I have seen this terminology for first time. Interesting wordplay. You're trying to dismantle the aspects of consciousness, empathy, moral compass altogether from the AI tools hardware or software. But default I think it is already in a dehyphenated form.
Before I bought my first car I used to think I would call my Moving Machine because yes it just moves, but then I go on to call it by name and take pride and feel happy when I ride it. It is humans who attach values and emotions to machines not the other way. And this would be the only distinguishing factor between the two of us for eternity.
If a meatbag uses this word in public, they're rightly going to get punched. It's not their word to claim. Is there any wonder the word has become associated with slavery and racism?
A clanker doesn't clank? It feels more like an emotional name — a pet name or derogatory name — as opposed to a name that evokes a tool-like view of the thing. A typer? an autocompleter? auto-sed? LLM? All probably would have less emotional feeling to it due to being rooted in the actions performed.
Strong agree with the point that "agents" shouldn't be called "agents", as agency lies with human. That being said, I don't like the word clanker either.
I prefer the term clanker. It accurately sketches LLMs for what they are: heartless machines. However, none of them make that tinman clank sound, so it is still a bit imprecise. Once we have them walking and talking, it will be perfect.
> The reason I liked the word is precisely the opposite of that use. I want language that prevents anthropomorphizing. I want a word that says: this is a tool, a machine of numbers and matrices
I understand wanting the language to reflect the actuality that these agents are just code doing predictive word guessing and not a feeling person. The irony is by using a word that culturally reflects and is used as a slur for AI rather than more precise wording that describes what an agent is to "creates distance from the machine", you are paradoxically humanizing it just enough to dehumanize it with the word.
I do find it funny how mad some people get about this.
Meanwhile, among the yonger generations, "clanker" has been fully established a a term to disapprove of AI for at least a year. though it remains to be seen if it will stick or just be a passing meme, like tge eponymous "6-7" or the now nearly dead "skibidi"
it reminds people (or specifically, very online gen Z and gen alpha) of the N-word because there was a popular TikTok account that mapped "clanker" to 1960s civil rights parody videos and it was clear he was substituting it for the N-word
The developers of Pi have some of the most reasonable and level-headed opinions about AI for someone in the AI space. It's always so refreshing to read their stuff.
As far as I'm concerned (and I know I won't be in the majority here, but whatever) non-ironic or non-meta usage of the term "clanker" for "agent" just makes one sound like a dumbass.
Tells something how we went from "meatbag" back when tech was cool and the Future etc to "clanker" now that tech is evil...
I agree with his rationale, but not with how he arrives there.
> Racism Is About Humans
This is generally true, but this argument also falls into a pattern that has enabled racism in the past: "It's a sin to harm any soul, but sadly the savages/blacks/natives/etc have no souls" etc etc.
I think it might get more muddy if you compare clankers to animals: Those too are clearly not humans, often have a vastly different "neural wiring" than humans and are at all kinds of different levels of cognition. But we still recognize that some level is protection is warranted and the reasoning goes beyond "puppies look so cute".
I'd at least like some deeper criterion than "they're not human" or "they're just next token predictors" - at least then this would depend how they predict the next token.
I still have no qualms using LLMs and I don't think they (or at least the current batch) needs any protection. What's a more convincing argument for me is that there isn't even a clear way to count how many individual "AI"s there are that could be conscious. The number of models is relatively clear-cut, but models are just numbers on a hard drive. They're about as conscious as raw DNA is.
The only thing where a consciousness could hypothetically exist is in the inference loop - and that lasts only very short time anyway - so it would more be like a "Mr. Meeseeks" kind of situation, where some kind of ephemeral entity gets spirited into being (by the GPU!), does some work and then vanishes again. Except, it did not even really "vanish": Because the conversation, i.e. context window is the entity, you could "wake it up" again by continuing the conversation. It seems very hard to imagine a consciousness inside that - and even then, it would necessarily be one with retrograde and anterograde amnesia.
So I'm not exactly worried, until we have models that are running in a constant loop, where the weights are continuously updated and that shows evidence of episodic memory. Will that change in the future? Who knows?
I'm really surprised that neither the post nor any other comments here seem to actually mention Star Wars. For those unaware, "Clanker" is the term the clone troopers used for the evil battle droids in the Clone Wars animated cartoon series. The show was fairly popular, and people have started to use the term for AI because it's funny.
The use Armin wants for clanker doesn't work because it is already widely being used as a derogatory term for LLMs. Co-opting the term back to the more neutral use he wants is a long uphill battle (and basically futile).
Regarding his wider point, I think something he's not quite taking into account is how personal language is. Yes, we should always keep in mind the LLM is a machine, but because we're human it's also fine to say "please" and to otherwise "play into" the human-like text that is being outputted. Many children still hug their stuffed animals, knowing that they are inanimate. Adults have been speaking politely to Siri and Alexa for years, which do not give off any vibes of possibly being intelligent or sentient. Purposefully adjusting the way you speak just because you know the thing you're speaking to is inanimate is actually what's strange, and human beings are creatures of habit. There is nothing wrong with speaking respectfully to an LLM even if you are under no delusions that its feelings can be hurt. There is no need to speak curtly or rudely just to remind yourself of this.
I'm skimming this, and I also agree - machines should not pretend to be human, and we should not pretend they are.
But I do think that, like, being mean to a machine and calling it slurs and telling it it's a piece of shit is bad not because the machine feels bad, but because despite our efforts, we do anthromorphize things, and it will normalize the idea that we can just call "someone" a piece of shit.
I used to call my mom's dog "the horrible hound", and eventually swithced to calling him "the honorable hound", and I explicitly told my kid: I know the dog doesn't know I'm insulting it, but I do, and that's not the kind of person I want to be.
17 comments
[ 4.0 ms ] story [ 41.5 ms ] threadA clanker doesn't clank? It feels more like an emotional name — a pet name or derogatory name — as opposed to a name that evokes a tool-like view of the thing. A typer? an autocompleter? auto-sed? LLM? All probably would have less emotional feeling to it due to being rooted in the actions performed.
I understand wanting the language to reflect the actuality that these agents are just code doing predictive word guessing and not a feeling person. The irony is by using a word that culturally reflects and is used as a slur for AI rather than more precise wording that describes what an agent is to "creates distance from the machine", you are paradoxically humanizing it just enough to dehumanize it with the word.
Meanwhile, among the yonger generations, "clanker" has been fully established a a term to disapprove of AI for at least a year. though it remains to be seen if it will stick or just be a passing meme, like tge eponymous "6-7" or the now nearly dead "skibidi"
It's not cool, it's not clever, it's just stupid.
I agree with his rationale, but not with how he arrives there.
> Racism Is About Humans
This is generally true, but this argument also falls into a pattern that has enabled racism in the past: "It's a sin to harm any soul, but sadly the savages/blacks/natives/etc have no souls" etc etc.
I think it might get more muddy if you compare clankers to animals: Those too are clearly not humans, often have a vastly different "neural wiring" than humans and are at all kinds of different levels of cognition. But we still recognize that some level is protection is warranted and the reasoning goes beyond "puppies look so cute".
I'd at least like some deeper criterion than "they're not human" or "they're just next token predictors" - at least then this would depend how they predict the next token.
I still have no qualms using LLMs and I don't think they (or at least the current batch) needs any protection. What's a more convincing argument for me is that there isn't even a clear way to count how many individual "AI"s there are that could be conscious. The number of models is relatively clear-cut, but models are just numbers on a hard drive. They're about as conscious as raw DNA is.
The only thing where a consciousness could hypothetically exist is in the inference loop - and that lasts only very short time anyway - so it would more be like a "Mr. Meeseeks" kind of situation, where some kind of ephemeral entity gets spirited into being (by the GPU!), does some work and then vanishes again. Except, it did not even really "vanish": Because the conversation, i.e. context window is the entity, you could "wake it up" again by continuing the conversation. It seems very hard to imagine a consciousness inside that - and even then, it would necessarily be one with retrograde and anterograde amnesia.
So I'm not exactly worried, until we have models that are running in a constant loop, where the weights are continuously updated and that shows evidence of episodic memory. Will that change in the future? Who knows?
Regarding his wider point, I think something he's not quite taking into account is how personal language is. Yes, we should always keep in mind the LLM is a machine, but because we're human it's also fine to say "please" and to otherwise "play into" the human-like text that is being outputted. Many children still hug their stuffed animals, knowing that they are inanimate. Adults have been speaking politely to Siri and Alexa for years, which do not give off any vibes of possibly being intelligent or sentient. Purposefully adjusting the way you speak just because you know the thing you're speaking to is inanimate is actually what's strange, and human beings are creatures of habit. There is nothing wrong with speaking respectfully to an LLM even if you are under no delusions that its feelings can be hurt. There is no need to speak curtly or rudely just to remind yourself of this.
But I do think that, like, being mean to a machine and calling it slurs and telling it it's a piece of shit is bad not because the machine feels bad, but because despite our efforts, we do anthromorphize things, and it will normalize the idea that we can just call "someone" a piece of shit.
I used to call my mom's dog "the horrible hound", and eventually swithced to calling him "the honorable hound", and I explicitly told my kid: I know the dog doesn't know I'm insulting it, but I do, and that's not the kind of person I want to be.