4 ms·

AGI hasn't been delayed. It's been overpromised. And the proponents will continue to overpromise right up till when it arrives. We need a breakthrough in AI res

AGI hasn't been delayed. It's been overpromised. And the proponents will continue to overpromise right up till when it arrives. We need a breakthrough in AI research of at least two magnitudes before even tackling the AGI problem. One is we need to understand the mechanisms of neural networks at least an order of magnitude better than we currently do. Best article on the subject that concisely describes NNs is The NN zoo[0]. I have no idea what needs to happen for us to collectively understand the mechanisms better. We need an Einstein moment I think where some researcher has a happy thought and a decade where the rest of us catch up (at minimum). The other is we need to understand the problem we are solving. At it's core it seems very easy to define intelligence but you very soon realize that it's an endless Matryoska with no discernible root. It's very easy to define intelligence in terms of solving elementary cogitation tasks and this is what we are doing writ-large. This leads to a culture of solving Weak AI tasks, hoping that the aggregation of the solutions will lead to AGI. Well, it's only going to do so if we are engineering solutions that will solve the larger problem and if we don't understand it well enough it's going to be a really random occurrence finding it. [0] https://www.asimovinstitute.org/neural-network-zoo/ https://www.asimovinstitute.org/neural-network-zoo/

FeepingCreature 7y ago
IMO the one thing we really need is some reliable way to do online learning. If we can get NNs to the point where you can just say, somewhere in code, "make this decision NNly, based on this set of features" and get reasonably good decisionmaking if the decision is in some sense approximable as a function of the features, I think GOFAI suddenly becomes feasible again. We would then enter an area of risk where we don't really understand NNs and we don't really understand the algorithms we can build with them either, but these algorithms could still have a lot of power. The aphorism used to be "hard things are easy and easy things are hard". Ie. arithmetic with large numbers is trivial, but opening a door or recognizing a label is the most difficult problem in the world. We're starting to close in on reliable, possibly prepackaged solutions for the latter; on the other hand, understanding the logic of intelligence turns out to have some teeth after all. Reality would do us a great favor if building genuine intelligence required actual understanding and knowledge of how to build reliable cognition; unfortunately our precedent suggests it's mostly done by kludging a half-baked barely-debugged motivation system onto a big heap of instinct-driven neural networks. I'm worried some spark will get the bright idea to combine some approximating variant of AIXI and a neural network for theory selection...
- Beltiras 7y ago
  To your first point I think that's actually where we are. Poking in the dark with no real understanding of the tools at our disposal. And the algorithms DO have power. We don't know WHY, except that we are emulating solutions given to us by biological evolution. To your second point, that's basically what evolution did. Poked in the dark for several billions of years and ended up with solution machines (us).
  - FeepingCreature 7y ago
    Yes, I'm aware, that's in large part why it worries me. And I think you undersell the importance of online learning. NNs are far from "plug in and forget."
keenmaster 7y ago
Has anyone tried aggregating top ML algorithms in each perceptual function (vision, audio, etc) and cognitive function into distinct cortices which are then governed by an AGI meta-algorithm? If that method could yield any success, it wouldn't be any time soon. I imagine we'd need more advanced algorithms for each cortex first (garbage in garbage out) and modern ML is still a young field. However, I'm still curious if that could work. Modeling each neuron seems like a more arduous and perhaps unnecessary task.
- ilaksh 7y ago
  Yes, along those lines is the concept of 'cognitive architecture' with efforts going back close to the beginning of practical AI research.
zamalek 7y ago
> we need to understand the mechanisms of neural networks at least an order of magnitude better than we currently do The problem seems more fundamental than this. We may understand intelligence, but we do not understand human-like intelligence. Human-like intelligence originates from more than a neural network; there are hormones and other things involved. Possibly even quantum mechanics, in some of the more esoteric explanations. > The other is we need to understand the problem we are solving. Indeed, we are asserting that we will soon have a solution (AGI) without first knowing what the problem is (human-like intelligence). > it's going to be a really random occurrence finding it. I'm placing my bets on this. Human-like intelligence (and by unavoidable association, experience) is, by definition, subjective. The tool that we use to observe and explain the universe around us is science. We have no such tool for the universe within us.
- hnick 7y ago
  Even dog-like intelligence would be extremely useful. It's interesting when training a dog. Sometimes you can see the gears turning and know their thought process and conclusion before they do, since we are just much smarter. And other times, they come up with something that makes no sense at all based on their "dog logic". I guess their brain just works differently. Anyway, focusing on simply humans may be short sighted. I often understand systems better by comparing them to similar systems and working out the differences. If anything bears fruit I think it'll be the researchers that are starting small and trying to replicate a worm brain. Then build from there, faster than evolution, because we can.
credit_guy 7y ago
> we need to understand the mechanisms of neural networks at least an order of magnitude better than we currently do Many, many orders of magnitude more. I can't provide a link, but I read a while ago about this fascinating experiment: you put a few cardiac cells in a petri dish. They start crawling around. When they find each other they stick to each other, and after a small group forms, they start beating together. We think of biological neurons as little more than their ANN cousins. Take a few inputs from a few dendrites, apply a RELU or some other non-linear function to them, send the result through the axon. I think a better way to think of them is they are living things themselves, humongously complex. It's very unlikely they are as simple as we make them to be.

FeepingCreature 7y ago

IMO the one thing we really need is some reliable way to do online learning. If we can get NNs to the point where you can just say, somewhere in code, "make this decision NNly, based on this set of features" and get reasonably good decisionmaking if the decision is in some sense approximable as a function of the features, I think GOFAI suddenly becomes feasible again. We would then enter an area of risk where we don't really understand NNs and we don't really understand the algorithms we can build with them either, but these algorithms could still have a lot of power. The aphorism used to be "hard things are easy and easy things are hard". Ie. arithmetic with large numbers is trivial, but opening a door or recognizing a label is the most difficult problem in the world. We're starting to close in on reliable, possibly prepackaged solutions for the latter; on the other hand, understanding the logic of intelligence turns out to have some teeth after all. Reality would do us a great favor if building genuine intelligence required actual understanding and knowledge of how to build reliable cognition; unfortunately our precedent suggests it's mostly done by kludging a half-baked barely-debugged motivation system onto a big heap of instinct-driven neural networks. I'm worried some spark will get the bright idea to combine some approximating variant of AIXI and a neural network for theory selection...

Beltiras 7y ago
To your first point I think that's actually where we are. Poking in the dark with no real understanding of the tools at our disposal. And the algorithms DO have power. We don't know WHY, except that we are emulating solutions given to us by biological evolution. To your second point, that's basically what evolution did. Poked in the dark for several billions of years and ended up with solution machines (us).
- FeepingCreature 7y ago
  Yes, I'm aware, that's in large part why it worries me. And I think you undersell the importance of online learning. NNs are far from "plug in and forget."

keenmaster 7y ago

Has anyone tried aggregating top ML algorithms in each perceptual function (vision, audio, etc) and cognitive function into distinct cortices which are then governed by an AGI meta-algorithm? If that method could yield any success, it wouldn't be any time soon. I imagine we'd need more advanced algorithms for each cortex first (garbage in garbage out) and modern ML is still a young field. However, I'm still curious if that could work. Modeling each neuron seems like a more arduous and perhaps unnecessary task.

ilaksh 7y ago
Yes, along those lines is the concept of 'cognitive architecture' with efforts going back close to the beginning of practical AI research.

zamalek 7y ago

> we need to understand the mechanisms of neural networks at least an order of magnitude better than we currently do The problem seems more fundamental than this. We may understand intelligence, but we do not understand human-like intelligence. Human-like intelligence originates from more than a neural network; there are hormones and other things involved. Possibly even quantum mechanics, in some of the more esoteric explanations. > The other is we need to understand the problem we are solving. Indeed, we are asserting that we will soon have a solution (AGI) without first knowing what the problem is (human-like intelligence). > it's going to be a really random occurrence finding it. I'm placing my bets on this. Human-like intelligence (and by unavoidable association, experience) is, by definition, subjective. The tool that we use to observe and explain the universe around us is science. We have no such tool for the universe within us.

hnick 7y ago
Even dog-like intelligence would be extremely useful. It's interesting when training a dog. Sometimes you can see the gears turning and know their thought process and conclusion before they do, since we are just much smarter. And other times, they come up with something that makes no sense at all based on their "dog logic". I guess their brain just works differently. Anyway, focusing on simply humans may be short sighted. I often understand systems better by comparing them to similar systems and working out the differences. If anything bears fruit I think it'll be the researchers that are starting small and trying to replicate a worm brain. Then build from there, faster than evolution, because we can.

credit_guy 7y ago

> we need to understand the mechanisms of neural networks at least an order of magnitude better than we currently do Many, many orders of magnitude more. I can't provide a link, but I read a while ago about this fascinating experiment: you put a few cardiac cells in a petri dish. They start crawling around. When they find each other they stick to each other, and after a small group forms, they start beating together. We think of biological neurons as little more than their ANN cousins. Take a few inputs from a few dendrites, apply a RELU or some other non-linear function to them, send the result through the axon. I think a better way to think of them is they are living things themselves, humongously complex. It's very unlikely they are as simple as we make them to be.