5 ms·

Also worth nothing that Ryan mentions > In addition to iterating on the training set, I also did a small amount of iteration on a 100 problem subset of the pub

Also worth nothing that Ryan mentions > In addition to iterating on the training set, I also did a small amount of iteration on a 100 problem subset of the public test set and > it's unfortunate that these sets aren’t IID: it makes iteration harder and more confusing It’s not unfortunate: generalizing beyond the training distribution is a crucial part of intelligence that ARC is trying to measure! Among other reasons, developing with test-set data is a bad practice in ML because it hides the difficulty this challenge. Even worse, writing about a bunch of tricks that help results on this subset is extending the test-set leakage the blog post's readers. This is why I'm glad the ARC Prize has a truly hidden test set

rfoo 2y ago
... and we know that if we really want to nail it we'd better just pay someone else to create 1,000,000 more harder problems for training (without looking at any in test set, of course). i.e. make the training set distribution similar to test set again. Because the thing we have now is data-hungry. Your brain is pre-trained on other similar challenges as well. What's the point of requiring it to "generalize beyond the training distribution" with so few samples? Really, I thought LLMs ended this "can we pretrain on in-house prepared private data for ILSVRC" flame war already.
- YeGoblynQueenne 2y ago
  The problem with that it is we know approaches that can generalise very well from very few examples, even one example, without any kind of pretraining, That requires a good background theory of the target domain (a "world model" in more modern parlance), and we don't know how to automatically generate that kind of theory; only human minds can do it, for now. But given such a theory the number of examples needed can be as few as 1. Clearly, if you can learn from one example, but find yourself using thousands, you've taken a wrong turn somewhere. The concern with the data-hungry approach to machine learning, that at least some of us have, is that it has given up on the effort to figure out how to learn good background theories and turned instead to getting the best performance possible in the dumbest possible way, relying on the largest available amount of examples and compute. That's a trend against everything else in computer science (and even animal intelligence) where the effort is to make everything smaller, cheaper, faster, smarter: it's putting all the eggs in the basket of making it big, slow and dumb, and hoping that this will somehow solve... intelligence. A very obvious contradiction. Suppose we lived in a world that didn't have a theory of computational complexity and didn't know that some programs are more expensive to run than others. Would it be the case in that world, that computer scientists competed in solving ever larger instances of the Traveling Salesperson Problem, using ever larger computers, without even trying to find good heuristics exploiting the structure of the problem and simply trying to out-brute-force each other? That world would look a lot like where we are now with statistical machine learning: a pell-mell approach to throwing all resources at a problem that we just don't know how to solve, and don't even know if we can solve.
  - advael 2y ago
    The formalism that data-driven machine learning leans on is empirical tuning of stochastic search to drive approximation of functions, and despite what Silicon Valley would have you believe, most of the significant advances have been in creating useful meta-structures for modeling certain kinds of problems (e.g. convolution for efficiently processing transformations that care about local structure across dimensions of data, or qkv attention for keeping throughlines of non-local correspondences intact through a long sequence). Neural networks as a flavor of empirical function approximation happened to scale well, and then a bunch of people who saw how much this scale improved the models' capabilities but couldn't be bothered to understand the structural component concluded that scale somehow magically gets you to every unsolved problem being solved. It's also convenient for business types that if you buy this premise, any unicorn they want to promise is just a matter of throwing obscene amounts of resources at the problem (through their company of course) I think probably the general idea of dynamic structures that are versatile in their ability to approximate functional models is at least a solid hypothesis for how some biological intelligence works at some level (I think maybe the "fluid/crystallized" intelligence distinction some psychology uses is informative here - a strong world model probably informs a lot of quick acquisition of relationships, but most intelligent systems clearly posess strong feedback mechanisms for capturing new models), though I definitely agree that a focus on how best to throw a ton of scale at these models doesn't seem like a fruitful path for actionably learning how to build or analyze intelligent systems in the way we usually think about, nor is it, well, sustainable. Moore's law appeals to business people because buying more computronium feels more like a predictable input-output relationship to put capital into, but even if we're just talking about raw computation speed advances in algorithms tend to dwarf advances in computing power in the long run. I think the same will hold true in AGI
    - YeGoblynQueenne 2y ago
      Yeah, very good points. To be fair there are people who have argued the big data side who have clearly solid knowledge of AI and are not just SV suits, for example I remember Yann LeCun in a debate with Christopher Manning, where Manning was arguing for the importance of "structure" and LeCun was arguing against it. Or see the "Bitter Lesson", mentioned in a parent comment. That may have become a total shibboleth of the Silicon bros but Rich Sutton, who wrote the eponymous article, is the guy who wrote the book on Reinforcement Learning (literally). And then Rodney Brooks' replied with his "Better Lesson" (https://rodneybrooks.com/a-better-lesson/ https://rodneybrooks.com/a-better-lesson/). So there's a lot of debate in this and I don't reckon we'll have a consensus soon. It should be clear which side I'm on- I work with firmly model-based AI ("planning is the model-based approach to autonomous behaviour" has become my shibboleth - see Bonnet and Geffner's book on planning: https://link.springer.com/book/10.1007/978-3-031-01564-9 https://link.springer.com/book/10.1007/978-3-031-01564-9) so maybe it's a deformation professionelle. And even LCun's recent plans for JEPA are very consciously model-based, except he wants to learn his models from data; which is not a bad idea I suppose.
- advael 2y ago
  You seem to misunderstand why generalization is important for making claims about intelligent systems. To illustrate this, we could really easily design a system that encodes all the test set questions and their answers, puts them in an enormous hash table, and looks up the correct answer to each challenge when presented with it. This could probably score 100% on ARC if given the entire test set. Would you call this AGI? What if I put it through a transformer as a hashing function? The mainstream attention LLMs have garnered has added a bunch of noise to the way we talk about machine learning systems, and unfortunately the companies releasing them are partially to blame for this. That doesn't mean we should change the definition of success for various benchmarks to better suit lay misunderstandings of how this all works
  - rfoo 2y ago
    First, LLMs are not AGI. Never will be. Can we talk now? > if given the entire test set. I don't want the entire test set. Or any single one in the test set. The problem here is ARC challenge deliberately give a training set with different distribution than both the public and the private test set. It's like having only 1+1=2, 3+5=8, 9+9=18 in training set and then 1+9=10, 5*5=25, 16/2=8, (0!+0!+0!+0!)!=24 in test set. I can see the argument of "giving the easy problems as demonstration of rules and then with 'intelligence' [1] you should be able to get harder ones (i.e. a different distribution)", but I don't believe it's a good way to benchmark current methods, mainly because there are shortcuts. Like I can teach my kids how factorial works and ! means factorial, instead of teaching them how addition works only and make them figure out how multiplication, division and factorial works and what's the notation. [1] Whatever that means.
    - advael 2y ago
      Okay I admit I'm confused and think I probably missed a crucial thing here. You're saying the publicly available problem set isn't indicative of the distribution of the test set? If so, I can see why you object to that. Still, it's potentially possible that the test's intention is to demonstrate something like progressive integration of compositionality given an abstract model. A lot of machine learning systems can do well as long as they've seen an example similar to the problem they've been presented, but can't do things like respond to a situation that presents them with a novel composition of two abstractions they seem to have already learned in the way a human can trivially. Like only having [1+1=2, 4+5=9, 2+10=12] in the training set and [2*5=10, 3/4=.75, 2^8=256] in the test set would be bad, but something like [1+1=2, 3+4*2=11, 5*3=15, 2*7=14, 1+3/5=1.8, 3^3=27] vs [2+4*3=14, 3+3^2+4=16, 2*3/4+2^3/2^4=2] might not be, depending on what they're trying to test Compositionality of information, especially of abstractions (like rules or models of a phenomenon), is a key criterion in a lot of people's attempts to operationally define "intelligence" (which I agree is overall a nebulous and overloaded concept, but if we're going to make claims about it we need at least a working definition for any particular test we're doing) I could see that meaning that the test set problems need to be "harder" in the sense that presenting compositions of rules in training doesn't preclude memorizing the combinations. But this is just a guess, I'm not involved in ARC and don't know, obviously*

rfoo 2y ago

... and we know that if we really want to nail it we'd better just pay someone else to create 1,000,000 more harder problems for training (without looking at any in test set, of course). i.e. make the training set distribution similar to test set again. Because the thing we have now is data-hungry. Your brain is pre-trained on other similar challenges as well. What's the point of requiring it to "generalize beyond the training distribution" with so few samples? Really, I thought LLMs ended this "can we pretrain on in-house prepared private data for ILSVRC" flame war already.

YeGoblynQueenne 2y ago
The problem with that it is we know approaches that can generalise very well from very few examples, even one example, without any kind of pretraining, That requires a good background theory of the target domain (a "world model" in more modern parlance), and we don't know how to automatically generate that kind of theory; only human minds can do it, for now. But given such a theory the number of examples needed can be as few as 1. Clearly, if you can learn from one example, but find yourself using thousands, you've taken a wrong turn somewhere. The concern with the data-hungry approach to machine learning, that at least some of us have, is that it has given up on the effort to figure out how to learn good background theories and turned instead to getting the best performance possible in the dumbest possible way, relying on the largest available amount of examples and compute. That's a trend against everything else in computer science (and even animal intelligence) where the effort is to make everything smaller, cheaper, faster, smarter: it's putting all the eggs in the basket of making it big, slow and dumb, and hoping that this will somehow solve... intelligence. A very obvious contradiction. Suppose we lived in a world that didn't have a theory of computational complexity and didn't know that some programs are more expensive to run than others. Would it be the case in that world, that computer scientists competed in solving ever larger instances of the Traveling Salesperson Problem, using ever larger computers, without even trying to find good heuristics exploiting the structure of the problem and simply trying to out-brute-force each other? That world would look a lot like where we are now with statistical machine learning: a pell-mell approach to throwing all resources at a problem that we just don't know how to solve, and don't even know if we can solve.
- advael 2y ago
  The formalism that data-driven machine learning leans on is empirical tuning of stochastic search to drive approximation of functions, and despite what Silicon Valley would have you believe, most of the significant advances have been in creating useful meta-structures for modeling certain kinds of problems (e.g. convolution for efficiently processing transformations that care about local structure across dimensions of data, or qkv attention for keeping throughlines of non-local correspondences intact through a long sequence). Neural networks as a flavor of empirical function approximation happened to scale well, and then a bunch of people who saw how much this scale improved the models' capabilities but couldn't be bothered to understand the structural component concluded that scale somehow magically gets you to every unsolved problem being solved. It's also convenient for business types that if you buy this premise, any unicorn they want to promise is just a matter of throwing obscene amounts of resources at the problem (through their company of course) I think probably the general idea of dynamic structures that are versatile in their ability to approximate functional models is at least a solid hypothesis for how some biological intelligence works at some level (I think maybe the "fluid/crystallized" intelligence distinction some psychology uses is informative here - a strong world model probably informs a lot of quick acquisition of relationships, but most intelligent systems clearly posess strong feedback mechanisms for capturing new models), though I definitely agree that a focus on how best to throw a ton of scale at these models doesn't seem like a fruitful path for actionably learning how to build or analyze intelligent systems in the way we usually think about, nor is it, well, sustainable. Moore's law appeals to business people because buying more computronium feels more like a predictable input-output relationship to put capital into, but even if we're just talking about raw computation speed advances in algorithms tend to dwarf advances in computing power in the long run. I think the same will hold true in AGI
  - YeGoblynQueenne 2y ago
    Yeah, very good points. To be fair there are people who have argued the big data side who have clearly solid knowledge of AI and are not just SV suits, for example I remember Yann LeCun in a debate with Christopher Manning, where Manning was arguing for the importance of "structure" and LeCun was arguing against it. Or see the "Bitter Lesson", mentioned in a parent comment. That may have become a total shibboleth of the Silicon bros but Rich Sutton, who wrote the eponymous article, is the guy who wrote the book on Reinforcement Learning (literally). And then Rodney Brooks' replied with his "Better Lesson" (https://rodneybrooks.com/a-better-lesson/ https://rodneybrooks.com/a-better-lesson/). So there's a lot of debate in this and I don't reckon we'll have a consensus soon. It should be clear which side I'm on- I work with firmly model-based AI ("planning is the model-based approach to autonomous behaviour" has become my shibboleth - see Bonnet and Geffner's book on planning: https://link.springer.com/book/10.1007/978-3-031-01564-9 https://link.springer.com/book/10.1007/978-3-031-01564-9) so maybe it's a deformation professionelle. And even LCun's recent plans for JEPA are very consciously model-based, except he wants to learn his models from data; which is not a bad idea I suppose.
advael 2y ago
You seem to misunderstand why generalization is important for making claims about intelligent systems. To illustrate this, we could really easily design a system that encodes all the test set questions and their answers, puts them in an enormous hash table, and looks up the correct answer to each challenge when presented with it. This could probably score 100% on ARC if given the entire test set. Would you call this AGI? What if I put it through a transformer as a hashing function? The mainstream attention LLMs have garnered has added a bunch of noise to the way we talk about machine learning systems, and unfortunately the companies releasing them are partially to blame for this. That doesn't mean we should change the definition of success for various benchmarks to better suit lay misunderstandings of how this all works
- rfoo 2y ago
  First, LLMs are not AGI. Never will be. Can we talk now? > if given the entire test set. I don't want the entire test set. Or any single one in the test set. The problem here is ARC challenge deliberately give a training set with different distribution than both the public and the private test set. It's like having only 1+1=2, 3+5=8, 9+9=18 in training set and then 1+9=10, 5*5=25, 16/2=8, (0!+0!+0!+0!)!=24 in test set. I can see the argument of "giving the easy problems as demonstration of rules and then with 'intelligence' [1] you should be able to get harder ones (i.e. a different distribution)", but I don't believe it's a good way to benchmark current methods, mainly because there are shortcuts. Like I can teach my kids how factorial works and ! means factorial, instead of teaching them how addition works only and make them figure out how multiplication, division and factorial works and what's the notation. [1] Whatever that means.
  - advael 2y ago
    Okay I admit I'm confused and think I probably missed a crucial thing here. You're saying the publicly available problem set isn't indicative of the distribution of the test set? If so, I can see why you object to that. Still, it's potentially possible that the test's intention is to demonstrate something like progressive integration of compositionality given an abstract model. A lot of machine learning systems can do well as long as they've seen an example similar to the problem they've been presented, but can't do things like respond to a situation that presents them with a novel composition of two abstractions they seem to have already learned in the way a human can trivially. Like only having [1+1=2, 4+5=9, 2+10=12] in the training set and [2*5=10, 3/4=.75, 2^8=256] in the test set would be bad, but something like [1+1=2, 3+4*2=11, 5*3=15, 2*7=14, 1+3/5=1.8, 3^3=27] vs [2+4*3=14, 3+3^2+4=16, 2*3/4+2^3/2^4=2] might not be, depending on what they're trying to test Compositionality of information, especially of abstractions (like rules or models of a phenomenon), is a key criterion in a lot of people's attempts to operationally define "intelligence" (which I agree is overall a nebulous and overloaded concept, but if we're going to make claims about it we need at least a working definition for any particular test we're doing) I could see that meaning that the test set problems need to be "harder" in the sense that presenting compositions of rules in training doesn't preclude memorizing the combinations. But this is just a guess, I'm not involved in ARC and don't know, obviously*