I have been dying to see how GPT-2 writes a Harry Potter script, and the result is glorious beyond my imaginings.pic.twitter.com/t4960bS28H
Можеш да добавиш към туитовете си информация за местоположението (например града ти или точното ти местоположение) през мрежата и чрез приложения на трети страни. Винаги можеш да изтриеш историята на местоположенията на туитовете си. Научи повече
I have been dying to see how GPT-2 writes a Harry Potter script, and the result is glorious beyond my imaginings.pic.twitter.com/t4960bS28H
Here's one of the outtakes: instead of writing a Harry Potter script, GPT-2 did a bot-written script about scriptwriters getting bots to write scripts.pic.twitter.com/BQsK741dA0
Huh, surprised it keeps the same narrative transcriptive style all the way through.
I was super impressed! The toy 117M model usually wanders away from the format & characters and descends into prose or video game playthrough transcripts.pic.twitter.com/siOf6jt48w
But eels.
And sudden bursts of fisticuffs
Was this actually retrained on the scripts or was it just prompted? I find even the 117M GPT-2 is good at writing fanfic for popular stuff like Harry Potter, Star Wars. But not good at outputting a script-like format, works better with normal fanfic style.
This is the full model, prompted with the "Open AI asked a bot" + 1st 2 lines of dialog.
Here are some Harry Potter and Star Wars script snippets I got with the 117M version.https://twitter.com/JanelleCShane/status/1096414471642324993 …
I'm impressed by the 117M version with popular fandoms. It will introduce additional characters that fit, etc. Though it tends to write Spock and Kird into any Star Trek story in my experience!
Wait this one is real? Not one of those "we made a computer watch 30,000 hours of Chipotle commercials and then had it write one" things? If this was almost anyone other than you, I'd be skeptical. But I trust that you know what you're talking about with this sort of thing.
This one is the real thing! From the full GPT-2 model. It's interesting to compare with the made-up version.
It’s not a party until Snaps shows up in his nightgown with just the tip of his wand showing ... 
I definitely had to read that part twice
This is going to make it more difficult to explain the fake bot scripts
I know! The memory thing is no longer a simple giveaway. At least the “what is in the training data” question still applies. Video in still doesn’t lead to a model that can write scripts with stage directions.
"Okay, listen" through "meant to imply" is shockingly coherent as a dialog, I could see it in a fic. Are the first two lines in each image prompted, or only the first two in the first image?
Only the first 2 lines in the 1st image were prompted. The rest is presented unchanged (I did indent the stage directions so they would be easier to read)
That's incredibly cool. Thanks for the clarification :)
Twitter може да е претоварен или да изпитва моментно затруднение. Опитай отново или виж Twitter Status за повече информация.