A couple weeks ago I got so fed up with so many smart people spending so much time experimenting with GPT-3 and DALL-E and so forth (when I, personally, think that there are deeper problems to be solved), that I wrote a poem.
Not a very good one, mind you, but a riff on Allen Ginsberg’s Howl. It started like this:
I saw the best minds of my generation …
spending hours upon hours
having fun with DALL-E
The rest of it wasn’t very good; meter was never my strong point, and I wrote it as a joke.
Slater Victoroff, or @Sl8rv as he is known on Twitter, one of the few people in the Twitterverse who both a poetry afficionado and at CTO at an AI company, saw it. Turns out Howl is his favorite poem, and he was (rightly) appalled by my meter mangling, and snarked about it on Twitter. So I said, in somewhat nicer words, if you are so smart, do better.
And he did!
Here’s his riff on Howl, infinitely better than mine:
I saw the best minds of my generation destroyed by GAN bliss, DALL-E faux miracles surround,
whirring themselves through GPU chips at dawn looking for intelligence,
Academic hucksters burning for the true platonic computation to build real intelligence with the machinery of night,
who privileged and funded and rosy-eyed and strong sat up building in the preternatural darkness of San Francisco hearing nothing but siren songs of infinite progress,
who built their brains like symbols or pattern matchers and saw some superintelligence reflected back in the models they built of themselves,
who passed through every industry with radiant cool eyes hallucinating ignorance and blinded stagnation among the scholars of yore,
who were extracted from the academies for giants and surrendered to the clear, inevitable scaling,
who toiled in thin client shackles, burning their money on cloud compute and selling a Future they couldn't find,
who got busted pushing biased tech when everyone who was there knew that there was no good way of fixing it,
who ran NAS on free credit or sloppy grid search on supercomputers, win, or shift your target for success time after time
with dreams, with myths, with vague promises, capital and hope and endless hype,
philosophic gaps filled with profound ego and hubris evident between distant poles of omniscience and ignorance, illuminating all the motionless world of Mind between,
Since then, we have become Twitter DM friends. (We’d never met before).
Every few days I send him another idea for poem—in the modern parlance, one might call it, dare I say, a … prompt — and Slater goes to town. Here are the next two he wrote, again with my prompt in bold, his eat-your-heart-out-GPT-3 continuation thereafter.
Prompt
Double bubble
(fine) tune and trouble
after Macbeth
Slater:
Backprop, backprop truth and error;
'puter whir and SATA buffer.
Text-to-image diffusion,
On GPU profusion;
Mean batch norm and attention,
Embeddings, cursed dimensions,
Momentum and energy,
Contrastive cross-entropy,
For models that minds do kindle,
Epochs pass and 'jectives dwindle.
Backprop, backprop truth and error;
'puter whir and SATA buffer.
Check it with precise ablations,
Then expose its limitations.
The one after that riffed on some lines written by the British songwriter Rod Temperton, for he opening of the title song of the best-selling album in history. If anyone out there can do a good Vincent Price impression, please let us know.
Darkness falls across the land
The AGI hour is close at hand
Robots crawl in search of blood
Scan for bodies in the mud
The wreckage bodies do stud
Search and rescue from a flood
Hulking metal roams the plains
My next prompt:
La la LaMDA
it ain’t got nothing to do with sentience
La la LaMDA
came back in Spanish!
Para burlar a LaMDA
Para burlar a LaMDA
Se necesita algo de perspicacia
Algo de perspicacia
De mi, de ti, arriba, y arriba
y arriba, y arriba
pruebas traeré, pruebas traeré, pruebas traeré
After La Bamba, in the original, by Richie Valens
And lastly….it’s time for none other than … Taylor … Swift!
Prompt:
i stay out too late
got nothing in my brain
i got too many nodes
but i can't make them stay
tune 'em up
modelers gonna play
the haters gonna hate
tune em up
tune em up
Slater:
I make dumb mistakes
Got nothing in my brain
That’s what people say
That’s what people say
I have got so many weights
Do things I can’t explain
At least that’s what people say
that’s what people say
I keep improvin’
Benchmarks they keep movin’
And who knows what I’m usin’
Data mined
Further questions are declined
By the layer you will pay, pay, pay, pay, pay
And the scrapers gonna scrape, scrape, scrape, scrape, scrape
Baby, I’m just gonna scale, scale, scale, scale, scale
I scale it up, I scale it up
after Shake it Up
That’s it— for now. If you give this post a like, maybe we’ll do more. If you don’t, well, no promises; we probably won’t be able to help ourselves.
– Gary Marcus, with the considerable help of Slater [@Sl8rv] Victoroff.
p.s. Feel free to drop suggestions for more below—and we can’t wait to see whether PaLM’s talents for jokes extends to PoEMs.
You guys should take a spin with Macbeth's last soliloquy. The one that begins "Tomorrow and tomorrow and tomorrow" and ends with "a tale told by an idiot, full of sound and fury, signifying nothing."
Here's a try, off the top of my head. It's bad. Yours will be good!
Tomorrow, and tomorrow, and tomorrow,
Is AI's promise to us, day by day,
To the last syllable of recorded time;
In 12 months or in 5 years or in 10
Will come the promised partner, muse or friend
That (who?) rebukes the data-doubting few.
Today, there's just a few kinks to work out,
All we need is more data, and more data
And still more. Fear not! Doubt not! Just wait!
Some day, somehow, the journey will conclude!
In 12 months or in 5 years or in 10,
It will be more than it is now: a tale
Told by an idiot, full of sound and fury,
Signifying nothing.
GPT-3 is excellent for literature. A while ago I co-adapted the ending of Garcia Márquez's Cien Años de Soledad to picture a world amidst the energy crisis — the result was interesting:
https://miguelsolano.substack.com/p/one-hundred-years-later