alt Hacker NewsI 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".
Replies
My criticism centers on the part of the paper they chose for their title, the “stochastic parrot” metaphor. And my criticism is that if you observe Claude code with opus 4.8 working through an entirely novel problem that nobody has ever worked on before and which certainly wasn’t in its training data, the choice to even metaphorically call them stochastic parrots turned out to be egregiously wrong.
And secondarily, and maybe only partially the authors’ fault, is the enormous tidal wave of morons that this paper minted who plague us with their misunderstandings to this day.
My main criticism of the paper is that it says LLMs work "haphazardly", using probabilistic information. That is a hypothesis, but it is stated as a known fact, a fundamental limitation.
It is true that LLMs often behave haphazardly, and do rely on statistics. But plenty of research has shown them behaving in methodical ways too. There are findings going both ways!
Granted, many of the strongest contradictory results appeared after the Stochastic Parrots paper, so it isn't like they were ignoring the literature at the time. But they did make a very strong claim, and in the half-decade since, a lot of evidence has come out against it.
Those costs have to be compared to the way things are currently done without AI.
They never are. Ever.
It is a good blog, not a good paper.
The contention that there is no grounding because the training data is linguistic and thus can only reference a world model is disproven in "This sentence has five words"- there's real, grounded information about what "five" means within that sentence. While that's a trivial counterexample, I don't know that it's an obvious one (I didn't come up with it myself).
It's not a criticism of the paper itself, but multimodal models came shortly after and provide grounding that is more of the sort the paper is getting at, and it didn't seem like anybody updated on that at all. If multimodal models were still stochastic parrots by the original argument, humans would have to be as well; we don't have any way to ground anything beneath sense data and evolution can't have programmed some innate grounding into us because it didn't either. But (and maybe this is my own misperception) nobody threw in the towel at that point.
I confess I never read the original paper until now, opting to absorb by osmosis instead, and I was quite surprised that they don't really make a deeper case than that. After just a few paragraphs about how they can't be grounded because humans don't express their thoughts directly, it lurches into a page about how they can be biased by training. And they certainly can be, but that has little to say about their stochastic nature- humans are biased as a rule with no exception. (For the record, I only read the Stochastic Parrots section before this reply.)
It's not really a bad paper, but I don't see why it ever carried the esteem it did. Hating on it is like hating on Taylor Swift- she's fine, yes, but for her level of success, one is inclined to question every dumb lyric where others get a pass. (Apologies to Swift fans, substitute a successful artist you don't care for here.)