6 ms·

It seems like Scott's bet was merely that our modern techniques would be able to make at least some nonzero progress in compositionality (and the terms of his b

It seems like Scott's bet was merely that our modern techniques would be able to make at least some nonzero progress in compositionality (and the terms of his bet spelled this out with how lenient it was), and Gary is treating it as if the bet was about compositionality being solved. It feels like a very bad faith reading from Gary.

t_mann 4y ago
His point that the test as described - with multiple statistical issues piled on top of each other - does not allow much of a meaningful inference in any direction is completely valid and independent of what hypotheses were being tested.
- ravi-delia 4y ago
  Sure, but the terms of the bet were known ahead of time. Like, Alexander never claimed composition was solved, just that he won the bet. Which he did.
  - t_mann 4y ago
    Even more so it is appropriate to point out that this victory is strictly limited to the specific terms of this particular bet (and strictly speaking not even that, since the terms were changed after the bet was placed), and do not provide statistically sound evidence of progress on compositionality. PS: in the end, Alexander claims that his experiment "provide(s) some evidence that simple scaling and normal progress are enough for compositionality gains". So he does in fact go significantly beyond just claiming victory on this particular bet.
plutonorm 4y ago
Gary Marcus is so deep into the "connectionism doesnt work" rabbit hole that he'd deny his own sentience if it turned out he was made of silicon. I just ignore him as he only appears to be getting more and more incorrect.
- mark_l_watson 4y ago
  Sure, he can sound strident but I still think Gary Marcus's riffing on the limitations of deep learning is important. The book "Rebooting AI" that he wrote with Ernest Davis is well worth reading if you are an AI practitioner (a term I use to describe myself). I think Marcus is also well worth following on Twitter to get a contrarian view (he re-tweeted me two weeks ago, so there is some overlap in our points of view). Way back when, I liked Roger Penrose's 1989 book "The Emperor's New Mind" even though some of the people I worked with thought he was a devil for writing that. I am much more optimistic than Marcus, but find his work useful and thoughtful.

t_mann 4y ago

His point that the test as described - with multiple statistical issues piled on top of each other - does not allow much of a meaningful inference in any direction is completely valid and independent of what hypotheses were being tested.

ravi-delia 4y ago
Sure, but the terms of the bet were known ahead of time. Like, Alexander never claimed composition was solved, just that he won the bet. Which he did.
- t_mann 4y ago
  Even more so it is appropriate to point out that this victory is strictly limited to the specific terms of this particular bet (and strictly speaking not even that, since the terms were changed after the bet was placed), and do not provide statistically sound evidence of progress on compositionality. PS: in the end, Alexander claims that his experiment "provide(s) some evidence that simple scaling and normal progress are enough for compositionality gains". So he does in fact go significantly beyond just claiming victory on this particular bet.

plutonorm 4y ago

Gary Marcus is so deep into the "connectionism doesnt work" rabbit hole that he'd deny his own sentience if it turned out he was made of silicon. I just ignore him as he only appears to be getting more and more incorrect.

mark_l_watson 4y ago
Sure, he can sound strident but I still think Gary Marcus's riffing on the limitations of deep learning is important. The book "Rebooting AI" that he wrote with Ernest Davis is well worth reading if you are an AI practitioner (a term I use to describe myself). I think Marcus is also well worth following on Twitter to get a contrarian view (he re-tweeted me two weeks ago, so there is some overlap in our points of view). Way back when, I liked Roger Penrose's 1989 book "The Emperor's New Mind" even though some of the people I worked with thought he was a devil for writing that. I am much more optimistic than Marcus, but find his work useful and thoughtful.