Stochastic Parrots: Frequently Unasked Questions (medium.com)
42 points by olalonde 3 days ago
siegecraft 11 minutes ago
> Most things we historically do with computing are not well approximated by extruding synthetic text.
I don't understand this point. I feel like almost everything associated with computing is extruding synthetic text.
loandbehold an hour ago
Sounds like increasing capabilities of LLMs over last 5 years proved her 2021 paper wrong but instead of admitting that she had been wrong she's trying to change/reinterpret what she wrote in 2021.
hellohello2 2 hours ago
"Text generated by an LM is not grounded in communicative intent, any model of the world, or any model of the reader’s state of mind."
Modelling text describing the world is not modelling (some aspect) of the world?
Modelling the probability that a reader likes or dislike a piece of text is not modelling (some aspect) of a reader's state of mind?
tootie 2 hours ago
No? There's no model involved. It's all just probabilistic. LLMs understand what you're thinking as well as a mood ring.
roenxi 2 hours ago
It isn't possible to have "just probabilistic" (maybe a philosophical exception could be made for a uniform random distribution or whatever provides the little dose of randomness required to get nondeterministic results). Probabilities are always in context of a model. LLMs model language but language itself is a model of something else. My money would have been on language modelling nonsense, but that is quite clearly not the case. Turns out it models the world and so do LLMs.
aoeusnth1 an hour ago
The model is the thing which is learned in order to make the probabilistic prediction with low entropy.
hellohello2 2 hours ago
The literal definition of a model is "an informative representation of an object, person, or system". I think you mean something else though, what are you trying to express exactly?
afthonos 2 hours ago
Nothing about an LLM is “just”. In what precise sense do you mean it is probabilistic?
NooneAtAll3 7 minutes ago
> Another common trope in the discourse around this phrase is to claim that stochastic parrot is an insult (or even a slur). On one reading, that would require LLMs to be the kind of thing that can take or feel offense, which they clearly aren’t.
isn't that circular reasoning?
"I can call anyone not smart enough to take offense because as I said those anyone aren't smart enough to take offense"?
(also disregarding that being offended has been shifted into "protection of the (perceived) weak" rather than "protection of self" for quite some time now)
---
but generally I always felt that this tension around the phrase was somewhat of perscriptive/descriptive difference, or maybe "level of detail in the model" type
just because there is knowledge of a more full understanding of the process doesn't mean other descriptions/modeling of the process are invalid or unuseful
newtonian gravity doesn't describe time dilation - and yet most of the time it is enough to use only it, so it's successfully studied in schools and undergrads
if output of LLM can be modeled (by intuition) as "some other being" for many practical uses *and model works* - then automatical blaming others for "using less precise model" and warning about it feels... strange
tibbar 21 minutes ago
I mean, we're pretty deep into Westworld/Blade Runner-style scifi at this point. It's actually a crazy, mind-bending question to try to grasp what is going on with chatclaudini at this point. Regardless of what labels we choose or properties we choose to affirm, we're far too deep into uncanny valley for it to be very helpful.
libraryofbabel 2 hours ago
It would have been nice to see some version of “I am very surprised by how far LLMs have come since I wrote the stochastic parrots paper, here is how I have revised my thinking.” But there is nothing like that and the author is just doubling down or trying to correct perceived “misinterpretations” of her work.
Meanwhile you have multiple Fields Medalists (Tau, Gowers) saying they’re very impressed by LLMs’ mathematical reasoning, something that the stochastic parrots thesis (if it has any empirically-predictive content at all) would predict was impossible. I doubt Tau and Gowers thought much of LLMs a few years ago either. But they changed their minds. Who do you want to listen to?
I think it’s time to retire the Stochastic Parrots metaphor. A few years ago a lot of us didn’t think LLMs would ever be capable of doing what they can do now. I certainly didn’t. But new methods of training (RLVR) changed the game and took LLMs far beyond just reducing cross entropy on huge corpuses of text. And so we changed our opinions. Shame Emily Bender hasn’t too.
Sigh.
marshray 31 minutes ago
The Parrots paper:
"Contrary to how it may seem when we observe its output, an LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot."
So perhaps this has always been a negative claim, about what language model AI is not.
libraryofbabel 4 minutes ago
Maybe, but a claim about what and LLM is not is still a claim about what it can or cannot do. And specifically:
> without any reference to meaning
is vague, but I read it as actually quite a strong claim about the limitations of LLMs. I don’t think it would be possible for LLMs to do long chains of correct mathematical reasoning about novel problems that they haven’t seen before “without any reference to meaning.” That simply isn’t possible just by regurgitating and remixing random chunks of training data. Therefore I consider the stochastic parrots picture of LLMs to be wrong.
It might have been an accurate picture in 2020. It is not an accurate picture now. What is often missed in these discussions is that LLM training now looks totally different than it did a couple years ago. RLVR completely changed the game, allowing LLMs to actually do math and code well, among other things.
gessha an hour ago
The appeal to authority is strong here. A tool stochastic parrot can be useful too.
tootie 2 hours ago
She says explicitly it's not an empirical hypothesis. It's just a label for how they function. Which hasn't really changed even as they've gotten more useful. I haven't followed the full drama but this post is her saying the term has been frequently misapplied and she's basically distancing herself from some critiques that were misinterpreting her intent.
libraryofbabel an hour ago
> She says explicitly it's not an empirical hypothesis. It's just a label for how they function.
Then… what’s the point of the label, if it’s not making any empirically-meaningful claims about LLMs at all? I know that LLMs involve sampling over a distribution of output logits. I’ve written code to do it. So what? I know they have statistical elements. Yet I don’t go around calling LLMs stochastic parrots, because that label implies a whole lot of claims about LLMs that I don’t think are true any longer, like that they are just regurgitating and remixing training data and can’t successfully model structured systems (like mathematics or programming).
roenxi an hour ago
mbauman 2 hours ago
> stochastic parrots thesis (if it has any empirically-predictive content at all
Did you read TFA? This is precisely one of the non-questions that she answers.
kalkin an hour ago
Yes, she addresses this by denying that she's made any empirical hypothesis, but in a way that's some combination of disingenuous and confused.
She also says:
> What I am trying to do... is to help people understand what these systems actually are
Can a phrase that has no empirical content aid people in understanding an empirical phenomenon?
> the astonishing willingness of so many to... turn to synthetic text... for all kinds of weighty decisions.
Why is this astonishing, if the nature of these models as "stochastic parrots" places no limitations whatosever on their empirical capabilities, reliability, etc?
> the field of linguistics is particularly relevant in this moment, as a linguist’s eye view on language technology is desperately needed to help make wise decisions about how we do and don’t use these products
Is it wise to make decisions about a product on the basis of information that has no relevance to how it is actually likely to behave?
(It may be, if one has ethical concerns with "data theft, the exploitative labor practices", etc -- but one could have such concerns about any kind of product, not just a "stochastic parrot", and linguists are not generally academia's experts on, e.g., labor practices.)
harpiaharpyja an hour ago
...did you read TFA?
seatsh 2 hours ago
Gowers, Tao and Lichtman are especially impressed by the funding of math.inc and the AI for Math Fund, a joint venture of Renaissance Philanthropies and XTX Markets.
Renaissance Philanthropies is a front for VC companies.
They never publish allocated computational resources, prior art or any novel algorithm that is used in the LLMs. For all we know, all accounts that are known to work on math stunts get 20% of total compute.
In other words, they ignore prior art, do not investigate and just celebrate if they get a vibe math result. It isn't science, it is a disgrace.
newtonsmethod 2 hours ago
Is your justification in dismissing Fields medalists that they are impressed by funding? Not even receiving it (I assume you say this because Tao is not funded by AI for Math, but rather an advisor for it)?
Not only would it be a leap to suggest that people automatically lose their integrity by taking funds for projects they believe are useful, especially after involvement with adjacent fields, but you are suggesting merely being impressed by a fund is enough to dismiss their views?
You also have no evidence that Renaissance Philanthropies is a front for VC companies. All news coverage indicates that they seek to be an alternative for high net worth individuals engaging in philanthropy.
Many people discovering Erdos results, engaging in Olympiads etc, are doing so with publicly available models and publish the resources used in the process.
sdf127 2 hours ago
leonidasv 2 hours ago
What a hill to die on.
_wire_ 3 days ago
Lovely article well worth attention by virtue of its regard for the cultural traits of terminology and its inflections, while also debunking the pervasive lore that "AI" devices are doing anything but the merest resemblance of thinking.
It's rare to read an author who can directly face Brandolini's Law of misinformation asymmetry and not only hold his own against the bullshit but overcome it.
CamperBob2 3 hours ago
TIL that the "merest resemblance of thinking" is enough to take gold at IMO.
radkZ 2 hours ago
Automated theorem provers are not new, in fact they are very old. One of the most automated is ACL2, which uses the well studied waterfall method (unrelated to waterfall development).
LLMs certainly use something similar, except they understand text as input. LLMs, especially used for marketing stunts, have way more computing power available than any theorem prover ever had. They probably do random restarts if a proof fails which amounts to partially brute forcing.
Lawrence Paulson correctly complained about some of the hype that Lean/LLMs are getting.
ACL2 even uses formulaic text output that describes the proof in human language, despite being all in Common Lisp and not a mythical clanker.
They do not think and use old and well established algorithms or perhaps novel ones that were added.
nsingh2 2 hours ago
scotty79 3 hours ago
And also create novel math proofs.
tom_ 2 hours ago
radkZ 2 hours ago
This is the first submission since a year that gives me some hope for humanity. It shows that linguistics is not obsolete. Maybe the last people capable of thinking will be linguists.