I still find myself confused that you don’t express thinking that the more complicated architectures provide a plausible guide for what to expect these things to look like in the brain. I have come to the conclusion that all the parts of intelligence—thinking and blindsight among them—have been tried by ML people at this point, but none of them at large enough scale or integrated well enough to produce something like us. It continues to seem strange to me that your objection is “they haven’t actually tried the right thing”, and that you are also optimistic about attention/memory/etc as being the right sort of thing to produce it. Do you think that thinking doesn’t have an obvious construction from available parts? What are the cruxes, the diff of our beliefs here?
I really wish you would give his argument for the claim that we (even plausibly) have all the pieces, Lahwran. I would also love to see an abridged transcript of a discourse wherein the two of you reached a double-crux. My best guess is that Lahwran is thinking of ‘only integrating existing systems’ as a triviality which can be automated by the market rather than what it actually is, a higher-level instance of the design problem.
That said, the idea that thinking has been tried seems so insane to me that I may be failing to steelman it accurately.
I was under the impression that things like “deliberative thinking” and “awareness” haven’t been simulated by ML thus far, so I think that’s the diff between us—though it’s not that strongly held, there are lots of ML advances I may just not have heard of.
At first I was very surprised that they got such good performance at answering questions about visual scenes (e.g. “what shape is the red thing?” “the red thing is a cube.”)
Then I noticed that they gave ground-truth examples not just for the answers to the questions but to the programs used to compute those answers. This does not sound like the machine “learned to reason” so much as it “learned to do pattern-recognition on examples of reasoning.” When humans learn, they are “trained” on examples of other people’s behavior and words, but they don’t get any access to the raw procedures being executed in other people’s brains. This AI did get “raw downloads of thinking processes,” which I’d consider “cheating” compared to what humans do. (It doesn’t make it any less of an achievement by the paper authors, of course; you have to do easier things before you can do harder things.)
That seems like weaseling out of the evidence to me. This is just another instance of neural networks being able to learn to do geometric computation to produce hard-edged answers, like alphago is; that they’re being used to generate programs seems not super relevant to that. I certainly agree that it’s not obvious exactly how to get them to learn the space of programs efficiently, but it seems surprising to expect it to be different in kind vs previous neural network stuff. This doesn’t seem that different to me vs attention models in terms of what kind of problem learning the internal behavior presents.
I still find myself confused that you don’t express thinking that the more complicated architectures provide a plausible guide for what to expect these things to look like in the brain. I have come to the conclusion that all the parts of intelligence—thinking and blindsight among them—have been tried by ML people at this point, but none of them at large enough scale or integrated well enough to produce something like us. It continues to seem strange to me that your objection is “they haven’t actually tried the right thing”, and that you are also optimistic about attention/memory/etc as being the right sort of thing to produce it. Do you think that thinking doesn’t have an obvious construction from available parts? What are the cruxes, the diff of our beliefs here?
I really wish you would give his argument for the claim that we (even plausibly) have all the pieces, Lahwran. I would also love to see an abridged transcript of a discourse wherein the two of you reached a double-crux. My best guess is that Lahwran is thinking of ‘only integrating existing systems’ as a triviality which can be automated by the market rather than what it actually is, a higher-level instance of the design problem.
That said, the idea that thinking has been tried seems so insane to me that I may be failing to steelman it accurately.
I was under the impression that things like “deliberative thinking” and “awareness” haven’t been simulated by ML thus far, so I think that’s the diff between us—though it’s not that strongly held, there are lots of ML advances I may just not have heard of.
An example of what I would mean by thinking: https://arxiv.org/pdf/1705.03633.pdf
Thanks for the paper!
At first I was very surprised that they got such good performance at answering questions about visual scenes (e.g. “what shape is the red thing?” “the red thing is a cube.”)
Then I noticed that they gave ground-truth examples not just for the answers to the questions but to the programs used to compute those answers. This does not sound like the machine “learned to reason” so much as it “learned to do pattern-recognition on examples of reasoning.” When humans learn, they are “trained” on examples of other people’s behavior and words, but they don’t get any access to the raw procedures being executed in other people’s brains. This AI did get “raw downloads of thinking processes,” which I’d consider “cheating” compared to what humans do. (It doesn’t make it any less of an achievement by the paper authors, of course; you have to do easier things before you can do harder things.)
That seems like weaseling out of the evidence to me. This is just another instance of neural networks being able to learn to do geometric computation to produce hard-edged answers, like alphago is; that they’re being used to generate programs seems not super relevant to that. I certainly agree that it’s not obvious exactly how to get them to learn the space of programs efficiently, but it seems surprising to expect it to be different in kind vs previous neural network stuff. This doesn’t seem that different to me vs attention models in terms of what kind of problem learning the internal behavior presents.