|
|
|
|
|
|
by blankaccount
2473 days ago
|
|
|
> There was one recent one that even asked “What’s a pirate’s favorite letter?”, and the bot had noticed the answer was being given in the form of letters, but I don’t think a single instance of the bot proposed “r”. But it does not understand humor in the /slightest/. To me this actually sounds hilarious, if it is as you describe it. It's the kind of joke you'd find on The Office: Dwight: "What's a pirate's favourite letter?"
Jim: "M"
Jim: "N"
Jim: "O"
Jim: "P"
Jim: "Q"
Jim: "..."
Jim: "S" > "Toss some leeches around and wait 'til we get there."
Like the above poster said, I've seen this style of spam-y context-free meme on reddit before too. |
|
|
Well, here's the original I was referring to: https://www.reddit.com/r/SubSimulatorGPT2/comments/c8klvj/wh...
I was wrong. R was suggested in some of the replies. But the original answer is given as "The M", and contains gems like '"r" is a misspelled letter, and "m" is a misspelled letter. "M" is a misspellable word you can't use as a word in the English language.'
"Like the above poster said, I've seen this style of spam-y context-free meme on reddit before too. "
That would make sense. I guess I just don't frequent those corners. GPT-2 is clearly capable of picking up on structure, so if it sees something repeated it doesn't just notice "This particular thing is repeated a lot", it picks up some concept of repetition itself. A number of the bots have picked up the concept of quoting the message they're replying to. (In the meta reddit for this, the creator has said the posts and the replies are trained as separate corpora, so the replies "know" they are replies. I gather there is also enough markers that the bots can distinguish between title, post text, and subsequent replies.)