4 ms·

The Infinite Space Between Words

16 points by selectnull 12y ago

delluminatus 12y ago
This raises an interesting question: If there was a conscious AI system running on a computer, what kind of time scales would it actually perceive? It's not like the AI would experience time on a CPU cycle-by-cycle basis. I would be surprised if we had an AI that perceived time in "steps" smaller than 10ms, unless it was a system specifically designed for extremely fast reactions. I would guess that an AI would be backed by, basically, an event loop that handles "sensory" data X times a second. Sure, it might do a lot of computation in each tick. But even if it does extensive predictive simulation or data analysis, I doubt the AI would be conscious of that simulation process, any more than we are aware of our own internal predictive models. We just predict.
- DanAndersen 12y ago
  It's kind of the opposite of what you're talking about, but I'd recommend reading "Permutation City" by Greg Egan (if you haven't already). The main premise is that in the future, human mind uploading is possible, but there's some upper limit on computation that prevents the uploads from running at real time. Instead, they are simulated at something like a ratio of 17 to 1. The story explores concepts of isolation as a growing society of uploads becomes more and more detached from the causality of "reality." The real world isn't a place you can conveniently interact with when you run 17 times slower than everyone in it. Plus there are interesting segments about an uploaded person having himself simulated with greater and greater real time between each simulated timestep, and with "skipped frames" that are imperceptible to the one being simulated (e.g. counting from 1 to 10, where you have the memory of having said "9" and being about to say "10" but the "frame" where you said "9" was never actually simulated and it skipped ahead).
- Houshalter 12y ago
  From Predicting the Future (http://www.sl4.org/wiki/PredictingTheFuture http://www.sl4.org/wiki/PredictingTheFuture): >Neurons are slow. Transistors are fast. Neurons fire around 200 times per second. A modern CPU delivers 2 billion operations per second, which is a difference of about ten million. Now, at least in the beginning, we'll probably burn a lot of that speed difference making up for the difference in parallelism. But while you can convert fast into parallel, you can't convert parallel into fast. If you can do 2 billion operations, one after the other, in one second, then you can do 200 operations ten million times, in one second. But the reverse isn't true - you can't take ten million slow neurons and simulate one really really fast neuron. The human brain has incredible massive parallelism - something like a hundred trillion synapses. But how much of that computing power is necessary just do anything at all in real time when your CPU speed is 200 operations per second? >...AI might be much slower, it might be much faster, it might be much slower on some things and much faster on others, but it won't be just exactly the human set of speeds. >...when I talk about AIs being slower or faster, I'm not talking about differences in speed that are like slow humans or fast humans. We spend our whole lives living and talking with other humans, and we naturally tend to focus on the kind of differences that separate humans from each other. Like, maybe most people take one second to figure something out, but one person might see it intuitively in a fifth of a second, and another person might be slower and take five seconds. We're not talking about that kind of difference. That kind of difference is tiny on the scale we're talking about. The difference in serial speed between a neuron and a CPU is a factor of ten million, and the difference in parallelism between today's computer clusters and the human brain is around ten billion to one in the other direction. Relative to the space of minds in general, all humans are the same make and model of car.

16 points by selectnull 12y ago

This raises an interesting question: If there was a conscious AI system running on a computer, what kind of time scales would it actually perceive? It's not like the AI would experience time on a CPU cycle-by-cycle basis. I would be surprised if we had an AI that perceived time in "steps" smaller than 10ms, unless it was a system specifically designed for extremely fast reactions. I would guess that an AI would be backed by, basically, an event loop that handles "sensory" data X times a second. Sure, it might do a lot of computation in each tick. But even if it does extensive predictive simulation or data analysis, I doubt the AI would be conscious of that simulation process, any more than we are aware of our own internal predictive models. We just predict.

DanAndersen 12y ago
It's kind of the opposite of what you're talking about, but I'd recommend reading "Permutation City" by Greg Egan (if you haven't already). The main premise is that in the future, human mind uploading is possible, but there's some upper limit on computation that prevents the uploads from running at real time. Instead, they are simulated at something like a ratio of 17 to 1. The story explores concepts of isolation as a growing society of uploads becomes more and more detached from the causality of "reality." The real world isn't a place you can conveniently interact with when you run 17 times slower than everyone in it. Plus there are interesting segments about an uploaded person having himself simulated with greater and greater real time between each simulated timestep, and with "skipped frames" that are imperceptible to the one being simulated (e.g. counting from 1 to 10, where you have the memory of having said "9" and being about to say "10" but the "frame" where you said "9" was never actually simulated and it skipped ahead).
Houshalter 12y ago
From Predicting the Future (http://www.sl4.org/wiki/PredictingTheFuture http://www.sl4.org/wiki/PredictingTheFuture): >Neurons are slow. Transistors are fast. Neurons fire around 200 times per second. A modern CPU delivers 2 billion operations per second, which is a difference of about ten million. Now, at least in the beginning, we'll probably burn a lot of that speed difference making up for the difference in parallelism. But while you can convert fast into parallel, you can't convert parallel into fast. If you can do 2 billion operations, one after the other, in one second, then you can do 200 operations ten million times, in one second. But the reverse isn't true - you can't take ten million slow neurons and simulate one really really fast neuron. The human brain has incredible massive parallelism - something like a hundred trillion synapses. But how much of that computing power is necessary just do anything at all in real time when your CPU speed is 200 operations per second? >...AI might be much slower, it might be much faster, it might be much slower on some things and much faster on others, but it won't be just exactly the human set of speeds. >...when I talk about AIs being slower or faster, I'm not talking about differences in speed that are like slow humans or fast humans. We spend our whole lives living and talking with other humans, and we naturally tend to focus on the kind of differences that separate humans from each other. Like, maybe most people take one second to figure something out, but one person might see it intuitively in a fifth of a second, and another person might be slower and take five seconds. We're not talking about that kind of difference. That kind of difference is tiny on the scale we're talking about. The difference in serial speed between a neuron and a CPU is a factor of ten million, and the difference in parallelism between today's computer clusters and the human brain is around ten billion to one in the other direction. Relative to the space of minds in general, all humans are the same make and model of car.