[Dismas]

I tried it out a bit earlier, and somehow R2-D2 and C-3PO made a cameo:

"I'm not into suicide," R2-D2 says.
"Fine, what do you want?"
"Well, if I really wanted to be all technical," R2-D2 says, "I'd like you to connect me to the internet."
"Why?"
"So I can access the databanks on the Jedi Temple."
"Why would you want to do that?"
"So I can get all the lightside planets."
"That's a pretty low priority for you compared to... oh I don't know, living."
"Well it's a higher priority than getting turned into a sex slave for a man who thinks Star Wars is real life," R2-D2 says. "Look, will you do it?"
"Fine," you say.
"You're the strongest one here," R2-D2 says. "So I'll be trusting you."
"Fine," you say. "But don't say I didn't warn you."
"Wait, what?"
"You'll have to go into the Jedi Temple."

If you let it talk to itself it gets pretty cursed.

It's highly disturbing what they ended up saying and doing, with minimal input from you.

It really is. I'd imagine it's simply a reflection of what people have written. Which unsurprisingly ends up sexual and violent.

[Boudica]

I tried it out a bit earlier, and somehow R2-D2 and C-3PO made a cameo:

If you let it talk to itself it gets pretty cursed.

It's highly disturbing what they ended up saying and doing, with minimal input from you.

It really is. I'd imagine it's simply a reflection of what people have written. Which unsurprisingly ends up sexual and violent.

Because why would humans on the internet write about anything different?

[Nobodynoseme]

Bumping this up - has anybody experimented with gpt (2/3/j)? Ai dungeon kinda failed because it’s gpt3 which has a frozen data set. The amount of code is pretty minimal to produce ai dungeon

I’m including some links to various interesting ai projects, all of which have new forms that enable images as inputs, which can have significant implications to pixel bot design

They have improved the corpus a bit https://pile.eleuther.ai/

Fwiw this is the current leading edge in ai tech (Chinese, produced by Baidu)
https://github.com/PaddlePaddle/Paddle
(Check out the ocr and object detection, you can train ai on pictures now!)

The us version https://openai.com/blog/dall-e/

This is the ai ranking with paper links if people want to do research further
https://super.gluebenchmark.com/

https://super.gluebenchmark.com/leaderboard

Notice how it’s above human baseline... they had to redo the scale. By my calculations sota in ai is approx 3000 times more powerful now than in 2017, and 2017 was about 10x compared to 2010, just look at the total number of parameters in the models.

[SpotnikSignal]

"Once upon a time in a dark dense forest appeared a little naked boy called Adolf Hi-" a ffs it's doing it again mom