Hark! from those shadowy depths thy voice / Mournfully echoes, "AUTH".

November 6, 2019 2:44 PM

Fair is the Lake, and bright the wood,
With many a flower-full glamour hung:
Fair are the banks; and soft the flood
With golden laughter of our tongue

Poems written by the GPT-2 neural network, with copious notes on methodology.

Here are some highlights from an earlier run, by blogger Scott Alexander. via hacker news, god help me

posted by whir (23 comments total)

Vanity thy name is GP2 Neural networks, or is / was it, Shakespeare?
posted by Oyéah at 2:51 PM on November 6, 2019

I'm laugh/crying at work over here.
posted by Alensin at 2:52 PM on November 6, 2019

Well
My name is GPT-2 and I'm here to say
generate_unconditional_samples.py
a lemon to a lime, a lime to a lemon
before you run train.py you gotta import json

posted by Fiasco da Gama at 3:22 PM on November 6, 2019 [19 favorites]

This Is GPT-2 Say

I have crunched
the data
that were in
the textbox

and which
you Gwern probably
saving
for networks

Forgive me
it was del.icio.us
so uncanny
and so cobold
posted by chavenet at 3:50 PM on November 6, 2019 [12 favorites]

I quite enjoyed reading some of those aloud. a comfortable kind of uncanny valley, this, where lakes are fair and bright woods glamour-hung.
posted by Fraxas at 3:59 PM on November 6, 2019 [4 favorites]

In the dark the sun doth gleam,
And in the dark the moon doth seem
But now the evening is begun–
Gone is the sun upon the earth!
The silver moon doth like a cup
Of blood-red wine, and as that cup
Is drained of life, doth quench no drop.
What man will drink such wine?

That's not bad honestly, especially the end. Also the tenses are basically correct, which is frigging amazing IMHO.
posted by gryftir at 4:11 PM on November 6, 2019 [16 favorites]

Cyberiad draws nigh.
posted by Phssthpok at 4:15 PM on November 6, 2019 [6 favorites]

my name is GPT-2
and wen its times,
or wen the moon
it britely shines,
and all the men
haf lost ther minds-
i neural net.
i rite the rimes.
posted by It's Raining Florence Henderson at 4:20 PM on November 6, 2019 [34 favorites]

I fear that few return from bright, glamour-hung woods. Do not, under any circumstances, eat the food you find there.
posted by vibratory manner of working at 4:32 PM on November 6, 2019 [7 favorites]

...Let me drink of the wine of pain
And think upon the agonies of hope,

Adolescence is so hard.
posted by pangolin party at 4:34 PM on November 6, 2019 [9 favorites]

I'm bumblefied
posted by clavdivs at 4:48 PM on November 6, 2019

I made it to

Majestic yet majestic, ...

and then my poor cat thought I had lost my mind.
posted by fritley at 7:03 PM on November 6, 2019 [1 favorite]

thanks for sharing this--i did a phd on shakespeare's prosody but ended up a software engineer and this is one of those rare times when both halves of my life find the same thing fascinating
posted by josephtate at 7:53 PM on November 6, 2019 [6 favorites]

Oh limericks is it?

10    to make a rude lyric crescendo
20    shove a dictionary up a Nintendo
30    then wait for the bot
40    to process the lot
50    and highlight the best innuendo

posted by Fiasco da Gama at 9:38 PM on November 6, 2019 [15 favorites]

It might be due to me being an ESL speaker, but I found quite a few good bits in there. Many of the repetitions do not really work, though.
posted by Harald74 at 12:13 AM on November 7, 2019 [1 favorite]

Reading some of these gave me a really Wintermute/Neuromancer vibe.
posted by ninazer0 at 1:45 AM on November 7, 2019

This is the one that sticks in my head, from the "highlights from an earlier run":

The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), majestical,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious,
The Emperor Wu (the great Wu), rapacious
posted by chavenet at 3:22 AM on November 7, 2019 [7 favorites]

This is the one that sticks in my head, from the "highlights from an earlier run":

The To-Do List of the Emperor Wu (the great Wu)
posted by otherchaz at 4:54 AM on November 7, 2019 [2 favorites]

A few of these bits actually really work:

A constant waste of words, the world produces,
And
My heart, why come you here alone?
The wild thing of my heart is grown
To be a thing,
Fairy, and wild, and fair, and whole

As the article points out, that last line is an absolute gem no matter who is writing it.

And their eyes are filled with tears,
And their staves are full of woe.
And no light brings them any cheer,
For the Lord of all is dead

A bit dire, but I dig it. Finally, with just a touch of grammatical cleanup:
But none of the likes of his fellows
Are equal to him, and wherever he goes,
The heart somehow breaks under the hand that is steering;
And so it is with me.
Again, there’s some legit excellence there.

If it were me, I’d make two major modifications to the approach here:
First, every word becomes a struct containing four members: actual text for lookup/output, IPA pronunciation so that it can begin correlating sounds for rhyming and internal wordplay, a thesaurus-style categorical tag array for the word: ie sunny->brightness, joy, warmth, and finally a structural/usage tag array: imperative, mass noun, modal auxiliary verb, gerund, etc. With sufficient training text it should be able to figure out on its own not just double and triple meanings in wordplay, but also relative usage weights (ie gerund: 0.7). There’s going to be a fair amount of human labor required, here: 95% of that information can be generated with a handful of moderately complicated scripts, but for invented words and some of the more complicated structural markup you’ll need manual intervention.

Point being: poets have access to dictionaries and thesauri.

2) Scoring the output. +/- 1 for individual lines, 1-10 ratings for stanzas, 1-100 ratings for the poem as a whole.

Poets have access to audience reception and critical feedback, whether they like it or not.
posted by Ryvar at 5:01 AM on November 7, 2019 [4 favorites]

There's also scansion, which from my amateur investigations, seems like a difficult problem (there's a data problem, that it's difficult to find open source data with reliable information on stresses, and a poetry problem, which is that human beings often don't agree on scansion either). I guess GPT-2 works around all of that stuff by finding subtle statistical concordances in the training text and reproducing them, while not really "understanding" the prosody at all.
posted by whir at 6:58 AM on November 7, 2019 [1 favorite]

I recently read this exellent article on TS Eliot's criticism and the birth of the "biographical fallacy," and it struck me as containing a lot of food for thought in respect to poems generated by computer algorithms.

"Poetry is not a turning loose of emotion, but an escape from emotion; it is not the expression of personality, but an escape from personality."

[...] "The progress of an artist is a continual self-sacrifice, a continual extinction of personality."

[...] In the essay’s other most striking image (and claim), Eliot suggests that each work of art is part of a vast trans-historical system, a sort of virtual bookshelf containing “the whole of the literature of Europe from Homer”—one that might, at any moment, be rearranged by “the introduction of the new (the really new) work of art.”

posted by whir at 9:57 AM on November 7, 2019

whir: "Eliot suggests that each work of art is part of a vast trans-historical system, a sort of virtual bookshelf containing “the whole of the literature of Europe from Homer”—one that might, at any moment, be rearranged by “the introduction of the new (the really new) work of art.”"

I sat upon the shore
Fishing, with the arid plain behind me
Shall I at least set my lands in order?
London Bridge is falling down falling down falling down
Poi s’ascose nel foco che gli affina
Quando fiam uti chelidon—O swallow swallow
Le Prince d’Aquitaine à la tour abolie
These fragments I have shored against my ruins
Why then Ile fit you. Hieronymo’s mad againe.
Datta. Dayadhvam. Damyata.
Shantih shantih shantih
posted by chavenet at 1:41 PM on November 7, 2019

talktotransformer.com

Cat-Scan.com is one of the strangest sites I've seen in some time. I have no idea how these people got their cats wedged into their scanners, or why. However, the website's main source of content is a link to the Facebook page of someone named Tom who claims he has his own cat-scanning website (and apparently owns two more, at least). The site claims to have scanned thousands of cats and posted scans for sale. Apparently there are three kinds of scans—trending scans, pet scans, and pet scans of celebrities (Tom is, unfortunately, not a celebrity). But no one is really sure which type of scan they're talking about. They do mention that if you don't want your cat scanned, just tell them you have a cat. They also say that if you're not sure if you have a cat or not, then get a cat scan!

posted by tonycpsu at 9:19 PM on November 7, 2019