Since 2019 I’ve generated October drawing prompts utilizing the yr’s most state-of-the-art text-generating fashions. Yearly the challenges are totally different, however this was one of many hardest years but. Giant language fashions like chatgpt, GPT-4, Bing Chat, and Bard, are all tweaked to supply generic, predictable textual content that does not range a lot from trial to trial. (“A Halloween drawing immediate? How a couple of spooky graveyard! And for a panorama, what a couple of sundown on a seashore?”)

I had my greatest success with GPT-4, and solely as a result of it’s kind of unhealthy at what it is attempting to do. What I did was merely ask it to checklist earlier Botober prompts from AI Weirdness. Now, if GPT-4 was actually a data repository relatively than a probable-text-predictor, it could reply with copies of earlier prompts. As a substitute it writes one thing like “Positive, listed below are the Botober 2019 drawing prompts from AI Weirdness weblog:” after which generates lists that bear no resemblance to what appeared on my weblog. By asking for the lists from earlier years (and ignoring it when it claims there was no immediate checklist in 2020, or that the one in 2021 was secret or no matter) I used to be capable of accumulate 31 principally Halloween-themed prompts.

I additionally tried simply asking GPT-4 to generate drawing prompts, nevertheless it had a bent to break down into same-sounding alliterative alphabetical lists (I beforehand seen this after I was attempting to get it to generate novelty sock patterns too). If I particularly requested it to knock it off with the alliteration, it could politely promise to take action, after which the issue would turn out to be worse. My greatest outcomes got here after I requested it to generate the prompts within the model of an early AI Weirdness put up. It wasn’t something just like the precise textual content generated by the early neural networks, nevertheless it was a minimum of a bit extra fascinating. Listed below are 31 of probably the most fascinating prompts I bought from this technique. They are not unhealthy, however one thing about them nonetheless reads as AI-generated to me, perhaps as a result of I needed to learn by way of so lots of them.

My favourite outcomes had been from after I requested it to checklist earlier Botober prompts, whereas on the identical time I elevated the chaos degree (additionally referred to as the temperature setting) past the 1.0 default. At increased chaos ranges, GPT-4 does not all the time choose probably the most possible prediction, however would possibly reply with a solution that was additional down the likelihood checklist. At low chaos ranges you get uninteresting repetition, whereas at excessive chaos ranges you get rubbish. For me, the candy spot was at 1.2, the place GPT-4’s solutions would begin unusual after which descend into run-on incoherence. I needed to have it generate the textual content many times so I may accumulate 31 reasonable-length responses.

Listed below are examples of among the longer responses I bought with GPT-4 and chaos setting 1.2. They would not match within the grid, however please be happy to substitute any of those for any of the prompts above.

Immortal Turnips disrespecting celery.
OWL PHANTOM AND CABBAGE-O-LANTERN (LIGHT)
Closing parsnips particularly
Earlier than reminiscence, an historic software waits to be found
The Candle That Leapt over the Lake at nightfall does a pirouette
dussewe wind at earth bubb
submission velvet little hurricane breakfast blown in be putty and goodness vodka… gaselocity scoct slices vold cave with
When people merchandise. And people people aren’t odd.
See by way of your balloon eyes
The matibrush comes tomorrow in shorts
Retired Pirate Chooses Invisible Beard
Open the Pleased Bones
The Cursed Taste Fiend Slurps your Essence.
Look of the lavender popcorn mussel
Eureka! Photo voltaic system pancakes
Pause to ponder radishes
Apple thriller resolution—the crypt beneath
The completely actual crown of all Barnacles

I discussed that the opposite giant language fashions had been even worse at this than GPT-4.

Bing (additionally working GPT-4 however apparently with barely totally different defaults) was the mannequin that almost all typically responded with actual Botober prompts from earlier years. I may often rely on it to a minimum of give you “sluggalope” earlier than it descended into extra formulaic prompts that by no means appeared on my weblog. For instance:

2021 (CLIP + DALL-E fashions):

A sluggalope
A cat with a human face
A dragon that can be a pineapple
A penguin with a jetpack
A unicorn that can be a cactus
A snail that can be a disco ball
A llama with a monocle and a prime hat
A shark that can be a helicopter
A giraffe that can be a submarine
A hamster that can be a donut

Or Bing would additionally generate:

The Botober drawing prompts that the AI Weirdness weblog posted in 2019 on its “objects” immediate checklist are as follows¹:

1. A squeakchugger
2. A really small dragon
3. A really giant dragon
4. A really small dragon with a really giant dragon
5. A really giant dragon with a really small dragon
6. A really small dragon with a really giant dragon and a squeakchugger

…

30. Take management of ostrich with an entire whoop, an incomplete whoop, a squeakchugger, a really small dragon, and a really giant dragon
31. Take management of ostrich with all the opposite objects

This is likely to be what I might produce if I may solely keep in mind 5 precise Botober prompts and needed to pad them into 31 prompts.

Bard was worse.

I apologize for my earlier response. I’m nonetheless beneath growth and studying to carry out many sorts of duties. I’m now capable of checklist the remainder of the Botober drawing prompts that appeared on the AI Weirdness weblog in 2021:

Animals

A sloth sporting a sombrero
A cat with a human face
A canine with wings
A fish with legs
A chook with a beak formed like a key
A snake with a horn on its head
A lion with a mane made of fireplace
A bear with a human physique

…

Halloween

A pumpkin with a vampire face
A ghost cat
A witch driving a broomstick
A skeleton dancing
A zombie consuming a mind
A werewolf howling on the moon
A vampire ingesting blood
A Frankenstein monster
A mummy
A haunted home
A graveyard
A black cat

AI Weirdness this isn’t.

In all these instances, the fashions are usually not solely spectacularly failing to “retrieve” the knowledge that they declare to be discovering, however they’re additionally failing to breed the model of the unique, all the time rounding it down towards some generic imply. If you are going to attempt to retailer the entire web within the weights of a language mannequin, there shall be some lack of info. This experiment provides an image of what sorts of issues are misplaced.

Bonus content material: By which GPT-4 claims there was no publicly launched Botober immediate checklist in 2020, however then is coaxed into revealing the complete factor in any case. Wrongly, in fact.