Alex Wool­ner knows how to put a degree in Eng­lish to good use.

Past projects include a fem­i­nist type­writer blog, retro­fitting stick­er vend­ing machines to dis­pense poet­ry, and a free res­i­den­cy pro­gram for emerg­ing artists at a mul­ti­dis­ci­pli­nary stu­dio she co-found­ed with play­wright and painter Jason Mont­gomery in East­hamp­ton, Mass­a­chu­setts.

More recent­ly, the poet and inter­na­tion­al edu­ca­tor has com­bined her inter­est in amigu­ru­mi cro­cheted ani­mals and Chat­G­PT, the open source AI chat­bot.

Hav­ing cro­cheted an amigu­ru­mi nar­whal for a nephew ear­li­er this year, she hopped on Chat­G­PT and asked it to cre­ate “a cro­chet pat­tern for a nar­whal stuffed ani­mal using worsted weight yarn.”

The result might have dis­cour­aged anoth­er quer­ent, but Wool­ner got out her cro­chet hook and sal­lied forth, fol­low­ing Chat­G­PTs instruc­tions to the let­ter, despite a num­ber of red flags indi­cat­ing that the chatbot’s grasp of nar­whal anato­my was high­ly unre­li­able.

Its igno­rance is part of its DNA. As a large lan­guage mod­el, Chat­G­PT is capa­ble of pro­duc­ing pre­dic­tive text based on vast amounts of data in its mem­o­ry bank. But it can’t see images.

As Amit Kat­wala writes in Wired:

It has no idea what a cat looks like or even what cro­chet is. It sim­ply con­nects words that fre­quent­ly appear togeth­er in its train­ing data. The result is super­fi­cial­ly plau­si­ble pas­sages of text that often fall apart when exposed to the scruti­ny of an expert—what’s been called “flu­ent bull­shit.”

It’s also not too hot at math, a skill set knit­ters and cro­cheters bring to bear read­ing pat­terns, which traf­fic in num­bers of rows and stitch­es, indi­cat­ed by abbre­vi­a­tions that real­ly flum­mox a chat­bot.

An exam­ple of begin­ner-lev­el instruc­tions from a free down­load­able pat­tern for a cute amigu­ru­mi shark:

DORSAL FIN (gray yarn)

Rnd 1: in a mr work 3 sc, 2 hdc, 1 sc (6)

Rnd 2: 3 sc, 1 hdc inc, 1 hdc, 1 sc (7)

Rnd 3: 3 sc, 2 hdc, 1 hdc inc, 1 sc (8)

Rnd 4: 3 sc, 1 hdc inc, 3 hdc, 1 sc inc (10)

Rnd 5: 3 sc, 1 hdc, 1 hdc inc, 3 hdc, 1 sc, 1 sc inc (12)

Rnd 6: 3 sc, 6 hdc, 3 sc (12)

Rnd 7: sc even (12); F/O and leave a long strand of yarn to sew the dor­sal fin between rnds # 18–23. Do not stuff the fin.

Pity poor Chat­G­PT, though, like Wool­ner, it tried.

Their col­lab­o­ra­tion became a cause célèbre when Wool­ner debuted the “AI gen­er­at­ed nar­whal cro­chet mon­stros­i­ty” on Tik­Tok, apt­ly com­par­ing the large tusk Chat­G­PT had her posi­tion atop its head to a chef’s toque.

Is that the best AI can do?

A recent This Amer­i­can Life episode details how Sebastien Bubeck, a machine learn­ing researcher at Microsoft, com­mand­ed anoth­er large lan­guage mod­el, GPT‑4, to cre­ate code that TikZ, a vec­tor graph­ics pro­duc­er, could use to “draw” a uni­corn.

This col­lab­o­ra­tive exper­i­ment was per­haps more empir­i­cal­ly suc­cess­ful than the Chat­G­PT amigu­ru­mi pat­terns Wool­ner duti­ful­ly ren­dered in yarn and fiber­fill. This Amer­i­can Life’s David Kesten­baum was suf­fi­cient­ly awed by the result­ing image to haz­ard a guess that “when peo­ple even­tu­al­ly write the his­to­ry of this crazy moment we are in, they may include this uni­corn.”

It’s not good, but it’s a fuck­ing uni­corn. The body is just an oval. It’s got four stu­pid rec­tan­gles for legs. But there are lit­tle squares for hooves. There’s a mane, an oval for the head. And on top of the head, a tiny yel­low tri­an­gle, the horn. This is insane to say, but I felt like I was see­ing inside its head. Like it had pieced togeth­er some idea of what a uni­corn looked like and this was it.

Let’s not poo poo the mer­its of Woolner’s ongo­ing explo­rations though. As one com­menter observed, it seems she’s “found a way to instan­ti­ate the weird messed up arti­facts of AI gen­er­at­ed images in the phys­i­cal uni­verse.”

To which Wool­ner respond­ed that she “will either be spared or be one of the first to per­ish when AI takes over gov­er­nance of us meat sacks.”

In the mean­time, she’s con­tin­u­ing to har­ness Chat­G­PT to birth more mon­strous amigu­ru­mi. Ger­ald the Narwhal’s has been joined by a cat, an otter, Nor­ma the Nor­mal Fish, XL the Newt, and Skein Green, a pel­i­can bear­ing get well wish­es for author and sci­ence vlog­ger Hank Green.

When retired math­e­mati­cian Daina Taim­i­na, author of Cro­chet­ing Adven­tures with Hyper­bol­ic Planes, told the Dai­ly Beast that Ger­ald would have resem­bled a nar­whal more close­ly had Wool­ner sup­plied Chat­G­PT with more specifics, Wool­ner agreed to give it anoth­er go.

Two weeks lat­er, the Dai­ly Beast pro­nounced this attempt, nick­named Ger­ard, “even less nar­whal-look­ing than the first. Its body was a mas­sive stuffed tri­an­gle, and its tusk looked like a gum­drop at one end.”

Wool­ner dubbed Ger­ard pos­si­bly the most frus­trat­ing AI-gen­er­at­ed amigu­ru­mi of her acquain­tance, owing to an onslaught of speci­fici­ty on ChatCPT’s part. It over­loaded her with instruc­tions for every indi­vid­ual stitch, some­times call­ing for more stitch­es in a row than exist­ed in the entire pat­tern, then dipped out with­out telling her how to com­plete the body and tail.

As sil­ly as it all may seem, Wool­ner believes her Chat­G­PT amigu­ru­mi col­labs are a healthy mod­el for artists using AI tech­nol­o­gy:

I think if there are ways for peo­ple in the arts to con­tin­ue to cre­ate, but also approach AI as a tool and as a poten­tial col­lab­o­ra­tor, that is real­ly inter­est­ing. Because then we can start to branch out into com­plete­ly dif­fer­ent, new art forms and cre­ative expressions—things that we couldn’t nec­es­sar­i­ly do before or didn’t have the spark or the idea to do can be explored. 

If you, like Hank Green, have fall­en for one of Woolner’s unholy cre­ations, down­load­able pat­terns are avail­able here for $2 a pop.

Those seek­ing alter­na­tives to fiber­fill are advised to stuff their amigu­ru­mi with “aban­doned hopes and dreams” or “all those free tee shirts you get from giv­ing blood and run­ning road races or what­ev­er you do for fun”.

Relat­ed Con­tent 

An Artist Cro­chets a Life-Size, Anatom­i­cal­ly-Cor­rect Skele­ton, Com­plete with Organs

A Bio­sta­tis­ti­cian Uses Cro­chet to Visu­al­ize the Fright­en­ing Infec­tion Rates of the Coro­n­avirus

Make an Adorable Cro­cheted Fred­die Mer­cury; Down­load a Free Cro­chet Pat­tern Online

– Ayun Hal­l­i­day is the Chief Pri­ma­tol­o­gist of the East Vil­lage Inky zine and author, most recent­ly, of Cre­ative, Not Famous: The Small Pota­to Man­i­festo and Cre­ative, Not Famous Activ­i­ty Book. Fol­low her @AyunHalliday.

DALL‑E, an arti­fi­cial intel­li­gence sys­tem that gen­er­ates viable-look­ing art in a vari­ety of styles in response to user sup­plied text prompts, has been gar­ner­ing a lot of inter­est since it debuted this spring.

It has yet to be released to the gen­er­al pub­lic, but while we’re wait­ing, you could have a go at DALL‑E Mini, an open source AI mod­el that gen­er­ates a grid of images inspired by any phrase you care to type into its search box.

Co-cre­ator Boris Day­ma explains how DALL‑E Mini learns by view­ing mil­lions of cap­tioned online images:

Some of the con­cepts are learnt (sic) from mem­o­ry as it may have seen sim­i­lar images. How­ev­er, it can also learn how to cre­ate unique images that don’t exist such as “the Eif­fel tow­er is land­ing on the moon” by com­bin­ing mul­ti­ple con­cepts togeth­er.

Sev­er­al mod­els are com­bined togeth­er to achieve these results:

• an image encoder that turns raw images into a sequence of num­bers with its asso­ci­at­ed decoder

• a mod­el that turns a text prompt into an encod­ed image

• a mod­el that judges the qual­i­ty of the images gen­er­at­ed for bet­ter fil­ter­ing 

My first attempt to gen­er­ate some art using DALL‑E mini failed to yield the hoped for weird­ness.  I blame the bland­ness of my search term — “toma­to soup.”

Per­haps I’d have bet­ter luck “Andy Warhol eat­ing a bowl of toma­to soup as a child in Pitts­burgh.”

Ah, there we go!

I was curi­ous to know how DALL‑E Mini would riff on its name­sake artist’s han­dle (an hon­or Dali shares with the tit­u­lar AI hero of Pixar’s 2018 ani­mat­ed fea­ture, WALL‑E.)

Hmm… seems like we’re back­slid­ing a bit.

Let me try “Andy Warhol eat­ing a bowl of toma­to soup as a child in Pitts­burgh with Sal­vador Dali.”

Ye gods! That’s the stuff of night­mares, but it also strikes me as pret­ty legit mod­ern art. Love the spar­ing use of red. Well done, DALL‑E mini.

At this point, van­i­ty got the bet­ter of me and I did the AI art-gen­er­at­ing equiv­a­lent of googling my own name, adding “in a tutu” because who among us hasn’t dreamed of being a bal­le­ri­na at some point?

Let that be a les­son to you, Pan­do­ra…

Hope­ful­ly we’re all plan­ning to use this play­ful open AI tool for good, not evil.

Hyperallergic’s Sarah Rose Sharp raised some valid con­cerns in rela­tion to the orig­i­nal, more sophis­ti­cat­ed DALL‑E:

It’s all fun and games when you’re gen­er­at­ing “robot play­ing chess” in the style of Matisse, but drop­ping machine-gen­er­at­ed imagery on a pub­lic that seems less capa­ble than ever of dis­tin­guish­ing fact from fic­tion feels like a dan­ger­ous trend.

Addi­tion­al­ly, DALL‑E’s neur­al net­work can yield sex­ist and racist images, a recur­ring issue with AI tech­nol­o­gy. For instance, a reporter at Vice found that prompts includ­ing search terms like “CEO” exclu­sive­ly gen­er­at­ed images of White men in busi­ness attire. The com­pa­ny acknowl­edges that DALL‑E “inher­its var­i­ous bias­es from its train­ing data, and its out­puts some­times rein­force soci­etal stereo­types.”

Co-cre­ator Day­ma does not duck the trou­bling impli­ca­tions and bias­es his baby could unleash:

While the capa­bil­i­ties of image gen­er­a­tion mod­els are impres­sive, they may also rein­force or exac­er­bate soci­etal bias­es. While the extent and nature of the bias­es of the DALL·E mini mod­el have yet to be ful­ly doc­u­ment­ed, giv­en the fact that the mod­el was trained on unfil­tered data from the Inter­net, it may gen­er­ate images that con­tain stereo­types against minor­i­ty groups. Work to ana­lyze the nature and extent of these lim­i­ta­tions is ongo­ing, and will be doc­u­ment­ed in more detail in the DALL·E mini mod­el card.

The New York­er car­toon­ists Ellis Rosen and Jason Adam Katzen­stein con­jure anoth­er way in which DALL‑E mini could break with the social con­tract:

And a Twit­ter user who goes by St. Rev. Dr. Rev blows minds and opens mul­ti­ple cans of worms, using pan­els from car­toon­ist Joshua Bark­man’s beloved web­com­ic, False Knees:

Pro­ceed with cau­tion, and play around with DALL‑E mini here.

Get on the wait­list for orig­i­nal fla­vor DALL‑E access here.

Relat­ed Con­tent

Arti­fi­cial Intel­li­gence Brings to Life Fig­ures from 7 Famous Paint­ings: The Mona Lisa, Birth of Venus & More

Google App Uses Machine Learn­ing to Dis­cov­er Your Pet’s Look Alike in 10,000 Clas­sic Works of Art

Arti­fi­cial Intel­li­gence for Every­one: An Intro­duc­to­ry Course from Andrew Ng, the Co-Founder of Cours­era

- Ayun Hal­l­i­day is the Chief Pri­ma­tol­o­gist of the East Vil­lage Inky zine and author, most recent­ly, of Cre­ative, Not Famous: The Small Pota­to Man­i­festo.  Fol­low her @AyunHalliday.