This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
I would consider myself an LLM evangelist, and have introduced quite a few not-particularly tech savvy people to them, with good results.
I've never been tempted to call them stochastic parrots. The term harms more than it helps. My usual shortcut is to tell people to act as if they're talking to a human, a knowledgeable but fallible one, and they should double check anything of real consequence. This is a far more relevant description of the kind of capabilities they possess than any mention of a "parrot".
The fact you've never been tempted to use the 'stochastic parrot' idea just means you haven't dealt with the specific kind of frustration I'm talking about.
Yeah the 'fallible but super intelligent human' is my first shortcut too, but it actually contributes to the failure mode the stochastic parrot concept helps alleviate. The concept is useful for those who reply 'Yeah, but when I tell a human they're being an idiot, they change their approach.' For those who want to know why it can't consistently generate good comedy or poetry. For people who don't understand rewording the prompt can drastically change the response, or those who don't understand or feel bad about regenerating or ignoring the parts of a response they don't care about like follow up questions.
In those cases, the stochastic parrot is a more useful model than the fallible human. It helps them understand they're not talking to a who, but interacting with a what. It explains the lack of genuine consciousness, which is the part many non-savvy users get stuck on. Rattling off a bunch of info about context windows and temperature is worthless, but saying "it's a stochastic parrot" to themselves helps them quickly stop identifying it as conscious. Claiming it 'harms more than it helps' seems more focused on protecting the public image of LLMs than on actually helping frustrated users. Not every explanation has to be a marketing pitch.
I still don't see why that applies, and I'm being earnest here. What about the "stochastic parrot" framing keys the average person into the fact that they're good at code and bad at poetry? That is more to do with mode collapse and the downsides of RLHF than it is to do with lacking "consciousness". Like, even on this forum, we have no shortage of users who are great at coding but can't write a poem to save their lives, what does that say about their consciousness? Are parrots known to be good at Ruby-on-rails but to fail at poetry?
My explanation of temperature is, at the very least, meant as a high level explainer. It doesn't come up in normal conversation or when I'm introducing someone to LLMs. Context windows? They're so large now that it's not something that is worth mentioning except in passing.
My point is that the parrot metaphor adds nothing. It is, at best, irrelevant, when it comes to all the additional explainers you need to give to normies.
I thought I explained it pretty well, but I will try again. It is a cognitive shortcut, a shorthand people can use when they are still modelling it like a 'fallible human' and expecting it to respond like a fallible human. Mode collapse and RLHF have nothing to do with it, because it isn't a server side issue, it is a user issue, the user is anthropomorphising a tool.
Yes, temperature and context windows (although I actually meant to say max tokens, good catch) don't come up in normal conversation, they mean nothing to a normie. When a normie is annoyed that chatgpt doesn't "get" them, the parrot model helps them pivot from "How do I make this understand me?" to "What kind of input does this tool need to give me the output I want?"
You can give them a bunch of additional explanations about mode collapse and max tokens that they won't understand (and they will just stop using it) or you can give them a simple concept that cuts through the anthropomorphising immediately so that when they are sitting at their computer getting frustrated at poor quality writing or feeling bad about ignoring the llms prodding to take the conversation in a direction they don't care about, they can think 'wait it's a stochastic parrot' and switch gears. It works.
A human fails at poetry because it has the mind, the memories and grounding in reality, but it lacks the skill to match the patterns we see as poetic. An LLM has the skill, but lacks the mind, memories and grounding in reality. What about the parrot framing triggers that understanding? Memetics I guess. We have been using parrots to describe non-thinking pattern matchers for centuries. Parroting a phrase goes back to the 18th century. "The parrot can speak, and yet is nothing more than a bird" is a phrase in the ancient Chinese Book of Rites.
Also I didn't address this earlier because I thought it was just amusing snark, but you appear to be serious about it. Yes, you are correct that a parrot can't code. Do you have a similar problem with the fact a computer virus can't be treated with medicine? Or that the cloud is actually a bunch of servers and can't be shifted by the wind? Or the fact that the world wide web wasn't spun by a world wide spider? Attacking a metaphor is not an argument.
I've explained why I think the parrot is a terrible metaphor above. And no, metaphors can vary greatly in how useful or pedagogical they are. Analyzing the fitness of a metaphor is a perfectly valid, and in this case essential, form of argument. Metaphors are not neutral decorations; they are cognitive tools that structure understanding and guide action.
A computer virus shares many properties with its biological counterpart, such as self-replication, transmission, damage to systems, the need for an "anti-virus". It is a good name, and nobody with a functional frontal lobe comes away thinking they need an N95 mask while browsing a porn site.
The idea of the Cloud at least conveys the message that the user doesn't have to worry about the geographical location of their data. Even so, the Cloud is just someone else's computer, and even AWS goes down on rare occasions. It is an okay metaphor.
The Parrot is awful. It offers no such explanatory power for the observed, spiky capability profile of LLMs. It does not explain why the model can write functional Python code (a task requiring logic and structure) but often produces insipid poetry (a task one might think is closer to mimicry). It does not explain why an LLM can synthesize a novel argument from disparate sources but fail to count the letters in a word. A user equipped only with the parrot model is left baffled by these outcomes. They have traded the mystery of a "fallible human" for the mystery of a "magical parrot".
I contend that as leaky generalizations go, the former is way better than the latter. An LLM has a cognitive or at least behavioral profile far closer to a human than it does to a parrot.
You brought up the analogy of "parroting" information, which I would assume involves simply reciting things back without understanding what they mean. That is not a good description of how the user can expect an LLM to behave.
On an object level, I strong disagree with your claims that LLMs don't "think" or don't have "minds". They clearly have a very non human form of cognition, but so does an octopus.
Laying that aside, from the perspective of an end-user, LLMs are better modeled as thinking minds.
The "fallible but knowledgeable intern" or "simulation engine" metaphor is superior not because it is more technically precise (though it is), but because it is more instrumentally useful. It correctly implies the user's optimal strategy: that performance is contingent on the quality of the instructions (prompting), the provided background materials (context), and a final review of the output (verification). This model correctly guides the user to iterate on their prompts, to provide examples, and to treat the output as a draft. The parrot model, in contrast, suggests the underlying process is fundamentally random mimicry, which offers no clear path to improvement besides "pull the lever again". It encourages users to conceptualize the LLM as a tool incapable of generalization, which is to ignore its single most important property. Replacing a user's anthropomorphism with a model that is descriptively false and predictively useless is not a pedagogical victory. It is swapping one error for another, and not even for a less severe one to boot.
We are looking at this from two different angles. My angle helps people. Your angle, which seems to prioritize protecting the LLM from the 'insult' of a simple metaphor, actively harms user adoption. My goal in using the parrot model is to solve a specific and very common point of frustration - the anthropomorphising of a tool. I know the parrot shortcut works, I have watched it work and I have been thanked for it.
The issue is that humans - especially older humans - have been using conversation - a LUI - in a very particular way their entire lives. They have conversations with other humans who are grounded in objective reality, who have emotions and memories, and therefore when they use a LUI to interact with a machine, they subconsciously pattern match the machine to other humans and expect it to work the same way - and when it doesn't they get frustrated.
The parrot model on the other hand, tells the user 'Warning: This looks like the UI you have been using your whole life, but it is fundamentally different. Do not assume understanding. Do not assume intention. Your input must be explicit and pattern-oriented to get a predictable output.' The parrot doesn't get anything. It has no intentions in the sense the person is thinking of. It can't be lazy. The frustration dissolves and is replaced by a practical problem solving mindset. Meanwhile the fallible intern exacerbates the very problem I am trying to solve by reinforcing the identification of the LLM as a conscious being.
The beauty is, once they get over that, once they no longer have to use the parrot model to think of it as a tool, they start experimenting with it in ways they wouldn't have before. They feel much more comfortable treating it like a conversation partner they can manipulate through the tech. Ironically they feel more comfortable joking about it being alive and noticing the ways it is like and unlike a person. They get more interested in learning how it actually works, because they aren't shackled by the deeply ingrained grooves of social etiquette.
You're right that metaphors should be analyzed for fitness, but that analysis requires engaging with the metaphor's intended purpose, not just attacking its accuracy literally. A metaphor only needs to illuminate one key trait to be effective, but the parrot goes a lot further than that. It is in fact fantastic at explaining the spiky profile of LLMs. It explains why an LLM can 'parrot' highly structured Python from its training data but write insipid poetry that lacks the qualia of human experience. Likewise I could train a parrot to recite 10 PRINT "BALLS"; 20 GOTO 10, but it could never invent a limerick. It explains why it can synthesize text (a complex pattern matching task) but can't count letters in a word (a character level task it's not trained to understand). Your analysis ignores this context, seemingly because the metaphor is offensive to an aspirational view of AI. But you're attacking a subway map for not being a satellite image. The resolution is drastically reduced yes - this is a selling point, not a flaw. Cultural cachet drastically outweighs accuracy when it comes to a metaphor's usefulness in real world applications.
And do you want to know another animal with a clearly non human form of cognition? A parrot. How did you skip over crows and dolphins to get to octupi, animals with an intelligence that is explicitly not language based, when we are talking about language models? Unlike an octopus, a parrot's intelligence is startlingly relevant here (my mentioning of parroting was just an example of how a parrot has been used as a metaphor for a non-thinking (or if you prefer, non-feeling) pattern matcher in the past.) Using a LUI a parrot can learn complex vocalisation. They can learn mimicry and memorisation. They can learn to associate words with objects and concepts (like colours and zero). They can perform problem solving tasks through dialogue. Is it just because octupus intelligence is cool and weird? Because that just brings me back to the difference between evangelising llms and helping people. You want to talk up llms, I want to increase their adoption.
Shaming users for not having the correct mental model is precisely how we end up with people who are afraid of their tools - the boomers who work out calculations on a pocket calculator before typing them into Excel, or who type 'Gmail login' into the Google search bar every single day. As social media amply demonstrates, technical accuracy does not aid in adoption, it is a barrier to it. We can dislike that from a nerd standpoint, which is why I admired your point in my original post (technically correct is the best kind of correct!) but user adoption will do a lot more for advancing the tech.
Touché. I walked into that one.
Look, come on. We are literally in a thread dedicated to avoiding Bulverism. Do you honestly think I'm out here defending the honor of a piece of software? My concern is not for the LLM's public image. Sam Altman is not sending me checks. I pay for ChatGPT Plus.
I think the charitable, and correct, framing is that we are both trying to help people use these things better. We just disagree on the best way to do that. My entire point is that the "stochastic parrot" model, while it might solve the one specific problem of a user getting frustrated, ultimately creates more confusion than it solves. It's a bad mental model, and I care about users having good mental models.
You're right that a metaphor is a subway map, not a satellite image. Its value is in its simplification. But for a subway map to be useful, it has to get the basic topology right. It has to show you which stations connect. The parrot map gets the topology fundamentally wrong.
It tells you the machine mimics, and that's it. It offers zero explanation for the weird, spiky capability profile. Why can this "parrot" debug Python but not write a good joke? Why can it synthesize three different academic papers into a novel summary but fail to count the letters in a word? The parrot model just leaves you with "I guess it's a magic parrot". It doesn't give the user any levers to pull. What's the advice? "Just keep feeding the parrot crackers and hope it says something different?"
Compare that to the "fallible but brilliant intern" model. It's also a simplification, but it's a much better map. It correctly predicts the spikiness. An intern can be a world-class expert on one topic and completely sloppy with basic arithmetic. That feels right. More importantly, it gives the user an immediate, actionable strategy. What do you do with a brilliant but fallible intern? You give them very clear instructions, you provide them with all the necessary background documents, and you always, always double-check their work for anything mission-critical. That maps perfectly onto prompt engineering, RAG, and verification. It empowers the user. The parrot model just leaves them shrugging.
I'm pretty sure I haven't done that. My frustration isn't with your average user. It's with people who really should know better using the term as a thought-terminating cliche to dismiss the whole enterprise.
If my own grandmother told me she was getting frustrated because "Mr. GPT" kept forgetting what she told it yesterday, I wouldn't lecture her on stateless architecture. I'd say something like, "Think of it as having the world's worst long-term memory. It's a total genius, but you have to re-introduce yourself and explain the whole situation from scratch every single time you talk to it."
That's also a simple, not-quite-accurate metaphor. But it's a better one. It's a better map. It addresses her actual problem and gives her a practical way to think that will get her better results next time. It helps her use the tool, which is the goal I think we both agree on.
I'm pretty sure you said people like me are less intelligent than a parrot and that you hope we get mauled by a tiger. You did not specify that it was only directed at those using it to dismiss using AI, it was anyone using the term unironically. If I felt shame like normal people I would have simply stopped doing it instead of defending it - and I would no longer be helping people stop anthropomorphising a tool.
You lay out your complex 'fallible intern' model as the superior model. It can debug code and synthesise academic papers, it has a mind, though unlike any we know. You say we need to teach people to give clear instructions, provide background documents, and verify all work. But when you imagine talking to your own grandmother - a perfect example of a novice user - what do you do? You drop the intern model completely in favour of a genius with the world's worst memory. Why?
Because you know the intern model is too complicated. You know it doesn't work for a normal person. You'd never actually saddle your grandmother with the mental load of dealing with an intern who is an amnesiac - and is also a compulsive liar who has mood swings, no common sense, and can't do math. You give her a simple tool for the problem. But your tool deals with the symptom, mine deals with the cause.
I believe that you are trying to help people too, but you really are prioritising defending your model first. It might work great with techbros or the techbro adjacent, but even you drop it when you imagine a real world situation with a novice.
And I have to say, if I told you I'm not biased towards Teslas, Elon doesn't send me cheques, and in fact I just paid money for one, how wide would your eyes go as you attempted to parse that?
Look, I think it's quite clear that my statement about tigers was hyperbole. You seem like a perfectly nice guy, while I wouldn't jump into the ring to save you, I'd throw rocks (at the tiger) and call for paramedics.
That is the nice thing about being able to compartmentalize one's combative online persona from being an actually nice and easy-going person in reality. There are very few people I would actually watch and let die, and they're closer to Stalin than they are to people I disagree with on a forum for underwater basket weaving.
If this were a perfectly realistic scenario, my conversation would go:
"What the fuck. Is that a ghost? I thought your ashes were somewhere in the Bay of Bengal by now."
Do you understand why I paraphrased what is usually a more nuanced context-dependent conversation IRL? If my granny was actually alive, I would probably teach her how to use the voice mode in her native language and let her chill.
Uh? I don't know. If you have a reputation for doing that, I genuinely do not recall. I am very active here, but I do not remember that without actually having to open your profile.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link