Janelle Shane @janellecshane@wandering.shop
I train neural networks to write humor at aiweirdness.com | Also a research scientist in optics. She/her
Joined Nov 2017
Are you not terrified and repulsed??
I prompted GPT-3 with some human proposals for marking a nuclear waste site, in a way that will still be forbidding millennia from now.
https://aiweirdness.com/post/629516384194461697/a-10000-year-warning
Janelle Shane
boosted
"I’m not sure how to feel about the fact that telling a neural net to pretend it’s a smaller, more-confused neural net is now easier than actually training the smaller neural net."
wow @janellecshane
welcome to the future I guess
Rather than collect 1000s of supervillain bios, train a neural net on them, and style-transfer them all to be about AI, I just had GPT-3 generate the resulting blog post.
I collected my favorite AI supervillains and illustrated them.
https://aiweirdness.com/post/628791398606471168/ai-supervillains
There are several ways of getting GPT-3 to generate drawing prompts.
For #Botober I chose to give it a short story to complete, in which it had to generate lines for an erratic station AI. The neural net took over after "1."
The neural net suggested #Botober drawing prompts based on how the words fit, and not on any concept of how the objects might look. This list makes that abundantly clear.
I'm hand-curating the GPT-3 generated #Botober prompts because many of the neural net's drawing prompts are terrible for one reason or another.
Here are some of the prompts that were unfairly difficult.
For the first #Botober collection I generated, I had GPT-3 generate "things". Here's an alternate collection in which I had the neural net generate "concepts". It's different, but I'm not sure if it's easier.
Introducing #Botober, a set of AI-generated drawing prompts for each day in October!
I generated these using GPT-3, a neural net trained on a huge amount of internet text, but not on drawing prompts specifically. (This explains a lot)
Janelle Shane
boosted
Certainly, the image of her ripping open a spaceship hull with arm strength alone would not have been good for marketing the EZ-Clean bot system. Rawr, she thought. And also, beep.
(by @janellecshane)
A single typo in an Open Street Map entry caused an ordinary 2-story building in Melbourne to become an eerie 212-story pillar.
The Melbourne Monolith is now the most famous landmark in Microsoft Flight Simulator 2020. A successful completion of the Monolith Challenge is to land neatly on its roof.
[images: fulltimespy]
https://aiweirdness.com/post/627344542828544000/planet-earth-from-above
In today's disturbing news, I discovered that GPT-3 can write AI Weirdness blog posts.
1st paragraph is my prompt. The rest is GPT-3 simulating the fumbling of a much less powerful neural net.
https://aiweirdness.com/post/626712039215202304/ai-ai-weirdness
birdsite user @ConcernedCraig wrote this and I'm so pleased:
my nams janelle
and when a i
at riting pomes
can onli try
and its atempts
ar blest or curst
i sneek rite in
i post the wurst
what inception level is gpt-3 describing a meme from 2019 in a pseudo-chaucerian poetry form from late 2016?
getting gpt-3 to generate poetry on command seemed like a good idea at the time
https://aiweirdness.com/post/626075920939220992/my-name-is-bot
Which ant-person are you? https://www.tryinteract.com/share/quiz/5f1df7f2db538e00144fd1a3
I used GPT-3 to generate personality quizzes, and it turned out better than I expected.
Which artificial intelligence are you?
https://www.tryinteract.com/share/quiz/5f1e41e09f0ba50014bc45c6
I train neural networks to write humor at aiweirdness.com | Also a research scientist in optics. She/her
Joined Nov 2017
