> Vague terms lend themselves well to motte and bailey style argument manipulation
I've never liked the trendiness of the label "motte and bailey", because to me it mostly signifies that someone wants to comment on rhetorical structures without having any prior knowledge of the field. "Motte and bailey" is a modern label that shares the same meaning as the many-centuries-older label "equivocation". It's also an opaque metaphor ("motte" and "bailey" are technical terms related to the geographical layout of feudal European settlements) where "equivocation" is much more transparent, literally referring to "calling [things] equal".
Your comment is an example of fashionable terminology blinding you to the underlying situation. Vague terms don't just "lend themselves" to equivocation; they are the same phenomenon.
(There is much to be said for using a term that you expect your interlocutors to be familiar with instead of using one that you don't. But how does a whole movement form around a modern repetition of a very ancient idea without anyone realizing that people knew about the ancient idea in the past?)
Sorry Michael, but I'm going to have to disagree with you here. While "strategic equivocation" is also an accurate description (and indeed older as you say), I specifically like "motte and bailey" because I like the analogy to how it is used to (unfairly) win arguments.
There you are, out in the bailey, engaging in the productive activity you want to be doing (scoring ideological points with bold claims). Someone comes to challenge you, and you run back to the safety of the motte (pretending that you were only making a much more limited and defensible claim). Then the attacker gets frustrated and goes away, and you go back into the bailey, pretending that you won the argument about the less defensible claims you are pushing.
(obviously, the "you" in this paragraph is not meant to represent you, Michael, but a generic combatant in the online discourse wars)
Also, strategic equivocation encompasses a broader range of tactics... for example strawmanning an argument is also a type of equivocation.
For these reasons, your annoyance at trendy terminology notwithstanding, I will continue to use this term because it conveys concepts with a certain specificity that I prefer.
Agreed. Motte and bailey has become so popular b/c it effectively describes a common rhetorical strategy that's hard to pin down. Although I hadn't heard of mottes or baileys before reading Scott Alexander's, it's a great metaphor that makes the concept memorable.
I'm surprised that Jesse didn't use it in this post. That's what he's really complaining about here. Even with stochastic parrots ... In the motte, it's obvious that AI (LLMs) are built on statistical algorithms and lots of human text to copy. But that doesn't get you to the bailey where LLMs are useless and ineffective.
One thing that allows progressives to punch above their weight is their success at so often creating a home field advantage when it comes the language-arena in which debates take place. With mystifying timidity moderates and conservatives cede that home field advantage, allowing progressives to choose and place their own definitions on the foundational monikers of a debate. Thus, an engineering department that is 60% male has a "Diversity" problem but a educational system that is 80% female does not. School admissions where people of Asian ancestry are over-represented vs their overall percent of the population lack "Equity," but instances where black Americans over-represent do not. Critics of DEI programs allow their positions to be labelled "anti-diversity" and "anti-inclusion." What they should be doing calling out the disingenuous and one-sided way progressives' define those terms.
Agreed but getting people to agree on terms or even understand what they themselves mean when they talk about DEI or just equity vs equality is a whole conversation that keeps having to happen before any debate can even begin. These terms have meanings but we are not all agreed on what they even are and the result is we talk past each other.
To make things even more confusing: parrots are not stochastic parrots. Irene Pepperberg's work with parrots has shown that parrots can actually grasp the conceptual meanings of language, and answer questions correctly with words.
There's a similar problem with the frog and boiling water metaphor. Unlike humans, frogs will jump out of hot water, even if it slowly increases temperature. Perhaps that's because the frogs don't watch Fox News.
Good point. And you can turn it around. Regardless of whether or not parrots are stochastic, they are intelligent. Stochastic parrot is a catchy term but a weird epithet. Like, yeah, maybe but so what?
If only everyone had majored in philosophy. Well, Anglo-American analytic philosophy, anyway. Taking a claim and systematically, exhaustingly, boringly, spelling out every possible thing it might mean and then addressing each possibility one at a time is analytic philosophy's bread and butter. It can be painful to drag yourself through, but it's how you make intellectual progress.
Jessie that piece was a jeremiad against an endless series of strawmen. He kept attacking positions people weren't arguing and misrepresenting what they actually were arguing. And this is your example of "how to argue better"? The piece also conspicuously fails to answer the question of real interest: do LLMs have animal-like cognition? And every LLM developer of note will tell you the answer is no.
Is this really about how to have a good debate? Or did you get butthurt because a bunch of people pushed back against your own AI maximalist piece, and so this piece is particularly attractive to you right now? Which, do you think, is a more rational read of this situation?
with the greatest of respect, it feels like you are arguing past the piece rather than engaging with it.
the whole point of the article is about clarifying arguments, not answering whether LLMs have "animal-like cognition." changing to that question misses the point of what the article is trying to do.
basically: you’re criticizing the article for not doing something it never set out to do.
the "strawman" claim is also questionable.
the argument here is that "stochastic parrot" is used in multiple inconsistent ways. to call that a strawman, you’d need to show those versions aren’t actually being used in online discourse about this subject.
people do in fact use these online - i can provide you with links if you don't believe me.
also: saying "every LLM developer of note would say no" leans on an undefined term while appealing to authority.
what counts as "animal-like cognition"?
depending on the definition, you’ll get very different answers. that’s exactly the kind of vagueness the piece is criticizing.
also: attributing petty motives to Jesse is probably not going to convince him (or anyone else here) of your claims.
you're better than the word "butthurt".
if anything, your response ends up reinforcing the article’s point: vague terms, unexamined assumptions, and talking past the actual claim.
I don’t know if I’m of note but I’m literally an LLM developer and when you asked whether LLMs have animal like cognition I said it was a bit like asking if submarines swim.
LLMs do have cognition, they continue to prove it by achieving stellar scores on things like the international math Olympiad where they are presented with problems they couldn’t have possibly seen before.
It doesn’t matter if it’s “animal-like,” that term is too vague anyway. It’s clearly doing something and calling it mere pattern matching is either vacuously reductionist or wrong, depending on the specifics of what you mean.
All this is a completely different discussion from whether AI will be transformative, or if it will replace humans.
Can you provide an example of a "mere pattern matching" argument that is not wrong, but is "vacuously reductionist"? While I know I wouldn't be able to provide definitions for "pattern matching" and "cognition" without looking them up, as far as I am concerned there is a difference between the terms and my understanding is that pattern matching describes LLMs better than cognition. That doesn't mean LLMs are not valuable/useful/impressive/etc. I do think that our ability to leverage them is impacted based on how we think of them and to that end it might be better to have an understanding of them based in reality (or maybe the opposite is true and we're better off interacting with them as if they are intelligent... this is where I tend to get lost :)
I inferred it as “how do I build a circle of trust?”. It was a jarring transition from the LLM point and didn’t really did deep into either, but I feel like I read a different article.
Speaking of AI - here's something horrific I just learned today:
Uber has switched to algorithmic payments for it's drivers where (to oversimplify) an AI tracks every driver and when a customer needs a ride the AI offers the driver as little payment as it thinks is necessary for the driver to accept the ride.
Two drivers could easily be paid different amounts for the exact same trip.
Meanwhile the percentage of the fare for that ride that Uber keeps varies each time and is always kept hidden from the drivers and the passengers (but has been found to exceed 50%).
Landlords are also feeding availability data into an AI and pricing as it recommends rather than competing with each other.
It's like my grandfather used to say: It's not a cartel if the AI does the colluding!
Can't wait for the grocery stores to start this shit.
From the standpoint of the Uber company, obviously it's advantageous for them to buy low (pay the driver as little as possible) and sell high (charge the rider as much as possible), constrained only on what price/payment levels cause the participants to refuse to participate.
this reminds me of everything I hate about "we live in an oppressive patriarchy" claims.
those who make these claims rarely define "patriarchy" (a textbook example of a floating signifier) and instead motte and bailey between something like the academic "patriachy is where men have all the power and women are largely excluded from it" and some kind of definition that relies on headcounts of politicians and ceos or (worse yet) the mere existence of gender roles.
“general, if someone makes a strong claim — especially a morally weighted one — and then refuses to define their terms or answer follow-up questions, I think this is an extremely useful heuristic”
I think those terms also have to be grounded at some point in something concrete.
You write: "In general, if someone makes a strong claim — especially a morally weighted one — and then refuses to define their terms or answer follow-up questions... this probably isn’t someone worth taking seriously"
Unwillingness to answer follow-up questions is a reasonable heuristic, but there's a difference between unwillingness to admit scrutiny and unwillingness to produce tidy definitions for moral rules which are, themselves, only heuristics. The only thing that can exist at the bottom of morality is preference. I can argue that I would prefer to live in one society than another without needing to apply any sort of rules based ethical system. The definitions only serve this loosely.
When you overextend these heuristics, there are predictable breakdowns. It's why many vegans won't eat honey, but it is also why the word transphobia terminates all discussion. See also, "colonialism".
People making strong claims and refusing to define their terms or answer follow-up questions is what I think of as sloganeering and is why, weirdly, I've given up protesting things in the street. Even posting a pithy statement on social media invites an actual conversation from which I might, gasp!, actually learn something.
Helpful thanks Jesse. But the fact that AI is becoming increasingly powerful does not mean that the intelligence bit of Artificial Intelligence is demonstrated. It would be better if AI stood for Artificial; Imitation - that's more nearly what it is. Ian
Are you suggesting that somebody could be asking in good faith what "transgender" means or what distinguishes Russian propaganda from legitimate free expression?
Vague terms lend themselves well to motte and bailey style argument manipulation, which is why people like them. M&B is powerful in public debate.
> Vague terms lend themselves well to motte and bailey style argument manipulation
I've never liked the trendiness of the label "motte and bailey", because to me it mostly signifies that someone wants to comment on rhetorical structures without having any prior knowledge of the field. "Motte and bailey" is a modern label that shares the same meaning as the many-centuries-older label "equivocation". It's also an opaque metaphor ("motte" and "bailey" are technical terms related to the geographical layout of feudal European settlements) where "equivocation" is much more transparent, literally referring to "calling [things] equal".
Your comment is an example of fashionable terminology blinding you to the underlying situation. Vague terms don't just "lend themselves" to equivocation; they are the same phenomenon.
(There is much to be said for using a term that you expect your interlocutors to be familiar with instead of using one that you don't. But how does a whole movement form around a modern repetition of a very ancient idea without anyone realizing that people knew about the ancient idea in the past?)
Sorry Michael, but I'm going to have to disagree with you here. While "strategic equivocation" is also an accurate description (and indeed older as you say), I specifically like "motte and bailey" because I like the analogy to how it is used to (unfairly) win arguments.
There you are, out in the bailey, engaging in the productive activity you want to be doing (scoring ideological points with bold claims). Someone comes to challenge you, and you run back to the safety of the motte (pretending that you were only making a much more limited and defensible claim). Then the attacker gets frustrated and goes away, and you go back into the bailey, pretending that you won the argument about the less defensible claims you are pushing.
(obviously, the "you" in this paragraph is not meant to represent you, Michael, but a generic combatant in the online discourse wars)
Also, strategic equivocation encompasses a broader range of tactics... for example strawmanning an argument is also a type of equivocation.
For these reasons, your annoyance at trendy terminology notwithstanding, I will continue to use this term because it conveys concepts with a certain specificity that I prefer.
Agreed. Motte and bailey has become so popular b/c it effectively describes a common rhetorical strategy that's hard to pin down. Although I hadn't heard of mottes or baileys before reading Scott Alexander's, it's a great metaphor that makes the concept memorable.
I'm surprised that Jesse didn't use it in this post. That's what he's really complaining about here. Even with stochastic parrots ... In the motte, it's obvious that AI (LLMs) are built on statistical algorithms and lots of human text to copy. But that doesn't get you to the bailey where LLMs are useless and ineffective.
One thing that allows progressives to punch above their weight is their success at so often creating a home field advantage when it comes the language-arena in which debates take place. With mystifying timidity moderates and conservatives cede that home field advantage, allowing progressives to choose and place their own definitions on the foundational monikers of a debate. Thus, an engineering department that is 60% male has a "Diversity" problem but a educational system that is 80% female does not. School admissions where people of Asian ancestry are over-represented vs their overall percent of the population lack "Equity," but instances where black Americans over-represent do not. Critics of DEI programs allow their positions to be labelled "anti-diversity" and "anti-inclusion." What they should be doing calling out the disingenuous and one-sided way progressives' define those terms.
Agreed but getting people to agree on terms or even understand what they themselves mean when they talk about DEI or just equity vs equality is a whole conversation that keeps having to happen before any debate can even begin. These terms have meanings but we are not all agreed on what they even are and the result is we talk past each other.
To make things even more confusing: parrots are not stochastic parrots. Irene Pepperberg's work with parrots has shown that parrots can actually grasp the conceptual meanings of language, and answer questions correctly with words.
There's a similar problem with the frog and boiling water metaphor. Unlike humans, frogs will jump out of hot water, even if it slowly increases temperature. Perhaps that's because the frogs don't watch Fox News.
also, if you throw a frog into a pot of boiling water it dies almost instantly.
Good point. And you can turn it around. Regardless of whether or not parrots are stochastic, they are intelligent. Stochastic parrot is a catchy term but a weird epithet. Like, yeah, maybe but so what?
If only everyone had majored in philosophy. Well, Anglo-American analytic philosophy, anyway. Taking a claim and systematically, exhaustingly, boringly, spelling out every possible thing it might mean and then addressing each possibility one at a time is analytic philosophy's bread and butter. It can be painful to drag yourself through, but it's how you make intellectual progress.
Jessie that piece was a jeremiad against an endless series of strawmen. He kept attacking positions people weren't arguing and misrepresenting what they actually were arguing. And this is your example of "how to argue better"? The piece also conspicuously fails to answer the question of real interest: do LLMs have animal-like cognition? And every LLM developer of note will tell you the answer is no.
Is this really about how to have a good debate? Or did you get butthurt because a bunch of people pushed back against your own AI maximalist piece, and so this piece is particularly attractive to you right now? Which, do you think, is a more rational read of this situation?
Freddie
with the greatest of respect, it feels like you are arguing past the piece rather than engaging with it.
the whole point of the article is about clarifying arguments, not answering whether LLMs have "animal-like cognition." changing to that question misses the point of what the article is trying to do.
basically: you’re criticizing the article for not doing something it never set out to do.
the "strawman" claim is also questionable.
the argument here is that "stochastic parrot" is used in multiple inconsistent ways. to call that a strawman, you’d need to show those versions aren’t actually being used in online discourse about this subject.
people do in fact use these online - i can provide you with links if you don't believe me.
also: saying "every LLM developer of note would say no" leans on an undefined term while appealing to authority.
what counts as "animal-like cognition"?
depending on the definition, you’ll get very different answers. that’s exactly the kind of vagueness the piece is criticizing.
also: attributing petty motives to Jesse is probably not going to convince him (or anyone else here) of your claims.
you're better than the word "butthurt".
if anything, your response ends up reinforcing the article’s point: vague terms, unexamined assumptions, and talking past the actual claim.
I don’t know if I’m of note but I’m literally an LLM developer and when you asked whether LLMs have animal like cognition I said it was a bit like asking if submarines swim.
LLMs do have cognition, they continue to prove it by achieving stellar scores on things like the international math Olympiad where they are presented with problems they couldn’t have possibly seen before.
It doesn’t matter if it’s “animal-like,” that term is too vague anyway. It’s clearly doing something and calling it mere pattern matching is either vacuously reductionist or wrong, depending on the specifics of what you mean.
All this is a completely different discussion from whether AI will be transformative, or if it will replace humans.
Can you provide an example of a "mere pattern matching" argument that is not wrong, but is "vacuously reductionist"? While I know I wouldn't be able to provide definitions for "pattern matching" and "cognition" without looking them up, as far as I am concerned there is a difference between the terms and my understanding is that pattern matching describes LLMs better than cognition. That doesn't mean LLMs are not valuable/useful/impressive/etc. I do think that our ability to leverage them is impacted based on how we think of them and to that end it might be better to have an understanding of them based in reality (or maybe the opposite is true and we're better off interacting with them as if they are intelligent... this is where I tend to get lost :)
love the submarine analogy.
Thanks. Can’t take full credit for it, I’ve seen it a few places before. But it is quite apt.
"attacking positions people weren't arguing"
You, of all people, are making the "no one is saying that" objection?
I inferred it as “how do I build a circle of trust?”. It was a jarring transition from the LLM point and didn’t really did deep into either, but I feel like I read a different article.
Did you get the World's Greatest Dad coffee mug that the BARpod comments section bought you?
Just to confirm, you still live at the corner of Sad and Faded?
If I were to write a piece like this, I would make sure to highlight a piece I DID NOT AGREE WITH. For basic credibility's sake.
Speaking of AI - here's something horrific I just learned today:
Uber has switched to algorithmic payments for it's drivers where (to oversimplify) an AI tracks every driver and when a customer needs a ride the AI offers the driver as little payment as it thinks is necessary for the driver to accept the ride.
Two drivers could easily be paid different amounts for the exact same trip.
Meanwhile the percentage of the fare for that ride that Uber keeps varies each time and is always kept hidden from the drivers and the passengers (but has been found to exceed 50%).
Landlords are also feeding availability data into an AI and pricing as it recommends rather than competing with each other.
It's like my grandfather used to say: It's not a cartel if the AI does the colluding!
Can't wait for the grocery stores to start this shit.
From the standpoint of the Uber company, obviously it's advantageous for them to buy low (pay the driver as little as possible) and sell high (charge the rider as much as possible), constrained only on what price/payment levels cause the participants to refuse to participate.
It sounds like you’re treating the information and power asymmetries here as normal or acceptable.
this reminds me of everything I hate about "we live in an oppressive patriarchy" claims.
those who make these claims rarely define "patriarchy" (a textbook example of a floating signifier) and instead motte and bailey between something like the academic "patriachy is where men have all the power and women are largely excluded from it" and some kind of definition that relies on headcounts of politicians and ceos or (worse yet) the mere existence of gender roles.
The core point about concretizing terms is a good one.
Even a bad argument is made less bad through clarity of terms.
“general, if someone makes a strong claim — especially a morally weighted one — and then refuses to define their terms or answer follow-up questions, I think this is an extremely useful heuristic”
I think those terms also have to be grounded at some point in something concrete.
Really enjoyed this piece.
OK I lied, I have one comment.
You write: "In general, if someone makes a strong claim — especially a morally weighted one — and then refuses to define their terms or answer follow-up questions... this probably isn’t someone worth taking seriously"
Unwillingness to answer follow-up questions is a reasonable heuristic, but there's a difference between unwillingness to admit scrutiny and unwillingness to produce tidy definitions for moral rules which are, themselves, only heuristics. The only thing that can exist at the bottom of morality is preference. I can argue that I would prefer to live in one society than another without needing to apply any sort of rules based ethical system. The definitions only serve this loosely.
When you overextend these heuristics, there are predictable breakdowns. It's why many vegans won't eat honey, but it is also why the word transphobia terminates all discussion. See also, "colonialism".
Lovely essay, no notes.
Ha! Great moniker for this comment.
People making strong claims and refusing to define their terms or answer follow-up questions is what I think of as sloganeering and is why, weirdly, I've given up protesting things in the street. Even posting a pithy statement on social media invites an actual conversation from which I might, gasp!, actually learn something.
Helpful thanks Jesse. But the fact that AI is becoming increasingly powerful does not mean that the intelligence bit of Artificial Intelligence is demonstrated. It would be better if AI stood for Artificial; Imitation - that's more nearly what it is. Ian
Are you suggesting that somebody could be asking in good faith what "transgender" means or what distinguishes Russian propaganda from legitimate free expression?
Yes. Ignorance is a thing that actually exists and should not be treated as a moral failing.