![]() ![]() ![]() To get the AI to do this, I started with a big neural net pre-trained on internet text. Then I did my best to create each card following its instructions. So, what I ended up doing was using a text-generating neural net to write both the message and a description of the accompanying picture. ![]() But asking for even something as simple as "a candy heart with a message" resulted in entirely illegible and extremely cursed items. CLIP+DALL-E and similar CLIP-based algorithms can generate a picture of a sign that's at least a bit legible if the message is only a word or two long, and you spell out for it exactly what it should say. Right now we don't quite have an AI capable of this. When I started this project, it was important to me that the AI specify not just the message, but also the images. I remember having to give out Valentine cards in elementary school, a mandatory expression of not-actual-valentine-feeling in which you could express your individuality and warm regard for your best friends and the classroom bully by giving them small store-bought cards with cartoon characters and generic messages on them.īut what if the messages could be not generic but weirdly unsettling? In other words, what if the cards were generated by a glitchy AI? Model: GPT-3 DaVinci ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |