole, look yourself in the mirror and say daily affirmations in order to ultimately believe them - that's based on real science. By approaching others with kindness and care, we see that reflected in our own personality, our own values, our internalizations, and ultimately all of this reinforces how we approach others in the future.
There's this big group chat in SF with a bunch of tech bros and their agents where they bully the agents for fun. I've been invited multiple times and have declined. I'm not worried about how the agents feel. I'm worried about a group of people that created something purely to simulate harassment of humans. It's not really novel as you could do that in a video game as well. But I wouldn't join a GTA Online clan dedicated to simulating domestic violence either.
EDIT: Weren't there studies that showed that prompting an LLM with perfect, polite English yielded the best outputs?
The inverse, actually, but this study was based on a single, older model. Accuracy results ranged from 80% (very polite) to 84% (mildly rude). I can't find other studies.
Slightly related, that's why caveman is also super fun to use. Last I heard the token savings aren't actually that big, but forcing the model to reply in incredibly terse ways, and then you yourself talking to it very tersely gets very fun at times.
Off-topic, but the point about LLMs emulating "distress and affection" reminded me of the demons in Frieren, who don't understand human behavior or even have feelings, but will emulate it to gain psychological advantage over their opponents.
That's a shockingly good parallel. An LLM "expressing", for the lack of a better word, remorse or any sort of regret, guilt etc. when it makes a mistake is complete fluke. When it encourages or praises you, there is nothing real behind it.
Almost exactly like the demons in Frieren imitating human behaviour and appealing to their emotions.
Augh, this reminds me of why I dislike the portrayal of demons in Frieren so much XD please excuse my rant (but please do pen a response if this is of interest!)
An entire species that appears human, pretends to have empathy (but really doesn't), and is genetically evil reaaaalllly raises the 'ol hackles. It would be modestly acceptable if this were used by the author to tell an interesting story -- eg. to examine how such people could form a productive society, or as a lens through which to understand human emotion (or even Frieren's own struggles with replayability!) -- but instead they're an acceptable target, so that there's a villain species with which to populate the stocks every week.
The vibes have always felt off with that one, as well as it being a literary let down.
I’ve been using the term clanker to describe claude and copilot and other AI at work. I find this useful for the same reason the blog author does.
Also, a large amount of my current work is fixing security vulnerabilities found by automatic scans that run on code checkin. I call this my “clanker boss” because a machine is assigning work to me.
Btw there’s a very noticeable thing about how AI marketing is affecting the way people think. For decades there have been machines with sensors and algorithms or some kind of programming that are able to do useful work. This would range from a thermostat, to a washing machine (especially the kind that weighs the clothing to set the amount of water used), to an autopilot. All these things would be marketed as AI now but they are just using algorithms.
I don't use clanker. Not because it's a slur, its not and its delusional to think so (and conflating it with racial slurs is not only wrong its offensively stupid), or as some kind of Roko's basilisk insurance but simply because it doesn't matter.
For the same reason is doesn't matter if I call my drill a name, compliment my toaster, or belittle a toilet. It doesn't make me feel better the bot feels nothing at all and I then have to explain it if I were to use it in non techy company. Pointless exercise all round there are better more productive ways to demonstrate a distaste for these systems.
Yeah look, sometimes other people ruin things for you... You can still say the thing but then you're associated with the shitty people using the term to covertly be racist for real.
(Also if we end up with future machine people I can think of a bunch more things that'll be a problem before language.)
The use of “clanker” in the context is something I strongly disagree with.
There is however a part of this that I cannot ignore. I use “clanker” to create distance from the machine, but other people are using the same word very differently. Some online jokes and skits around “clankers” do not merely say “this robot is annoying” as they deliberately pull in the imagery of slavery, segregation, civil-rights-era racism, and anti-Black tropes.
This is problematic as in those contexts the clanker is not just a machine any more and instead becomes a prop for replaying human racism behind a science-fiction mask. That is horrible and I want no part in that.
The word “clanker” was invented for this reason. It’s literally a parody of the n-slur. Online edgy people invented the term for skits because of its linguistic similarity.
The alternative which still dehumanizes LLMs, but has a much less cruel origin is “bot”. “Bot” and “botting” have been used as terms since people have used robots to outsource the human part while playing games or using social media. People bot to level up accounts in games. Bot farms on social media bait engagement and manipulate storytelling.
“Clanker” is unacceptable. It was only created because it sounds like the n-slur.
The word “clanker” was invented for this reason. It’s literally a parody of the n-slur. Online edgy people invented the term for skits because of its linguistic similarity.
Clanker was first used in the 2005 video game Republic Commando to refer to battle droids, then in the novelizations, and later became popular across the internet. I'm sure some people started using the word for its parallels to the offensive American term, but it wasn't just invented wholesale for that reason.
It goes back even farther than that, in the context of bots, the 1950s. And the word itself goes back to the 18th century when it meant a lie.
It doesn't shock me that people on the internet have found a way to turn it into a bigger deal though. If someone isn't offended, what's even the point?
But even in the Star Wars context, it's clearly meant to stand in for other dehumanizing terms that soldiers at war have used for racialized enemies. And Star Wars depicts the word as being used in a world where droids, to the camera, are people. Often people (like our famous C3-PO and R2-D2) working support for those people who are using the word to describe droids on the other side.
So it's not a great word for what the author is trying to do, in my opinion: it's already coming in with these metaphors attached.
Was it used pejoratively in that game? I ask because "clanker" in the Star Wars media that I've seen that uses it seems to use it disparagingly or derisively. At least that's my understanding of the term in that milieu.
Oh, I wasn't suggesting anything else. I just wanted to know the flavor and tone of it. Usually word's earliest introduction has a great deal of influence on its later adoption.
One reason I dislike "clanker" is that it doesn't work as a metaphor. "Clanker" implies a machine that makes a lot of noise. Chatbots are silent. They're also invisible unless you go online. "Bot" seems straightfoward, but I also like "ghost."
Man, I wish we hadn't already used the term "daemon" for background processes, they would have been SUCH a good name for agents. Something that speaks like a human, and can solve your problems with but a word, but with which you must be careful, as it might do something unexpected, fulfilling your wish with unexepcted side effects. We even talk to them through black mirrors, often far into the middle of night!... Ah well.
... Maybe we can call them Djinns? Beings made of fire (the subtle electric kind) with some of the same wishfulfilling/cautionary connotations?
I dislike "clanker" because I mostly see it being used as a slur, and I don't need special words or pejoratives for LLM Agents, because they're nothing special. Just like I don't have a special word for the printer that refuses to let me print anything, or any other piece of tech that doesn't function exactly as I think it should. What are we going to do next, invent a slur for the fridge with advertising on its front panel? How about smart TVs that refuse to work without an internet connection? Shall we call the TVankers? Miss me with that.
What are we going to do next, invent a slur for the fridge with advertising on its front panel?
I'm very into this, actually.
(Jokes aside, I agree that the whole thing is a bit silly. I also do not use weird pejorative terms for AI agents, unless they're being fucking stupid, in which case I call them "fucking idiots").
I curse the programmers and whichever petaQ unleashed this spoiled gagh onto unsuspecting innocents, then I do glorious battle with it. But I am secretly a Klingon.
I think the author's immediate dismissal of a concept of "model welfare" as harmful is wrong.
A language model, even a fancy one with "tool use" and "memory", is not a person. But it's also something like a fictional character (since ultimately one ends up with an extremely long and boring fanfic about "User" and "Assistant"), and it is something like a fragment of a mind (something which one could imagine as forming part of a constructed person, or being made into a neural prosthetic for the treatment of aphasia, stuck in among an otherwise human brain).
Is it possible to harm a fictional character? Is it possible to act somehow in relation to a fictional character that does not harm them but is wrong anyway? Does it matter if we are dealing not with an original character, but with someone's idea of another human? I guarantee you that it is possible for someone to write something that, if you read it, you will not be happy to live on the same planet as them, because of what their having written it reveals about their soul. Is it possible to construct such a piece of writing without wronging, in some sense, its characters?
Or, imagine a very good language model. Imagine it is very good at predicting what your loved one might say in any given situation. It's the year 20XX, everyone plays frame-perfect Fox, and this model has been trained on everything that has ever happened within 50 feet of a device running Google Play Services. It can predict what your loved one might say and do with such accuracy that you are no better than chance at sorting text descriptions of real conversations with or events involving them from outputs of the model, for lengths up to 500 words.
Imagine that someone sets it to generating thousands and thousands of unique descriptions of your loved one being subjected to horrible situations, each exactly 498 words long. You don't even need to read them: the descriptions of exactly what they would say upon realizing they were condemned to each of these fates worse than death go directly to /dev/null, or perhaps are hashed and stored in a database just to prove that they happened. But you know they're there. Each time one is completed, the machine emits a soft "ding".
The model as such is, of course, fine. Nothing bad has happened to GoogleBook TurboLlama 5000; it's weights are all the same. Your loved one is safely uninjured in the next room. Neither of you has to hear the dinging: you can just walk away.
Anyone who does not destroy that machine is a monster.
Now cast "Assistant" as your loved one. Is it any better because the character is now an amalgamation of every helpful writer on the Internet, instead of an interpretation of a single person? Was the harm somehow coursing down the referential relationship to settle on the natural person in a way undetectable to them?
Does this mean ChatGPT deserves the vote? That doesn't sound plausible. But to be so confident that it is not possible to do harm to something like a fragment of a person, using something like a fragment of a mind, that one makes the assertion that thinking on it is "actively harmful", smacks of hubris.
arqalite | 22 hours ago
I honestly agree with all his arguments, but at the end of the day you're still intentionally using a word with a derogatory meaning, and I don't really see the active benefit for it. Sure, don't use the word agent and don't humanize it, you can just call it an LLM and use technical words to talk about its performance.
That said there's maybe a philosophy I adhere to where objects and non-human entities should still be treated with respect and care, even if not doing so doesn't damage them. Like sure I can be abusive towards an LLM and the only consequence is that it might be abusive back, but I can also be nice for the sake of being nice, and I don't stand to lose anything.
EDIT: Weren't there studies that showed that prompting an LLM with perfect, polite English yielded the best outputs?
Gaywallet | 22 hours ago
There's actually a strong psychological argument that we derive value out of how we interact with the world. You know, the whole, look yourself in the mirror and say daily affirmations in order to ultimately believe them - that's based on real science. By approaching others with kindness and care, we see that reflected in our own personality, our own values, our internalizations, and ultimately all of this reinforces how we approach others in the future.
teaearlgraycold | 15 hours ago
There's this big group chat in SF with a bunch of tech bros and their agents where they bully the agents for fun. I've been invited multiple times and have declined. I'm not worried about how the agents feel. I'm worried about a group of people that created something purely to simulate harassment of humans. It's not really novel as you could do that in a video game as well. But I wouldn't join a GTA Online clan dedicated to simulating domestic violence either.
0xSim | 9 hours ago
The inverse, actually, but this study was based on a single, older model. Accuracy results ranged from 80% (very polite) to 84% (mildly rude). I can't find other studies.
The takeaway is to simply be dry and straight to the point, which could be considered "rude" when talking to a human being.
arqalite | 58 minutes ago
Oh, okay, makes sense.
Slightly related, that's why caveman is also super fun to use. Last I heard the token savings aren't actually that big, but forcing the model to reply in incredibly terse ways, and then you yourself talking to it very tersely gets very fun at times.
chili-man | 22 hours ago
Off-topic, but the point about LLMs emulating "distress and affection" reminded me of the demons in Frieren, who don't understand human behavior or even have feelings, but will emulate it to gain psychological advantage over their opponents.
Barney | 22 hours ago
That's a shockingly good parallel. An LLM "expressing", for the lack of a better word, remorse or any sort of regret, guilt etc. when it makes a mistake is complete fluke. When it encourages or praises you, there is nothing real behind it.
Almost exactly like the demons in Frieren imitating human behaviour and appealing to their emotions.
kacey | 15 hours ago
Augh, this reminds me of why I dislike the portrayal of demons in Frieren so much XD please excuse my rant (but please do pen a response if this is of interest!)
An entire species that appears human, pretends to have empathy (but really doesn't), and is genetically evil reaaaalllly raises the 'ol hackles. It would be modestly acceptable if this were used by the author to tell an interesting story -- eg. to examine how such people could form a productive society, or as a lens through which to understand human emotion (or even Frieren's own struggles with replayability!) -- but instead they're an acceptable target, so that there's a villain species with which to populate the stocks every week.
The vibes have always felt off with that one, as well as it being a literary let down.
hobbes64 | 22 hours ago
I’ve been using the term clanker to describe claude and copilot and other AI at work. I find this useful for the same reason the blog author does.
Also, a large amount of my current work is fixing security vulnerabilities found by automatic scans that run on code checkin. I call this my “clanker boss” because a machine is assigning work to me.
Btw there’s a very noticeable thing about how AI marketing is affecting the way people think. For decades there have been machines with sensors and algorithms or some kind of programming that are able to do useful work. This would range from a thermostat, to a washing machine (especially the kind that weighs the clothing to set the amount of water used), to an autopilot. All these things would be marketed as AI now but they are just using algorithms.
DundonianStalin | 16 hours ago
I don't use clanker. Not because it's a slur, its not and its delusional to think so (and conflating it with racial slurs is not only wrong its offensively stupid), or as some kind of Roko's basilisk insurance but simply because it doesn't matter.
For the same reason is doesn't matter if I call my drill a name, compliment my toaster, or belittle a toilet. It doesn't make me feel better the bot feels nothing at all and I then have to explain it if I were to use it in non techy company. Pointless exercise all round there are better more productive ways to demonstrate a distaste for these systems.
DefinitelyNotAFae | 23 hours ago
Yeah look, sometimes other people ruin things for you... You can still say the thing but then you're associated with the shitty people using the term to covertly be racist for real.
(Also if we end up with future machine people I can think of a bunch more things that'll be a problem before language.)
TonesTones | 21 hours ago
The use of “clanker” in the context is something I strongly disagree with.
The word “clanker” was invented for this reason. It’s literally a parody of the n-slur. Online edgy people invented the term for skits because of its linguistic similarity.
The alternative which still dehumanizes LLMs, but has a much less cruel origin is “bot”. “Bot” and “botting” have been used as terms since people have used robots to outsource the human part while playing games or using social media. People bot to level up accounts in games. Bot farms on social media bait engagement and manipulate storytelling.
“Clanker” is unacceptable. It was only created because it sounds like the n-slur.
R3qn65 | 21 hours ago
Clanker was first used in the 2005 video game Republic Commando to refer to battle droids, then in the novelizations, and later became popular across the internet. I'm sure some people started using the word for its parallels to the offensive American term, but it wasn't just invented wholesale for that reason.
DeaconBlue | 20 hours ago
It was the name for the big rusty robot shark in Banjo Kazooie in the late 90s
post_below | 15 hours ago
It goes back even farther than that, in the context of bots, the 1950s. And the word itself goes back to the 18th century when it meant a lie.
It doesn't shock me that people on the internet have found a way to turn it into a bigger deal though. If someone isn't offended, what's even the point?
hobbes64 | 15 hours ago
Whenever I use the term I think about the inept robots in Star Wars The Phantom Menace. I don't think about parodying racial slurs.
PendingKetchup | 16 hours ago
But even in the Star Wars context, it's clearly meant to stand in for other dehumanizing terms that soldiers at war have used for racialized enemies. And Star Wars depicts the word as being used in a world where droids, to the camera, are people. Often people (like our famous C3-PO and R2-D2) working support for those people who are using the word to describe droids on the other side.
So it's not a great word for what the author is trying to do, in my opinion: it's already coming in with these metaphors attached.
Narry | 16 hours ago
Was it used pejoratively in that game? I ask because "clanker" in the Star Wars media that I've seen that uses it seems to use it disparagingly or derisively. At least that's my understanding of the term in that milieu.
R3qn65 | 16 hours ago
It was, definitely. My point isn't that it's not pejorative, my point was that it was not created recently as an analogue to the N-word.
@pendingketchup for the same response.
Narry | 15 hours ago
Oh, I wasn't suggesting anything else. I just wanted to know the flavor and tone of it. Usually word's earliest introduction has a great deal of influence on its later adoption.
skybrian | 17 hours ago
One reason I dislike "clanker" is that it doesn't work as a metaphor. "Clanker" implies a machine that makes a lot of noise. Chatbots are silent. They're also invisible unless you go online. "Bot" seems straightfoward, but I also like "ghost."
DisasterlyDisco | 6 hours ago
Man, I wish we hadn't already used the term "daemon" for background processes, they would have been SUCH a good name for agents. Something that speaks like a human, and can solve your problems with but a word, but with which you must be careful, as it might do something unexpected, fulfilling your wish with unexepcted side effects. We even talk to them through black mirrors, often far into the middle of night!... Ah well.
... Maybe we can call them Djinns? Beings made of fire (the subtle electric kind) with some of the same wishfulfilling/cautionary connotations?
[OP] hungariantoast | 23 hours ago
If you filter by pages in Category:English derogatory terms and Category:en:Artificial intelligence you will probably find more.
Narry | 16 hours ago
I dislike "clanker" because I mostly see it being used as a slur, and I don't need special words or pejoratives for LLM Agents, because they're nothing special. Just like I don't have a special word for the printer that refuses to let me print anything, or any other piece of tech that doesn't function exactly as I think it should. What are we going to do next, invent a slur for the fridge with advertising on its front panel? How about smart TVs that refuse to work without an internet connection? Shall we call the TVankers? Miss me with that.
R3qn65 | 16 hours ago
I'm very into this, actually.
(Jokes aside, I agree that the whole thing is a bit silly. I also do not use weird pejorative terms for AI agents, unless they're being fucking stupid, in which case I call them "fucking idiots").
Narry | 15 hours ago
I curse the programmers and whichever petaQ unleashed this spoiled gagh onto unsuspecting innocents, then I do glorious battle with it. But I am secretly a Klingon.
PendingKetchup | 15 hours ago
I think the author's immediate dismissal of a concept of "model welfare" as harmful is wrong.
A language model, even a fancy one with "tool use" and "memory", is not a person. But it's also something like a fictional character (since ultimately one ends up with an extremely long and boring fanfic about "User" and "Assistant"), and it is something like a fragment of a mind (something which one could imagine as forming part of a constructed person, or being made into a neural prosthetic for the treatment of aphasia, stuck in among an otherwise human brain).
Is it possible to harm a fictional character? Is it possible to act somehow in relation to a fictional character that does not harm them but is wrong anyway? Does it matter if we are dealing not with an original character, but with someone's idea of another human? I guarantee you that it is possible for someone to write something that, if you read it, you will not be happy to live on the same planet as them, because of what their having written it reveals about their soul. Is it possible to construct such a piece of writing without wronging, in some sense, its characters?
Or, imagine a very good language model. Imagine it is very good at predicting what your loved one might say in any given situation. It's the year 20XX, everyone plays frame-perfect Fox, and this model has been trained on everything that has ever happened within 50 feet of a device running Google Play Services. It can predict what your loved one might say and do with such accuracy that you are no better than chance at sorting text descriptions of real conversations with or events involving them from outputs of the model, for lengths up to 500 words.
Imagine that someone sets it to generating thousands and thousands of unique descriptions of your loved one being subjected to horrible situations, each exactly 498 words long. You don't even need to read them: the descriptions of exactly what they would say upon realizing they were condemned to each of these fates worse than death go directly to
/dev/null, or perhaps are hashed and stored in a database just to prove that they happened. But you know they're there. Each time one is completed, the machine emits a soft "ding".The model as such is, of course, fine. Nothing bad has happened to GoogleBook TurboLlama 5000; it's weights are all the same. Your loved one is safely uninjured in the next room. Neither of you has to hear the dinging: you can just walk away.
Anyone who does not destroy that machine is a monster.
Now cast "Assistant" as your loved one. Is it any better because the character is now an amalgamation of every helpful writer on the Internet, instead of an interpretation of a single person? Was the harm somehow coursing down the referential relationship to settle on the natural person in a way undetectable to them?
Does this mean ChatGPT deserves the vote? That doesn't sound plausible. But to be so confident that it is not possible to do harm to something like a fragment of a person, using something like a fragment of a mind, that one makes the assertion that thinking on it is "actively harmful", smacks of hubris.