|
@OpenAI | |||||
|
We've fine-tuned GPT-2 using human feedback for tasks such as summarizing articles, matching the preferences of human labelers (if not always our own). We're hoping this brings safety methods closer to machines learning values by talking with humans. openai.com/blog/fine-tuni…
|
||||||
|
||||||
|
𝔊𝔴𝔢𝔯𝔫
@gwern
|
19. ruj |
|
𝘠𝘦𝘴𝘴𝘴𝘴𝘴𝘴.
Excited to see if the code is usable for improving poetry generation.
|
||
|
|
||
|
Andre Infante
@AndreTI
|
19. ruj |
|
Surely if you can do this, you can use another transformer as an adversarial module to clean up some of the artifacts (repetition) right? No reason these supervision signals have to come from a human.
|
||
|
|
||
|
Dean P
@DeanPlbn
|
19. ruj |
|
Such a cool update to GPT-2. What I like about this is the qualitative examples showing improvement. So much better than “we’ve beat the SOTA by 0.01%”
|
||
|
|
||
|
rdwrt
@rdwrt
|
20. ruj |
|
Just adding "safety" to this tweet and hope that's enough to fool everyone into assuming we're up to something good.
|
||
|
|
||
|
Louis Maddox
@permutans
|
20. ruj |
|
Well it's in their paper too, they have been outspoken on the dangers of large language models able to imitate humans. 'Safety' is also RE: biases in the text gen. models following incompletely specified objective functions, here they phrase it in terms of rules, like "don't lie"
|
||
|
|
||
|
Cem Say
@say_cem
|
19. ruj |
|
Will you be sharing any examples of this phenomenon? pic.twitter.com/EGVQRTFAPe
|
||
|
|
||
|
Tony Abram
@antonabramov
|
19. ruj |
|
Is there a fast way to try these advances online? Summarization seems to be pretty easy to test on a website
|
||
|
|
||
|
Gus the Winged Wild Dog
@Gushousekai195
|
19. ruj |
|
I can't wait for the day it can write me stories about whatever I ask it to.
|
||
|
|
||
|
sahil
@Sahil_Verma_97
|
20. ruj |
|
😎😉
|
||
|
|
||
|
Josef Vojtek
@AdAstraMadCat
|
19. ruj |
|
Who wrote this tweet? :)
|
||
|
|
||