So have you thought about what “data points” mean? If the data is random samples from the mandelbrot set, the maximum information the AI can ever learn is just the root equation used to generate the set.
Human agents control a robotics system where we take actions and observe the results on our immediate environment. This sort of information seems to lead to very rapid learning especially for things where the consequences are near term and observable. You are essentially performing a series of experiments where you try action A vs B and observe what the environment does. This let’s you rapidly cancel out data that doesn’t matter, its how you learn that lighting conditions don’t affect how a rock falls when you drop it.
Point is the obvious training data for an AI would be similar. It needs to manipulate, both in sims and reality, the things we need it to learn about
I’ve thought about it enough to know I’m confused! :)
I like your point about active learning (is that the right term?). I wonder how powerful GPT-3 would be if instead of being force-fed random internet text from a firehose, it had access to a browser and could explore (would need some sort of curiosity reward signal?). Idk, probably this isn’t a good idea or else someone would have done it.
I don’t know that GPT-3 is the best metric for ‘progress towards general intelligence’. One example of the agents receiving ‘active’ data that resulted in interesting results is thisOpenAI experiment.
In this case the agents cannot emit text—which is what GPT-3 is doing that makes us feel it’s “intelligent”—but can cleverly manipulate their environment in complex ways not hardcoded in. The agents in this experiment are learning both movement to control how they view the environment and to use a few simple tools to accomplish a goal.
To me this seems like the most promising way forward. I think that robust agents that can control real robots to do things, with those things becoming increasingly complex and difficult as the technology improves, might in fact be the “skeleton” of what would later allow for “real” sentience.
Because from our perspective, this is our goal. We don’t want an agent that can babble and seem smart, we want an agent that can do useful things—things we were paying humans to do—and thus extend what we can ultimately accomplish. (yes, in the immediate short term it unemploys lots of humans, but it also would make possible new things that previously we needed lots of humans to do. It also should allow for doing things we know how to do now but with better quality/on a more broader scale. )
More exactly, how do babies learn? Yes, they learn to babble, but they also learn a set of basic manipulations of their body—adjusting their viewpoint—and manipulate the environment with their hands—learning how it responds.
We can discuss more, I think I know how we will “get there from here” in broad strokes. I don’t think it will be done by someone writing a relatively simple algorithm and getting a sudden breakthrough that allows for sentience, I think it will be done by using well defined narrow domain agents that each do something extremely well—and by building higher level agents on top of this foundation in a series of layers, over years to decades, until you reach the level of abstraction of “modify thy own code to be more productive”.
As a trivial example of abstraction, you can make a low level agent that all it does is grab things with a simulate/real robotic hand, and an agent 1 level up that decides what to grab by which grab has the highest predicted reward.
We can discuss more, I think I know how we will “get there from here” in broad strokes. I don’t think it will be done by someone writing a relatively simple algorithm and getting a sudden breakthrough that allows for sentience, I think it will be done by using well defined narrow domain agents that each do something extremely well—and by building higher level agents on top of this foundation in a series of layers, over years to decades, until you reach the level of abstraction of “modify thy own code to be more productive”.
I’d be interested to hear more about this. It sounds like this could maybe happen pretty soon with large, general language models like GPT-3 + prompt programming + a bit of RL.
So have you thought about what “data points” mean? If the data is random samples from the mandelbrot set, the maximum information the AI can ever learn is just the root equation used to generate the set.
Human agents control a robotics system where we take actions and observe the results on our immediate environment. This sort of information seems to lead to very rapid learning especially for things where the consequences are near term and observable. You are essentially performing a series of experiments where you try action A vs B and observe what the environment does. This let’s you rapidly cancel out data that doesn’t matter, its how you learn that lighting conditions don’t affect how a rock falls when you drop it.
Point is the obvious training data for an AI would be similar. It needs to manipulate, both in sims and reality, the things we need it to learn about
I’ve thought about it enough to know I’m confused! :)
I like your point about active learning (is that the right term?). I wonder how powerful GPT-3 would be if instead of being force-fed random internet text from a firehose, it had access to a browser and could explore (would need some sort of curiosity reward signal?). Idk, probably this isn’t a good idea or else someone would have done it.
I don’t know that GPT-3 is the best metric for ‘progress towards general intelligence’. One example of the agents receiving ‘active’ data that resulted in interesting results is thisOpenAI experiment.
In this case the agents cannot emit text—which is what GPT-3 is doing that makes us feel it’s “intelligent”—but can cleverly manipulate their environment in complex ways not hardcoded in. The agents in this experiment are learning both movement to control how they view the environment and to use a few simple tools to accomplish a goal.
To me this seems like the most promising way forward. I think that robust agents that can control real robots to do things, with those things becoming increasingly complex and difficult as the technology improves, might in fact be the “skeleton” of what would later allow for “real” sentience.
Because from our perspective, this is our goal. We don’t want an agent that can babble and seem smart, we want an agent that can do useful things—things we were paying humans to do—and thus extend what we can ultimately accomplish. (yes, in the immediate short term it unemploys lots of humans, but it also would make possible new things that previously we needed lots of humans to do. It also should allow for doing things we know how to do now but with better quality/on a more broader scale. )
More exactly, how do babies learn? Yes, they learn to babble, but they also learn a set of basic manipulations of their body—adjusting their viewpoint—and manipulate the environment with their hands—learning how it responds.
We can discuss more, I think I know how we will “get there from here” in broad strokes. I don’t think it will be done by someone writing a relatively simple algorithm and getting a sudden breakthrough that allows for sentience, I think it will be done by using well defined narrow domain agents that each do something extremely well—and by building higher level agents on top of this foundation in a series of layers, over years to decades, until you reach the level of abstraction of “modify thy own code to be more productive”.
As a trivial example of abstraction, you can make a low level agent that all it does is grab things with a simulate/real robotic hand, and an agent 1 level up that decides what to grab by which grab has the highest predicted reward.
I’d be interested to hear more about this. It sounds like this could maybe happen pretty soon with large, general language models like GPT-3 + prompt programming + a bit of RL.