4 ms·

We do have a near endless supply of human generated data. But another path is to train each component individually and then let the network combine them into a

We do have a near endless supply of human generated data. But another path is to train each component individually and then let the network combine them into a unified whole. I suspect synthetic data will go a long way here. The problem is one of inducing the right kind of information dynamics in the network. We can potentially synthesize unlimited data that require specific information dynamics to solve efficiently. In my view, we will know we reached AGI when a network can learn to play chess well through self-directed analysis after only being given the description of the game. Obviously chess is a solved computational problem, but only through direct programming of minimax search. If we can teach a network with enough generality such that it can self-induce the right computational structure within itself to learn, e.g. induce a self-play subnetwork, then this is plainly AGI, superintelligence even. But I see a direct path from where we are now to such a network. This is why I think AGI is imminent.

YeGoblynQueenne 4y ago
Well, chess is not a "solved computational problem". We have chess engines that can play better than every human, but there is no system that has mapped the entire chess game tree. That's what is usually meant by "solved" in the context of game-playing AI. MuZero, one of the successors of AlphaGo, learned to play chess (and a few other games besides) better than its predecessors while only receiving feedback about the end-state of a game (win, loss, draw), and not even the rules of the game, so it basically learned to play just by initially playing many (many) illegal moves. No "description of the game" needed. That didn't lead to AGI. Progress in game-playing AI has not led to any progress outside game-playing. It's easy to see why. For chess and other classic board games there are specialised algorithms (minimax and Monte-Carlo Tree Search) that are sufficient to dominate any opponent, given enough computing power. Those algorithms are useless outside of board games (even for card games they're no good and dominant players use different approaches). For computer games such as Atari, where AI players often have superhuman performance, the complexity of the game world is tiny, compared to the real world, and techniques (Reinforcement Learning, basically) do not transfer easily from one to the other. AI pioneers, like Alan Turing, John McCarthy, Claude Shannon and Donald Michie were interested in game playing as a gateway into understanding of the human mind, because it was assumed that the only way to play chess, and other classic board games, was to play them like the human mind plays them. It turns out, that's not the case. You can play very strong chess with a brute force search in alpha-beta minimax, and there is no human whose mind can process alpha-beta minimax well enough to beat a computer. That still tells us nothing about how the human mind works, or how to replicate it. What we have learned from all those successes in game-playing AI is that it's no use as a way towards general AI. >> We do have a near endless supply of human generated data. Yes, but what kind? We train systems on very narrow tasks: object identification in images, text prediction, speech recognition, document classification, etc. Those are not examples of "human intelligence", by which we could train a system to simulate "human intelligence", as a distinct process. They are examples of very specialised tasks that humans can perform using our intelligence, but it's not clear that intelligence is even needed for many of those tasks. That's why, for example, we have damn good image classifiers, but they are not AGI. The problem with statistical machine learning algorithms is that they have to be trained to do one thing, and one thing only. They're one-trick ponies. If you want them to do multiple things together, you have to train them with examples of all those things at once. See for example Google's Gato, which had to be trained on multiple modalities as one big batch of token sequences. Google claimed that Gato is a "generalist agent", but it, too, was only capable of performing a restricted set of tasks, the ones it was trained on, and no others. Humans, by contrast, don't need to be trained on all modalities at once: we can learn to read, and later learn to play ball, and later still to drive. And each of those things is a hundred different tasks engaging different "modalities" (our senses). No AI system can do that. AGI is not that easy as it perhaps might appear.
- hackinthebochs 4y ago
  This doesn't really engage with my points. Yes of course programming a narrow chess engine doesn't extrapolate to domains outside of games. But my point is once a system can induce a self-play dynamic in itself simply from being prompted with the rules of chess, then this represents a significant qualitative change. I only chose chess for the irony of it wrongly being represented as AGI-complete in the past. None of your points are relevant to this claim. The issue of how the human mind works is tangential, but its also not true that computer chess has no relation to how the human mind works. Neural net based evaluations result in moves that more closely match human expert choices[1], which suggests a similarity in mechanism compared to prior evaluation mechanisms. >The problem with statistical machine learning algorithms is that they have to be trained to do one thing, and one thing only. They're one-trick ponies. LLMs show the beginnings of breaking out of this limitation. People who argue LLMs are just another example in the history of narrow AI don't seem to have a strong response to the example of in-context learning. The straightforward understanding is that we're seeing the beginnings of true generalization. I have not seen a convincing argument that in-context learning is just another case of what we've seen before. [1]https://www.cs.toronto.edu/~ashton/pubs/maia-kdd2020.pdf https://www.cs.toronto.edu/~ashton/pubs/maia-kdd2020.pdf
  - YeGoblynQueenne 4y ago
    >> This doesn't really engage with my points. Sorry. I thought you described a system that did what MuZero did. It still seems to me that's what you're describing. "In-context learning" is a big potato. It's just some jargon that OpenAI came up with to play up the capabilities of GPT-3, just another example of the time-honoured tradition in AI of inventing grandiloquent terms to describe something much simpler. See Drew McDermot's "Artificial Intelligence meets Natural Stupidity", a paper from 1976, that castigates this already very common habit of AI researchers: https://dl.acm.org/doi/10.1145/1045339.1045340 https://dl.acm.org/doi/10.1145/1045339.1045340 To clarify, "in-context learning" is what we used to call "sampling from a model" back in my time (in 2014, when I did my Master's and studied neural nets and language modelling, and also game-playing algos; not coincidentally, because I usually engage in conversation about things that I understand, so you won't catch me talking about quantum computers, say, or physics). Models got better and sampling from them gets better results, so OpenAI figured they could embiggen it and call it something more impressive than "sampling". But it's sampling, not learning. >> People who argue LLMs are just another example in the history of narrow AI don't seem to have a strong response to the example of in-context learning. I don't think it's very productive to try and fit opinions into "people who say" kinds of boxes. I don't treat my interlocutors like that. Perhaps I haven't yet heard all the arguments on any particular claim. I'll never know if I assume everyone is just repeating the same old stuff. Anyway there's no point in arguing anything here. If you've used language models before, you know how sampling works and it's obvious it's what OpenAI call "in-context learning". They're not even pretending it's not sampling. They're just claiming it's different because they're sampling from a large Language model. Edit: >> Neural net based evaluations result in moves that more closely match human expert choices[1], which suggests a similarity in mechanism compared to prior evaluation mechanisms. The paper you linked describes a system that was explicitly trained to predict human moves. That's probably why it can do better at simulating the way humans play chess than a system that wasn't trained to simulate the way humans play chess.

YeGoblynQueenne 4y ago

Well, chess is not a "solved computational problem". We have chess engines that can play better than every human, but there is no system that has mapped the entire chess game tree. That's what is usually meant by "solved" in the context of game-playing AI. MuZero, one of the successors of AlphaGo, learned to play chess (and a few other games besides) better than its predecessors while only receiving feedback about the end-state of a game (win, loss, draw), and not even the rules of the game, so it basically learned to play just by initially playing many (many) illegal moves. No "description of the game" needed. That didn't lead to AGI. Progress in game-playing AI has not led to any progress outside game-playing. It's easy to see why. For chess and other classic board games there are specialised algorithms (minimax and Monte-Carlo Tree Search) that are sufficient to dominate any opponent, given enough computing power. Those algorithms are useless outside of board games (even for card games they're no good and dominant players use different approaches). For computer games such as Atari, where AI players often have superhuman performance, the complexity of the game world is tiny, compared to the real world, and techniques (Reinforcement Learning, basically) do not transfer easily from one to the other. AI pioneers, like Alan Turing, John McCarthy, Claude Shannon and Donald Michie were interested in game playing as a gateway into understanding of the human mind, because it was assumed that the only way to play chess, and other classic board games, was to play them like the human mind plays them. It turns out, that's not the case. You can play very strong chess with a brute force search in alpha-beta minimax, and there is no human whose mind can process alpha-beta minimax well enough to beat a computer. That still tells us nothing about how the human mind works, or how to replicate it. What we have learned from all those successes in game-playing AI is that it's no use as a way towards general AI. >> We do have a near endless supply of human generated data. Yes, but what kind? We train systems on very narrow tasks: object identification in images, text prediction, speech recognition, document classification, etc. Those are not examples of "human intelligence", by which we could train a system to simulate "human intelligence", as a distinct process. They are examples of very specialised tasks that humans can perform using our intelligence, but it's not clear that intelligence is even needed for many of those tasks. That's why, for example, we have damn good image classifiers, but they are not AGI. The problem with statistical machine learning algorithms is that they have to be trained to do one thing, and one thing only. They're one-trick ponies. If you want them to do multiple things together, you have to train them with examples of all those things at once. See for example Google's Gato, which had to be trained on multiple modalities as one big batch of token sequences. Google claimed that Gato is a "generalist agent", but it, too, was only capable of performing a restricted set of tasks, the ones it was trained on, and no others. Humans, by contrast, don't need to be trained on all modalities at once: we can learn to read, and later learn to play ball, and later still to drive. And each of those things is a hundred different tasks engaging different "modalities" (our senses). No AI system can do that. AGI is not that easy as it perhaps might appear.

hackinthebochs 4y ago
This doesn't really engage with my points. Yes of course programming a narrow chess engine doesn't extrapolate to domains outside of games. But my point is once a system can induce a self-play dynamic in itself simply from being prompted with the rules of chess, then this represents a significant qualitative change. I only chose chess for the irony of it wrongly being represented as AGI-complete in the past. None of your points are relevant to this claim. The issue of how the human mind works is tangential, but its also not true that computer chess has no relation to how the human mind works. Neural net based evaluations result in moves that more closely match human expert choices[1], which suggests a similarity in mechanism compared to prior evaluation mechanisms. >The problem with statistical machine learning algorithms is that they have to be trained to do one thing, and one thing only. They're one-trick ponies. LLMs show the beginnings of breaking out of this limitation. People who argue LLMs are just another example in the history of narrow AI don't seem to have a strong response to the example of in-context learning. The straightforward understanding is that we're seeing the beginnings of true generalization. I have not seen a convincing argument that in-context learning is just another case of what we've seen before. [1]https://www.cs.toronto.edu/~ashton/pubs/maia-kdd2020.pdf https://www.cs.toronto.edu/~ashton/pubs/maia-kdd2020.pdf
- YeGoblynQueenne 4y ago
  >> This doesn't really engage with my points. Sorry. I thought you described a system that did what MuZero did. It still seems to me that's what you're describing. "In-context learning" is a big potato. It's just some jargon that OpenAI came up with to play up the capabilities of GPT-3, just another example of the time-honoured tradition in AI of inventing grandiloquent terms to describe something much simpler. See Drew McDermot's "Artificial Intelligence meets Natural Stupidity", a paper from 1976, that castigates this already very common habit of AI researchers: https://dl.acm.org/doi/10.1145/1045339.1045340 https://dl.acm.org/doi/10.1145/1045339.1045340 To clarify, "in-context learning" is what we used to call "sampling from a model" back in my time (in 2014, when I did my Master's and studied neural nets and language modelling, and also game-playing algos; not coincidentally, because I usually engage in conversation about things that I understand, so you won't catch me talking about quantum computers, say, or physics). Models got better and sampling from them gets better results, so OpenAI figured they could embiggen it and call it something more impressive than "sampling". But it's sampling, not learning. >> People who argue LLMs are just another example in the history of narrow AI don't seem to have a strong response to the example of in-context learning. I don't think it's very productive to try and fit opinions into "people who say" kinds of boxes. I don't treat my interlocutors like that. Perhaps I haven't yet heard all the arguments on any particular claim. I'll never know if I assume everyone is just repeating the same old stuff. Anyway there's no point in arguing anything here. If you've used language models before, you know how sampling works and it's obvious it's what OpenAI call "in-context learning". They're not even pretending it's not sampling. They're just claiming it's different because they're sampling from a large Language model. Edit: >> Neural net based evaluations result in moves that more closely match human expert choices[1], which suggests a similarity in mechanism compared to prior evaluation mechanisms. The paper you linked describes a system that was explicitly trained to predict human moves. That's probably why it can do better at simulating the way humans play chess than a system that wasn't trained to simulate the way humans play chess.