ChatGPT (and I assume other LLMs as well) uses an embeddings layer.
This is how it can translate concepts between domains. It is working with abstractions in a high dimensional semantic space. (12288 dimensions for the Davinci model)
they say: >>Embeddings are numerical representations of concepts converted to number sequences, which make it easy for computers to understand the relationships between those concepts.
Beyond that claim, there is nearly no information on how this works, and the examples give hints of not any kind of concept, but an additional abstraction layer of clustering words with similar meanings. It is not the abstraction of an understanding, but a more far-reaching thesaurus lookup.
(edit) It's an astonishingly good map of words, and this makes it better. The clusters may sort of map resembling concepts, but the kinds of errors it makes show that it's not actually wielding concepts to think.
That said, a journey of a thousand miles begins with a single step, and this looks like a good step.
In the middle of the road, it outputs very sensible, grammatical, and mostly accurate text. It looks impressive, and can be genuinely helpful.
As the same questions get into more obscure stuff, it starts making errors, then on anything near advanced cases, it hallucinates/confabulates/lies.
Even with simple math, you get this:
Q: what is 17 plus 34? A: 17 plus 34 equals 51. (correct) Q: what is 297438 plus 927465? A: 297438 plus 927465 equals 1224903. (correct) Q: what is 38460 times 2398756? A: 38460 times 2398756 equals 9218928960. (wrong, correct is 92256155760)
It is merely outputting a very grammatical word salad that has the same relationship to the actual concepts as a broken clock to the time (right twice a day; this is just better with insanely more parameters).
In the areas where it has ingested massive amounts and has billions of relevant vectors, it'll output good stuff. But in the edges, where the input is more scarce, it'll just guess, and that is were you see that there is no abstraction there.
I've observed it doing the same thing with legal and scientific topics that I have or know experts. When there is a great body of text that it has ingested, the predictions and grammar engine and idiom are amazing, sometimes astonishing, and often useful.
But the kinds of errors show that it has no shred of a concept. It if did, it would not have the kind of problems with facts. We would not get entirely hallucinated/confabulated references to entire papers, authors, APIs. etc., including math answers. We'd get something like "I should have that knowledge, but it is out of scope".
Your example is very cool, but pretty much the same as translating from English to French, it's just car idiom to curling idiom. When the LLM translates "Car" to "voiture", it is not abstracting the concept of a car, it is making a prediction. When we translate as novices in a second language, we have to hunt around for the vocabulary equivalent, and when we speak a second language, we aren't translating, but speaking from our own concepts. LLMs are not human, and they learn and produce nothing like humans, but humans have an extremely powerful tendency to anthropomorphize and infer qualities that are not there.
Also, from an entirely different angle, in every paper on LLM technology I've read or skimmed, I've seen zero reference to any technology that would abstract and wield concepts.
So, I find it extremely unlikely that these models even come close to anything resembling use of abstractions.