[453qtgreq]

Your original comment made a big deal of a poem-writer being able to incorporate unique words that you gave to it in the first place. No, that's not very interesting.

Writing a "4-line humourous poem" is also trivial. I'm sure my four year old could manage it. Praising a machine learning algorithm for this, that took over 10,000 Exaflops to perform its training routine seriously demeans human ability.

[fenomas]

I didn't claim the poem was interesting, I gave it as an example of something which could not possibly have been regurgitated straight out of the training set, which it clearly was.

Also: it was a 4-stanza poem, three rhyming couplets per stanza. It had a recognizable beginning/middle/end and a reasonably humorous punchline. The rhyme scheme was AABBCC - your four-year old will need to know that when they replicate the feat ;)

[453qtgreq]

You're amazed because a training set, largely based on the contents of the internet since about 2011, is able to find words that rhyme with arbitrary word endings?

4 stanzas. My bad. Still not impressed.

[fenomas]

> You're amazed because [stuff nobody said]?

Comments like this add nothing to the conversation, please consider not posting them.

[453qtgreq]

>stuff nobody said

>Given that the names were made-up and unique, and have zero hits on google

You said it yourself. I just don't see the novelty that a machine learning algorithm, trained on huge amounts of stuff scraped off the internet, can find rhyming words.

That aside, there's an enormous amount of poetry in the training set. Given the relatively strict constraints of the style, there's not much interesting about slotting in words that grammatically make sense.

[fenomas]

What I said: that outputting a poem structured around having end rhymes for rhyme made-up words was an obvious counterexample to the claim that the system is just regurgitating its training data. What I did not say: that I was amazed because it rhymed arbitrary words. HTH.

[453qtgreq]

>an obvious counterexample

ChatGPT uses sub-word tokens. This is not a surprising outcome.