[jchmrt]

You're missing the point of the parent comment: GPT-3 can understand this idiom because it was trained on a corpus in which the context for this idiom already existed. If a new idiom would emerge, the system would not necessarily be able to handle it if it can not understand it from the context it was trained on. Therefore, a translating AI needs to be continuously updated.

[kbelder]

Is that any different than a human? They can be pwnded by new idioms.

[ufmace]

I think Humans are pretty good at at least recognizing that a particular phrasing doesn't make sense as a literal statement and so must be a reference to something. Often you can get an idea of what it's meant to mean just by context. Sometimes if you see a dozen or so usages, you get the idea of what it means without ever having it explicitly explained.

[lupire]

One of the main differences between humans and ML is that humans learn in far fewer examples than machines.

"Poverty of the stimulus"

https://en.wikipedia.org/wiki/Poverty_of_the_stimulus

[andai]

Until the models become more sophisticated than human brains. Then humans will need more examples than a machine would to learn the same thing.

[jchmrt]

No it isn't of course, but that speaks for the argument that AGI is necessary for human-level translation :)

[andai]

Or just an up-to-date GPT?

[TillE]

Good luck accurately machine translating the infinitely growing idiom of jokey Twitter conversation, which most native speakers can pick up pretty quickly.

[xyzzyz]

Like, for example, “to be pwned”, which I believe is not going to be understood by majority of English speakers.

[mcguire]

And even fewer ML proggies.

[andai]

Ackshually... :)

https://files.catbox.moe/yvhqwd.jpeg