Let me preface this by saying I'm stupid, I can't even do my own research work well, let alone comment on cutting edge stuff with any degree of confidence:

I don't think this is incompatible with the concept of the stochastic parrot. Like, by the time "On the dangers of stochastic parrots" came out, it was already known that language models have rich representations of language structure:

(from the stochastic parrot paper):There are interesting linguistic questions to ask about what exactly BERT, GPT-3 and their kin are learning about linguistic structure from the unsupervised language modeling task, as studied in the emerging field of ‘BERTology’ [...]

So I don't think we can take this, or the probing/interpretability work later in the paper as a refutation of LLMs as stochastic parrots, because it was never about memorization:

(From the ICRL paper the blog post is based on): A potential explanation for these results may be that Othello-GPT is simply memorizing all possible transcripts. To test for this possibility, we created a skewed dataset of 20 million games to replace the training set of synthetic dataset[...] Othello-GPT trained on the skewed dataset still yields an error rate of 0.02%. Since Othello-GPT has seen none of these test sequences before, pure sequence memorization cannot explain its performance.

the concept is useful primarily as a way of delimiting how far this "understanding" really goes:

(stochastic parrot paper again) If a large LM, endowed with hundreds of billions of parameters and trained on a very large dataset, can manipulate linguistic form well enough to cheat its way through tests meant to require language understanding, have we learned anything of value about how to build machine language understanding or have we been led down the garden path?

the metaphor of the crow is kind of apt, I think. Like an LLM, it is working only with form, not meaning:

(blog post) At this point, it seems fair to conclude the crow is relying on more than surface statistics. It evidently has formed a model of the game it has been hearing about, one that humans can understand and even use to steer the crow's behavior. Of course, there's a lot the crow may be missing: what makes a good move, what it means to play a game, that winning makes you happy, that you once made bad moves on purpose to cheer up your friend, and so on.

(stochastic parrot paper) Furthermore, as Bender and Koller argue from a theoretical perspective, languages are systems of signs, i.e. pairings of form and meaning. But the training data for LMs is only form; they do not have access to meaning. Therefore, claims about model abilities must be carefully characterized.

Someone smarter please feel free to correct/dunk on me.