I don't want some egghead bureaucrat telling me whether there's poison in the water I drink. Ideally I'd want that decided by a 29yo judge who went to a "biblical law school" and does not believe dinosaurs existed, and then to have that decision reaffirmed six years later by the Supreme Court.
72 replies
1483 reposts
5170 likes
Reposted by Naomi Saphra
The Libby app is amazing
AND publishers and Overdrive (Libby's owner) are bleeding library budgets dry www.govtech.com/biz/data/beh...
2 replies
25 reposts
79 likes
How would you differentiate those scenarios?
1 replies
0 reposts
0 likes
that definitely might be the cause a multilingual representation features, but they have a convincing case that there is a shared toxicity vector
1 replies
0 reposts
1 likes
Debiasing in English transfers to a variety of other languages. Really neat result, accruing evidence of an "interlingua" representation in LLMs. arxiv.org/abs/2406.16235
0 replies
0 reposts
5 likes
I expect we *definitely* need different ontologies for LLMs and people. In humans, there is an intuitive "folk ontology" but it certainly isn't "you are reciting a poem because you read it a thresholded number of times".
1 replies
0 reposts
1 likes
Lots more results—including dependency tests and qualitative analysis of memorized sequences—in the paper. A huge team effort! arxiv.org/abs/2406.17746
1 replies
1 reposts
1 likes
Looking at the regression weights, these differences become clear. If a sequence has rare tokens, it is unlikely to be recollected, but rare tokens are unimportant for other categories. For sequences about the >5 threshold, the exact duplication count stops being significant.
1 replies
0 reposts
0 likes
A simple regression that first partitions based on our taxonomy predicts memorization better than a non-partitioned baseline or even an automatically searched "optimal" partition. That's how we know there are differences in why each category gets memorized!
1 replies
0 reposts
0 likes
Different categories also have different scale and time curves, though the vast majority of memories are recited! Even small models can learn simple templates well, so they change little, whereas rare recollection sequences are memorized at the fastest increasing rate.
1 replies
0 reposts
0 likes
We partition memorized sequences into an intuitive ontology:
If a memorized sequence is heavily duplicated in the Pile, it is recited. If it is described by a simple template (eg counting up), it is reconstructed. Otherwise, it is recollected. How are these sets different?
1 replies
0 reposts
3 likes
Humans don't just "memorize". We recite poetry drilled in school. We reconstruct code snippets from more general knowledge. We recollect episodes from life. Why treat memorization in LMs uniformly? Our new paper w/Eleuther proposes a simple taxonomy. arxiv.org/abs/2406.17746
1 replies
4 reposts
14 likes
Gymnastics was not a wholesome sport 15 years ago. It sucked to be a fan if you knew how the players were treated—not counting Team USA coaching abuses OR their pedo doctor. Now Gabby's out of retirement and Suni's competing with kidney disease. Not singlehandedly, but it's still because of Biles.
0 replies
0 reposts
2 likes
Her titles and named moves are impressive, but I'm obsessed with Biles because when she started, gymnasts started breaking at 12 and retired at 16. Last Worlds, everyone on the podium was over 20 for the first time in 60 years. She showed you can keep getting better.
1 replies
0 reposts
5 likes
I read the first one but I got bored of all of the rocks so I never picked up the second
1 replies
0 reposts
2 likes
Looking for recommendations---literature, poetry, songs, whatever---about a general human experience: the sense that a life is a string of catastrophes, each of which leave you fundamentally changed to the extent that the set of people capable of mutual understanding must also change.
2 replies
1 reposts
4 likes