AI cake fails are unlike any cake fails I've seen before. What do you get when
your cake is being generated by something that's seen lots of pictures labeled
"cake" but has never had a cake or experienced physics?
CLIP+VQGAN are two algorithms
You know those games that go around on platforms like Twitter, where people
share pictures of themselves looking like the leader of an underground
resistance, or search google images for their name + fantasy armor or somesuch?
I assume that if you successfully start one of these games, fame and fortune
People have noted [https://twitter.com/moultano/status/1418256870259580934?s=20]
that when using giant internet-trained AIs like CLIP+VQGAN to generate images,
you get much nicer-looking images if you include an artist byline.
Here's "Internet Infrastructure"
And here's "Internet Infrastructure by J*
Like other ambiguous [https://www.aiweirdness.com/the-art-of-asking-nicely/]
image-generation
[https://www.aiweirdness.com/why-is-generated-furniture-so-cursed/] prompts
[https://www.aiweirdness.com/internet-grab-bag/], asking for "a gothic wardrobe"
had the AI hedging its bets.
In CLIP+VQGAN's internet training, those words might go with pictures of clothes
in styles
Beetle kill pine is a popular wood in Colorado, salvaged from trees killed by
pine bark beetles. While the dead trees stand before harvesting, a fungus
colonizes them, giving the wood interesting blue-grey streaks called spalting.
It's a really pretty wood.
But AI apparently thinks "a chair
There are upsides to working with a neural net that trained on a huge collection
of internet images and text. One is that, instead of ominous grey geometric
blobs
[https://aiweirdness.com/post/177091486527/this-ai-is-bad-at-drawing-but-will-try-anyways]
when it doesn't understand your prompt (there is a free interactive demo of
One of my favorite ways to mess around with text-generating AI is to see what it
does with new games that didn't exist on the internet back when its training
data was collected.
OpenAI's GPT-3 models were trained with data that cut off in late 2019
The US National Oceanic and Atmospheric Administration releases its planned
hurricane names years in advance, and includes a pronunciation guide
[https://www.nhc.noaa.gov/pdf/aboutnames_pronounce_atlc.pdf] for each name. An
excerpt from NOAA's 2026 Atlantic Basin Storm Name Pronunciations:
> Arthur AR-thur
Bertha BUR-thuh
The neural net GPT-3 has seen a lot of dating profiles during its internet training. But has it seen the Bachelorette contestant variety of astonishingly pedestrian dating profile?
One way to generate buzz for your new food truck: a unique theme or gimmick. But
with new food trucks popping up everywhere, how can you ensure that your idea is
creative enough?
Lindsay Diamond [https://neurdy.com/] suggested a possible neural net solution.
So as a demonstration, I wrote