Science & Tech

Neural Net Generates Most Ridiculous Candy Hearts Ever

by Matthew Hart
Feb 18 2020 • 7:00 AM

Writer’s note: I’ve included here all of the candy heart messages generated by GPT-2 in their capitalized form for the sake of authenticity. Sorry FOR ALL OF THE LOUD READING you’ll be doing in your head. 

What do you get when you—metaphorically—feed lists of existing candy heart messages to an advanced neural network that’s frighteningly good at predicting language patterns? “BUNS BUNS BUNS,” apparently. Also, “MOUTHY HAMPSTER.” And, of course, “CANDY HOLE.” And those are just a few of the G-rated(?) messages the neural network spat out after being given real candy heart messages as references; the NSFW onesOpens in a new tab, we promise, are even better.

In a Valentine’s Day blog postOpens in a new tab from Janelle Shane (which comes via FuturismOpens in a new tab), the research scientist, speaker, writer, and “not a robot,” talked about how she used GPT-2Opens in a new tab, an unsupervised language generator developed by OpenAI, to create these completely insane, hilarious, and maybe even somewhat romantic, candy heart messages. And all we can say in response is let’s bust out the “KITTEN BAND” and get the “BODY PARTY!” started, ’cause Shane has done some fine work here.

Shane notes in her post that GPT-2 “had already learned a lot about [how] to write English from scanning millions of web pages”—8 million pages, to be more exactOpens in a new tab, or roughly the equivalent of all of Shakespeare’s works combined multiplied by 8,000—before she entered her list of candy heart messages; a different approach relative to how she generated the little pithy love blurbs back in 2018Opens in a new tab, when she utilized a neural net that had zero knowledge of English, and was only able to generate messages based on the 336 candy heart messages Shane gave it as training data.


“Keep in mind that [GPT-2] was not a neural net specially trained on candy hearts,” Shane says in her post, adding that “What I was doing was kind of like walking up to someone and shouting ‘HOT STUFF! COOL DUDE! LOVE BIRD!Opens in a new tab CALL ME! WHIZ KID! AS IF!'” and expecting GPT-2 to provide sensible candy heart messages in return.

And while the messages are certainly not, shall we say, conventional, they are endlessly entertaining. And isn’t that what candy hearts are all about, after all? It’s not like people eat them for the taste or actually use them to deliver romantic messages. (Although to be fair, if Valentine’s Day ain’t going so swell, pounding a box of these saccharine, symbolic blood pumpsOpens in a new tab can and will most definitely happen for some of us.)

A video describing how GTP-2 works. 

For those looking for even more entertainment, may we direct you to Talk to TransformerOpens in a new tab, the tool Shane used to interface with GPT-2. On the Talk to Transformer page, you can plug in any kind of prompt you’d like and have GPT-2 spit out line after line of, very loosely, related text. Or you can just check out the other awesome experiments Shane has done with neural nets, like generating lists of very convincing-sounding PokémonOpens in a new tab and even Dungeons & Dragons monstersOpens in a new tab.

What do you think about this list of candy heart messages generated by a neural net? Give us a “MILD FEEL” of your thoughts in the comments!

Feature image: Janelle Shane / AI WeirdnessOpens in a new tab