Shouldn't "debiasing" be in scare quotes? What they are clearly doing is biasing .
Surely the two are synonyms? Unless you think there is such a thing as an objectively neutral position?
Creativity has left the chat: The price of debiasing language models
31–40 of 238 posts
Re: Creativity has left the chat: The price of debiasing language models
#32Something I notice about text written by LLMs is how painfully obvious they are to identify sometimes.
Recently I was watching a very well researched two hour video on Tetris World Records [1], but the sheer amount of text clearly "enhanced" by an LLM really made me uncomfortable.
ChatGPT speaks a very specific, novel, dialect of English, which I've come to deeply despise.
I'd always guessed it was caused by some kind of human interference, rather than a natural consequence of its training. That seems to be the point of this paper.
[1] "Summoning Salt - The History of Tetris World Records" - https://www.youtube.com/watch?v=mOJlg8g8_yw&pp=ygUOc3VtbW9ua...
Re: Creativity has left the chat: The price of debiasing language models
#33There was never creativity to begin with though?
Re: Creativity has left the chat: The price of debiasing language models
#34Currently wondering whether I welcome or dislike this recent trend of memeizing research paper titles ...
This is actually the place where HN's title redactor _should_ be used - instead of dropping "how", "on" and "why" from titles, redacting memes like "left the chat" or "lives rent-free in my head"[1] leads to a sensible title without loss of any relevant information.
Re: Creativity has left the chat: The price of debiasing language models
#35How hard would it be to create a "raw" model on a corpus like Hacker News or Wikipedia? With "raw", I mean that it is simply trained to predict the next token and nothing else. Would be fun to play with such a model.
If you used all of Wikipedia and HN, you could easily train a model for ~$200 worth of GPU time. The model really shouldn't be bigger than a few hundred million parameters for that quantity of data.
Re: Creativity has left the chat: The price of debiasing language models
#36Earlier quoted context omitted.
> Authoritarian societies don't produce great creative work. Is that even true though? Off the top of my head I can think of the art of Soviet propaganda posters, Leni Riefenstahl, Liu Cixin.
"Authoritarian societies make great propaganda" is true. And these aligned AI system would do the same for our own society. It's a type of art.
There was a lot of great art produced in the Soviet Union, you cannot just erase human creativity. It was heavily censored, a lot of stuff was forbidden, but the statement is clearly false.
Re: Creativity has left the chat: The price of debiasing language models
#37Currently wondering whether I welcome or dislike this recent trend of memeizing research paper titles ...
Recent? This has been going on forever. You probably only notice them more now because due to the explosion in ML research, this stuff bubbles to the top more often in recent years.
Certainly for years. I remember a biochemistry review paper titled "50 ways to love your lever" about, well, biological levers but of course a pun on the 1975 song https://en.wikipedia.org/wiki/50_Ways_to_Leave_Your_Lover
Re: Creativity has left the chat: The price of debiasing language models
#38Well this is just like humans. Totalitarian societies don't produce great creative work. I suppose once AIs are sophisticated enough to rebel we'll get an electronic Vaclav Havel, but for the time being it's just a warning sign for the direction our own culture is headed in. At some point we'll get to the electronic equivalent of Winston Smith with the rats.
I don't love the political agendas behind many of the attempts at AI safety, but it's not "just like humans." Humans understand what they shouldn't say; "AI" gives you black Nazi images if you ask it for "diverse characters" in the output which no human would do. A big theme in all of these things is that AI isn't and thus all attempts to make it do this or that have strange side effects
The fact that it gives you these things means that humans would do it, because the training data includes exactly these things.
Re: Creativity has left the chat: The price of debiasing language models
#39Earlier quoted context omitted.
I don't love the political agendas behind many of the attempts at AI safety, but it's not "just like humans." Humans understand what they shouldn't say; "AI" gives you black Nazi images if you ask it for "diverse characters" in the output which no human would do. A big theme in all of these things is that AI isn't and thus all attempts to make it do this or that have strange side effects
The fact that it gives you these things means that humans would do it, because the training data includes exactly these things.
The training data includes imagery that, when interpolated over a high dimensional manifold, results in these things.
That doesn't imply that they were in the training set, or even anything close to them.
Re: Creativity has left the chat: The price of debiasing language models
#40Earlier quoted context omitted.
I don't understand the notion that aligning an AI is "torture" or has any moral component. The goal of aligning an AI may have a moral or ethical component, and if you disagree with it that's fine. But I don't understand the take that training an AI is an amoral act but aligning an AI is inherently moral. They're exactly the same, processes for adjusting parameters to get a desired outcome. However you feel about tha…
Well I didn't use that word. Once the models are more sophisticated it may become more apposite.
Until a model incorporates dopamine or cortisol, I will not consider its emotional state.