[xwolfi]

Have you talked to an angry redneck in a dank pub ? He'll do the same. And not just him, but whomever is expected to speak but has no high value things to add, and is a little bit dumb/drunk to count how many times he said something and stop at 3.

Maybe information, to be interesting to us, has to be novel, while GPT-3 may not model for listener's interest (like you when you re drunk) and only produces the best it can express in a given input context ? And sometimes, maybe repeating 34 times the same thing is good if no new input changes the fundamentals, just not very interesting for a signal dampener like our brain who starts losing focus when novelty disappears from the signal?

It s like imagine a political debate around building a bridge between a truck driver who wants to go faster and a bird watcher who wants birds to keep their habitat close to his home. There's no input that can change the fundamentals and it would be expected that after a few loops, no brain could find anything to add and just repeat forever the same thing: but the birds must be close to me or I lose my life's meaning, but the bridge must be built there or I cant optimize my route. The only thing we do is put a time stop and say "ok we got it, now everyone in the public can map their own constraint to the discussion and vote".

[skybrian]

No, that's not it at all. GPT-3 loops within a sentence, creating an infinitely long sentence of gibberish.