Forwarded from Axis of Ordinary
"Is AI task performance a type of submartingale, like a stock market index that goes up over time, but where each particular movement is intrinsically unpredictable?" https://www.lesswrong.com/posts/G993PFTwqqdQv4eTg/is-ai-progress-impossible-to-predict
Forwarded from Axis of Ordinary
"After showing a few examples, large language models can translate natural language mathematical statements into formal specifications."
[...]
"It handles the negation “there is no function” by proof-by-contradiction. It understands the phrase “into itself” and correctly formalizes the co-domain of f."
[...]
"Our finding hence shows a very surprising capability of these models. They learned very general and transferable knowledge that allows them to work with low-resource formal language."
Thread: https://threadreaderapp.com/thread/1529886847953870848.html
[...]
"It handles the negation “there is no function” by proof-by-contradiction. It understands the phrase “into itself” and correctly formalizes the co-domain of f."
[...]
"Our finding hence shows a very surprising capability of these models. They learned very general and transferable knowledge that allows them to work with low-resource formal language."
Thread: https://threadreaderapp.com/thread/1529886847953870848.html
Forwarded from Nox Populi
I think this counterpoint still stands
On the Paradox of Learning to Reason from Data
https://twitter.com/HonghuaZhang2/status/1528963938825580544
https://arxiv.org/abs/2205.11502
On the Paradox of Learning to Reason from Data
https://twitter.com/HonghuaZhang2/status/1528963938825580544
https://arxiv.org/abs/2205.11502
Twitter
Honghua Zhang
Can language models learn to reason by end-to-end training? We show that near-perfect test accuracy is deceiving: instead, they tend to learn statistical features inherent to reasoning problems. See more in arxiv.org/abs/2205.11502 @LiLiunian @TaoMeng10 @kaiwei_chang…
Forwarded from Just links
Inverse Occam’s razor https://www.nature.com/articles/s41567-022-01575-2
https://arxiv.org/abs/2204.08284
https://arxiv.org/abs/2204.08284
Nature
Inverse Occam’s razor
Nature Physics - Scientists have long preferred the simplest possible explanation of their data. More recently, a worrying trend to favour unnecessarily complex interpretations has taken hold.
Forwarded from Axis of Ordinary
"Google just released a 442-author paper about a monster new test suite for evaluating Large Language Models (GPT-3 and the like), and in particular, their study of the language models' performance on their test suite as the number of parameters is scaled. As a striking example, see below for a neural net's ability to guess a movie from emojis (rather than, say, outputting random nonsense) as the number of parameters is gradually scaled from 2 million all the way to 128 billion." (via Scott Aaronson )
https://github.com/google/BIG-bench/blob/main/docs/paper/BIG-bench.pdf
This should be a little bit worrying because it makes it difficult to predict future progress. They are not just getting predictably better with more training, data, and parameters but sometimes capabilities emerge in a jumpy and unpredictable way. There can be sudden phase transitions.
https://github.com/google/BIG-bench/blob/main/docs/paper/BIG-bench.pdf
This should be a little bit worrying because it makes it difficult to predict future progress. They are not just getting predictably better with more training, data, and parameters but sometimes capabilities emerge in a jumpy and unpredictable way. There can be sudden phase transitions.