What Emily Bender meant by "stochastic parrots"

> in part because Google fired two of the authors, Timnit Gebru

I remember being angry about this situation when I first saw it on social media, until I read the details: This person submitted a list of demands to her employer and said that if they weren’t met, she quit. Google wasn’t going to meet her demands so they considered it acceptance of her resignation. There has been a movement trying to debate whether it was a firing or resignation ever since.

The original paper they published gets recirculated every year or two as some landmark history of AI safety, but as other commenters have noted it wasn’t really a great paper nor was it groundbreaking at the time. If not for the controversy surrounding the resignation/firing (depending on your POV), I don’t think it would have been notable.

Here is what Jeff Dean said about the firing at the time: https://docs.google.com/document/d/1f2kYWDXwhzYnq8ebVtuk9CqQ...

> With the octopus thought experiment, I initially had told the story in terms of a dolphin, because dolphins clearly are intelligent animals. My co-author on that paper, Alexander Koller, said it should be an octopus, because first of all, the environment that octopuses live in is much more distinct from where people live. It makes the metaphor more vivid, that the octopus is just feeling these pulses in the cable and has no way to look at what the people are looking at.

On a completely tangential sidenote, octopusses are actually very very intelligent: https://www.nhm.ac.uk/discover/octopuses-keep-surprising-us-...

I paid a bit of attention to this paper and the phrase 'stochastic parrots' when it came out and i thought this was worth saying and doing at that time. their suggestions about financial and environmental costs are worth studying, their concern about carefully evaluating datasets to feed to the model rather than feeding the entire internet is fully justified. so - to everyone saying this was a bad paper; if you have actually read the paper then please list a few criticisms. all i have seen is "oh this wasn't that good of a paper" or "can't believe how bad this paper was".

Bender's paper had this to say about stochastic parrots:

"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."

This was not even a correct criticism in 2021. She is right that, at the time, the pretraining -- where it learns to predict missing words in pre-existing corpuses of text -- is basically a stochastic parrot.

But nowhere in her paper does the term "reinforcement" come up. At the time, this was done mainly through RLHF (reinforcement learning from human feedback) - after the initial training is done, you then tune the model's responses based on human grading. Humans imbue their own meanings into the parameter weights through their judgment.

At this point, they aren't really stochastic parrots anymore, because parameter weights have been shaped beyond the text corpus. It's not purely probabilistic in the sense of using the probabilities of the underlying text sequences. (It still is probabilistic in its output, but that is a pointless claim, because all events in the universe are also probabilistic; it is not enough to merely claim that probability is involved in some way in the outputs.)

RLHF was already in use prior to the paper, and was written about by Christiano in 2017 "Deep reinforcement learning from human preferences," so it's surprising that Bender apparently didn't know about this well-known paper.

RLHF was also, of course, a precursor to a more advanced form of parameter shaping - reinforcement learning with verified rewards, or RLVF, which has driven a lot of the gains in verifiable domains lately. That was not done in 2021 when she wrote the paper. But if you knew about RLHF -- and knew how Alpha Zero worked, with training neural nets on game rollouts -- you could squint and see that it might be useful for language models.

So after being proven to not only having a limited understanding of the field at the time, but also not being able to forecast the field, she's now walking back what she meant by "stochastic parrot," I assume because she believes readers will not read what she wrote. But despite the protests, her original claim was that it is a parrot because the text has no meaning -- a direct quote from the paper, which only really makes sense if training stops at the pretrain.

I don’t see a problem with the “stochastic parrot” label. It just turns out stochastic parrots are incredibly useful.

At a minimum it’s probably more accurate than “AI”.

Personally, I've always read that paper as a political criticism of industry and industrialized research and capitalism. After decades in academic (and industrialized research) I've learned that smart people can write convincing takedowns of things they hate- and those takedowns, due to being well written, often punch above their weight in terms of impact on the community.

I think this paper would have been best split off from the conjoined criticism of environmental effects (which could have been its own paper, but not one published by Google, since their leadership's fundamental beliefs disagree with the paper's environmental impact premise. And the remaining part on text models could have been a bit more focused on the technical issues associated with statistical text processing and meaning, rather than criticism of the power structure that is loosely associated with the current AI push.

I'm sorry but I do tend to feel like this muddies up the discussion on "what this technology really is".

I think "artificial" is actually a pretty good term to describe the output of the models. That output does appear to resemble at least some definition of the word "intelligence" - there is some ability there to do cognition over information that's been provided to them in-context.

What is it to understand, then? If they can work in complex domains and produce coherent output, it would seem to necessitate at least some definition of "understanding" of the corpus, even if that understanding is unlike how a human's brain would understand it.

What else should we call them then? They model language and information in ways that allow them to manipulate it on the fly. They do so 'unnaturally' from a human's point of reference.

I legitimately can't come up with a better term than 'artifical intelligence' -- not to be confused with artificial consciousness, which I don't think exists (yet).

This all sounds like a lot of backpedaling and “well actually” kind of stuff.

“Stochastic parrot got picked up and interpreted by other people as a minimization or an insult. It was not meant that way. Other people might be using it that way but that’s not how I intended it”.

Yeah that’s because it was chosen to be an insulting phrase.. Parroting is only ever used as a pejorative phrase. But sure, everyone else mindlessly parroting this line is the problem here.

This paper was always lousy, but it has really not aged well. We are living in a world when where an LLM has solved an Erdos problem. In a world where LLMs produce novel results that rival human thinking any conceptual reduction of an LLM is going to start inviting some unpleasant comparisons with human thinking.

Her language consistently defines LLMs in negative terms like “synthetic text extruder” but she claims she’s not trying to denigrate it. What’s missing for me are similar terms from her about how humans create sentences and thoughts. Judging by the state of the internet humans are quite capable of making shit up to argue their point (see latest Fox News apology). She talks about sycophantic AI but give me a car battery and some cables and I can train a sycophantic human (no I can’t but there are people who can). She’s pretty much a walking counter argument for her own claims.

Doesn't really matter that much what they're called as long as they're useful, and LLMs (particularly when harnessed) are already ridiculously useful. But it also begs the question: are stochastic parrots useful?

> when OpenAI imposed ChatGPT on the world...

OpenAI offered ChatGPT to the world. A large, monied cross-section of the world had yet to throw its capital behind the Large Language Model technology that made the ChatBot possible. While it is fair to see AI development now as a global imposition, OpenAI did not have the agency as a 2022 startup to impose on the scale we see now.

I respect you and parrots, please don’t use parrots as an insult.

Five years on, which term do we see as less accurate to describe LLMs? Artificial Intelligence or Stochastic Parrot? I guess it's still an open debate.

it annoys me how eager people are to hurl the word stochastic as pejorative. Statistics are a great tool for gleaning information from stochastic processes; statistics don't contribute randomness. Random sampling is necessary in order not to bias a sample, it's not used to contribute randomness to the sample but to preserve/measure the underlying distribution. (not meant to imply that training is random sampling)

[dead]

[flagged]

[flagged]

[dead]

'Stochastic parrots' is a great term, but reading it now, it's quite apparent how bad this paper is.

The term is not very useful since most humans are stochastic parrots... At least most of the time.

Not suggesting that I don't say stuff on autopilot sometimes but for many people, it's their only mode of operation. They never actually think about anything from first principles. Their whole approach to language is just chaining catchphrases together. It's how a toddler thinks; it seems like many people never moved past that stage of development.

> It argued that large language models (LLMs) generate text by statistically predicting likely sequences of words rather than understanding what they are saying—a process the authors captured with the metaphor of a “stochastic parrot,” a system that repeats patterns without comprehension.

I don't understand what we're setting the record straight on. This is the core point of dispute, and the author just blazes past it to focus on other things. I'm glad to hear "stochastic parrot" isn't intended as an insult, and I agree that it's not right to think of LLMs as a box with a little homunculus inside replying to you. But to me it seems obvious that LLMs are not repeating patterns without comprehension and do understand what they are saying; otherwise they would not be capable of doing things they routinely do.

For context, here's the main quote:

> 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.

I think this metaphor is so strained as to not be useful. I think key here is that the authors say "without any reference to meaning", which is a heavily loaded term, that does definitely apply to parrots, but does not apply when you apply it to immense bodies of text.

Namely that language embeds meaning in language. A sentence being written by a human (as a starting point) is designed to have consistent meaning. While it is possible to write syntactically correct meaningless text, that is not what most of human language has done; the meaning cannot be removed from the text.

This I think is clarifying, from the same paragraph in the text:

> ... the training data never included sharing thoughts with a listener, nor does the machine have the ability to do that.

That's just facially incorrect. The training data is entirely about sharing thoughts with a listener. Else why is the text being written?

What I have been doing in many places—the octopus thought experiment, stochastic parrots, the phrase “synthetic text-extruding machines”—it’s all about trying to make vivid to people who aren’t in the business of building language technology what these systems actually do

> Meanwhile, O, a hyper-intelligent deep-sea octopus who is unable to visit or observe the two islands, discovers a way to tap into the underwater cable and listen in on A and B’s conversations. O knows nothing about English initially, but is very good at detecting statistical patterns. Over time, O learns to predict with great accuracy how B will respond to each of A’s utterances. O also observes that certain words tend to occur in similar contexts, and perhaps learns to generalize across lexical patterns by hypothesizing that they can be used somewhat interchangeably. Nonetheless, Ohas never observed these objects, and thus would not be able to pick out the referent of a word when presented with a set of (physical) alternatives.

This seems kind of obviously wrong at least in the context of coding agents. These models get trained on actual output of the previous version of the model doing its job, often "IRL" on a real computer/project. It's like O is in the conversation for years now and learning from his own interactions between A <-> O <-> B, where A is the human and B is the computer.

The idea O ontologically has never "observed" "these objects" or referents is philosophically strained. Have I observed the moon, or a finger pointing at the moon? Have I observed `sed` more than Fable?

I think this is the most measured take I've seen from Bender, and I think it summarizes her only compelling point well (technologies should be referred to specifically rather than generally as AI, and that referring to everything as AI is not useful and helps hype the technology in a way that benefits those selling it).

In her previous interviews, I've found her assertion that LLMs aren't useful and will never be good at anything totally uncompelling. Also laughed at this quote as she's been pretty harsh IMO on "the people who like the systems".

> it’s all about trying to make vivid to people who aren’t in the business of building language technology what these systems actually do, which is not the same thing as insulting the systems or insulting the people who like the systems.

After having used LLMs for some time now, I don't agree with the concept they are just token generators, unless you think that's all humans are too. The way we test in most schools is just picking the right token. We also give them unique problems that they never saw in their training, which is the nature of programming. I realize they are probabilistic token generator models, but I find it harder and harder to accept that somehow there isn't something more going on. I'm not sure whether they are intelligent or not, but for the most part token generation is how you get degrees too. The thing is a parrot just says things it has already heard, it doesn't perform complex reasoning on novel situations and then explain it succinctly.