People choose google, because they like it , not because of some monopoly influence.
People choose Google because they like it.
They like Google because Google provides good results.
Google provides good results because they have huge datacenters and extreme amounts of data.
The cost of acquiring huge datacenters and extreme amounts of data provides a barrier to entry.
This barrier to entry produces a natural monopoly.
And so it is. Not all monopolies have to be ordained by the State.
If the computer was infinitely fast, there would be two programming languages.
1. Nonlinear optimization.
2. Teach-by-example, e.g. a neural net of maximum size for the RAM constraints, where all the weights are found by brute force.
The former is for when you know exactly what you want, the latter is for when you don't and use supervised learning instead.
For that matter, an infinitely fast computer is a hypercomputer, so 2. could easily be "the smallest Turing machine that produces the desired outputs given the inputs".
The market has no regret... some agents will, others won't.
That would be regret in the game theoretical sense. The regret of a strategy is the best payoff you could get minus the best you got; the "opportunity cost". There's an example here.
Since the market is based on individual predictors, the best it can do is somehow knowing the best predictor at each instant. That would correspond to zero regret. Any algorithm based on experts advice would have a regret greater than or equal to zero, even though it may not have an "emotional" regret.
If you consider the market to be a kind of weighted voting algorithm that exponentially amplifies the predictions of good experts (as you've said), then it doesn't greatly matter what the weights correspond to in the real world. All that matters is that good experts get their weights amplified, bad experts "go broke", that you have enough players to begin with to catch some good experts, and that the players can't do Sybil attacks to get around going broke.
As sites like StackOverflow show -- or MMOs for that matter -- there are plenty of people that are incentivized simply by getting a high score, even if that high score doesn't translate into real money. So while it may seem counterintuitive that play market should work, it's not that weird if you consider it from the point of view of an exponential weighting algorithm.
I was talking about real prediction markets, not as use in training NNs with virtual prediction markets. I can't think of any situation where this would even make sense... but maybe?
I might have been a bit hasty in my informal impossibility argument, as it were, but my line of thought was like this:
- The claim is that a prediction market can act as well as the best participant of that market, i.e. have zero regret.
- This should hold irrelvant of what the inputs are, whether the inputs are predictions by people or by other simpler algorithms.
- If the claim were true, you could dump a lot (and I mean an extreme amount) of comparatively simple algorithms into a virtual prediction market, and by the claim, the market would act just as well as the very best algorithm of the lot, producing a super AI.
- Since we don't have such super AIs in the real world, the claim must be wrong (and it is, because of the log(N) term).
More generally, if prediction markets were to have zero regret compared to the best expert, every kind of ensemble method would be easy. Just dump the individual methods into a prediction market and you'd get at least as good performance as the best individual method. The prediction market algorithm is the same irrespective of whether its inputs are by people or by computer algorithms, after all.
If anything Swarm AI is a prediction market, but with equal waiting for every "expert, and no reward feedback mechanism to promote the accurate and remove the inaccurate players.
Perhaps surprisingly, it's often hard to do better than nonadaptive methods. See for instance the Variance method of An empirical comparison of algorithms for aggregating expert predictions. Then again, I have no idea what particular algorithm Swarm AI uses; it may be a bad nonadaptive method.
If you're in the US, you've probably never had a chance to use a prediction market, because they are generally illegal there. I have no idea why.
I imagine the reason they're illegal is that they can produce perverse incentives. Suppose, for instance, that there's a prediction market on when the next act of terrorism will occur in the US, as defined by some criterion. You predict tomorrow, then go (or get a dupe to go) snipe off some people and send a manifesto to the newspapers. Since the market considered tomorrow to be very unlikely, you get a lot of money doing this, and a strategy like that might seem rather tempting for someone who's poor or desperate enough.
I'd thus expect play money prediction markets to be more easily accepted, since there's no reason to murder for Shiny Points or whatever. But I wouldn't be surprised if the law just refers to prediction markets in general rather than real money prediction markets in particular.
So, because it is a meta-prediction technique that optimally combines the results of all participating prediction techniques, it is always going to be better than any given underlying prediction system, or at worst case be as good as the best underlying prediction method.
No, because you're always going to take some hit determining what the optimal ensemble is (and even moreso if situations change or if there's noise). In exponential reweighting, the regret has an ln(N) term, which is what keeps you from just making an optimal AI by dumping an exponential number of algorithms into a virtual prediction market and getting a result "as good as the best underlying prediction method".
In practice, if you run the prediction market with real people, you're at risk of all the fun stuff that sometimes makes real markets behave poorly, such as bubbles. This happens because there's a feedback loop where people alter their predictions based on everybody else's predictions, as in a Keynesian beauty contest. In contrast, "the best underlying prediction method" would not have such problems because there would be no market for it to be influenced by.
I'm having a hard time picturing any culture that celebrates the "everyman". The "strong leader" is pretty much a universal cultural archetype.
A lot of cultures have social regulations of the form "don't think too much of yourself, or you'll be cut down to size". E.g. Japanese nails that stick up, tall poppies, and Jante. Isn't that a darker version of celebration of the everyman? Such cultures would be more suspicious of people trying to appear to be great leaders, and would encourage humility instead.
Even if you could hack the universe, how could you tell an unintended feature (a hack) from an intended feature (a law of physics) from the inside? Science is descriptive, not prescriptive...
What year would that Nobel Peace Prize have been awarded? I can't seem to find him anywhere on this list nor on this list (in case Wikipedia has been vandalized).
If someone invents an auto-bug-finder, won't every software company run it on their software before releasing it, knowing that if they don't, the malware creators will?
As a concrete example: suppose P != NP. Then no simulation can solve NP-hard problems in polynomial outside time, so most simulations will have the same constraint as the parent universe (where we reside) and also be unable to solve NP-hard problems in polynomial inside time.
The "hey, this universe looks like a simulation too!" argument would be: every simulation we make has P!=NP. Our universe has P!=NP... sure looks like what a simulation would be like. Obviously, every simulation's rules is heavily correlated with the rules of the parent universe. But that doesn't mean that the parent universe is also a simulation.
To take this example with respect to time: it might be that time is fuzzy in our universe. If so, it puts a constraint on how "sharp" time in a simulation can be as well without excessive penalties to runtime (like the poly inside time = exponential outside time example above for any simulation with P=NP). But if time is fuzzy inside and time is fuzzy outside, that doesn't mean that the outside is also a simulation, at least not unless you've controlled for the effect that every simulation in this universe will be greatly biased towards being like the universe.
Now, the outside (our universe) might seem to be like the inside (a simulation) in surprising ways, e.g. things seeming to be lazy-evaluated with how quantum mechanics weirdness works. But mathematical patterns might persist in seemingly disparate structures. So it may be, say, that it's easier to lazy-evaluate small systems (like simulations in RAM) because it's hard to change large amounts of information at once, and it's hard to change large amounts of information at once because that's how our universe works.
In a way, not taking this correlation into account when speculating about whether the universe is a simulation begs the question.
Higher up in the thread, someone mentioned the Butlerian Jihad. In light of the quote:
Once we get there, what reason is there to have most of us sweaty, non-machine-owning meatbags around?
we might as well ask what reason there would be for the completely automated AIs to have even the machine-owning meatbags around. What purpose do the Titans, err, the.1% serve once one of them invents an Omnius?
I'd imagine the most likely explanation to be that the statistical probability of event x happening in the future given that y has happened in the past is not the same as the probability of y happening in the future given that x happened in the past; but at very small scales, the difference is too small to see at any small time delta.
If so, try speeding up the video or yes, look at a higher level.
I think, and my thoughts cross the barrier into the synapses of the machine, just as the good doctor intended. But what I cannot shake, and what hints at things to come, is that thoughts cross back. In my dreams, the sensibility of the machine invades the periphery of my consciousness: dark, rigid, cold, alien. Evolution is at work here, but just what is evolving remains to be seen.
It's lethal in rats. Fatal familial insomnia is also pretty horrible, although one might argue that it's the prion that makes it lethal, not the sleep deprivation.
- Minimax type AI has spectacularly good micro, but sucks at macro. E.g. chess AIs, or see minimax used on RTSes - quote: " RTMM plays perfect short term micro-scale game, but plays a very bad high-level (long term) strategy..."
- UCT type AI has somewhere between consistently poor and consistently average play on both macro and micro, see e.g. Go programs prior to AlphaGo.
- Neural net AI has good macro but suck at micro. See AlphaGo: as long as it could do a death of a thousand cuts type play against Sedol, it won; but when Sedol forced it into a tactical trap in game 4, it failed badly. And from the TD-Gammon article on Wikipedia: "TD-Gammon's strengths and weaknesses were the opposite of symbolic artificial intelligence programs and most computer software in general: it was good at matters that require an intuitive "feel", but bad at systematic analysis." A problem with neural net AI is that it has to be designed to fit the problem. AlphaGo used a UCT hybrid with convolutional neural networks while TD-Gammon used temporal difference learning.
Humans are also much better at figuring out a game without being told how it works, or at constructing models for situations where trial and error is out of the question; this probably contributes to why we're not seeing fire-and-forget AI for say, governance or management. There's no rulebook for that kind of "game", and a neural net AI can't train itself by self-play unless it knows what the rules are to begin with.
An AI that can only optimize is not a very general AI. Optimize what? That's a matter of ethics, and that's just as hard, if not moreso, as the instrumental intelligence part.
Slashdot Mirror
People choose google, because they like it , not because of some monopoly influence.
People choose Google because they like it.
They like Google because Google provides good results.
Google provides good results because they have huge datacenters and extreme amounts of data.
The cost of acquiring huge datacenters and extreme amounts of data provides a barrier to entry.
This barrier to entry produces a natural monopoly.
And so it is. Not all monopolies have to be ordained by the State.
Those who rob Peter to pay Paul can always count on the support of Peter....
Why would Peter always support being robbed?
If the computer was infinitely fast, there would be two programming languages.
1. Nonlinear optimization.
2. Teach-by-example, e.g. a neural net of maximum size for the RAM constraints, where all the weights are found by brute force.
The former is for when you know exactly what you want, the latter is for when you don't and use supervised learning instead.
For that matter, an infinitely fast computer is a hypercomputer, so 2. could easily be "the smallest Turing machine that produces the desired outputs given the inputs".
That would be regret in the game theoretical sense. The regret of a strategy is the best payoff you could get minus the best you got; the "opportunity cost". There's an example here.
Since the market is based on individual predictors, the best it can do is somehow knowing the best predictor at each instant. That would correspond to zero regret. Any algorithm based on experts advice would have a regret greater than or equal to zero, even though it may not have an "emotional" regret.
Again perhaps surprisingly, play money vs real money doesn't seem to have much of an effect (See also this post).
If you consider the market to be a kind of weighted voting algorithm that exponentially amplifies the predictions of good experts (as you've said), then it doesn't greatly matter what the weights correspond to in the real world. All that matters is that good experts get their weights amplified, bad experts "go broke", that you have enough players to begin with to catch some good experts, and that the players can't do Sybil attacks to get around going broke.
As sites like StackOverflow show -- or MMOs for that matter -- there are plenty of people that are incentivized simply by getting a high score, even if that high score doesn't translate into real money. So while it may seem counterintuitive that play market should work, it's not that weird if you consider it from the point of view of an exponential weighting algorithm.
I might have been a bit hasty in my informal impossibility argument, as it were, but my line of thought was like this:
- The claim is that a prediction market can act as well as the best participant of that market, i.e. have zero regret.
- This should hold irrelvant of what the inputs are, whether the inputs are predictions by people or by other simpler algorithms.
- If the claim were true, you could dump a lot (and I mean an extreme amount) of comparatively simple algorithms into a virtual prediction market, and by the claim, the market would act just as well as the very best algorithm of the lot, producing a super AI.
- Since we don't have such super AIs in the real world, the claim must be wrong (and it is, because of the log(N) term).
More generally, if prediction markets were to have zero regret compared to the best expert, every kind of ensemble method would be easy. Just dump the individual methods into a prediction market and you'd get at least as good performance as the best individual method. The prediction market algorithm is the same irrespective of whether its inputs are by people or by computer algorithms, after all.
Perhaps surprisingly, it's often hard to do better than nonadaptive methods. See for instance the Variance method of An empirical comparison of algorithms for aggregating expert predictions. Then again, I have no idea what particular algorithm Swarm AI uses; it may be a bad nonadaptive method.
I imagine the reason they're illegal is that they can produce perverse incentives. Suppose, for instance, that there's a prediction market on when the next act of terrorism will occur in the US, as defined by some criterion. You predict tomorrow, then go (or get a dupe to go) snipe off some people and send a manifesto to the newspapers. Since the market considered tomorrow to be very unlikely, you get a lot of money doing this, and a strategy like that might seem rather tempting for someone who's poor or desperate enough.
I'd thus expect play money prediction markets to be more easily accepted, since there's no reason to murder for Shiny Points or whatever. But I wouldn't be surprised if the law just refers to prediction markets in general rather than real money prediction markets in particular.
No, because you're always going to take some hit determining what the optimal ensemble is (and even moreso if situations change or if there's noise). In exponential reweighting, the regret has an ln(N) term, which is what keeps you from just making an optimal AI by dumping an exponential number of algorithms into a virtual prediction market and getting a result "as good as the best underlying prediction method".
In practice, if you run the prediction market with real people, you're at risk of all the fun stuff that sometimes makes real markets behave poorly, such as bubbles. This happens because there's a feedback loop where people alter their predictions based on everybody else's predictions, as in a Keynesian beauty contest. In contrast, "the best underlying prediction method" would not have such problems because there would be no market for it to be influenced by.
30% of 0 is still 0.
A lot of cultures have social regulations of the form "don't think too much of yourself, or you'll be cut down to size". E.g. Japanese nails that stick up, tall poppies, and Jante. Isn't that a darker version of celebration of the everyman? Such cultures would be more suspicious of people trying to appear to be great leaders, and would encourage humility instead.
Even if you could hack the universe, how could you tell an unintended feature (a hack) from an intended feature (a law of physics) from the inside? Science is descriptive, not prescriptive...
What year would that Nobel Peace Prize have been awarded? I can't seem to find him anywhere on this list nor on this list (in case Wikipedia has been vandalized).
If someone invents an auto-bug-finder, won't every software company run it on their software before releasing it, knowing that if they don't, the malware creators will?
As a concrete example: suppose P != NP. Then no simulation can solve NP-hard problems in polynomial outside time, so most simulations will have the same constraint as the parent universe (where we reside) and also be unable to solve NP-hard problems in polynomial inside time.
The "hey, this universe looks like a simulation too!" argument would be: every simulation we make has P!=NP. Our universe has P!=NP... sure looks like what a simulation would be like. Obviously, every simulation's rules is heavily correlated with the rules of the parent universe. But that doesn't mean that the parent universe is also a simulation.
To take this example with respect to time: it might be that time is fuzzy in our universe. If so, it puts a constraint on how "sharp" time in a simulation can be as well without excessive penalties to runtime (like the poly inside time = exponential outside time example above for any simulation with P=NP). But if time is fuzzy inside and time is fuzzy outside, that doesn't mean that the outside is also a simulation, at least not unless you've controlled for the effect that every simulation in this universe will be greatly biased towards being like the universe.
Now, the outside (our universe) might seem to be like the inside (a simulation) in surprising ways, e.g. things seeming to be lazy-evaluated with how quantum mechanics weirdness works. But mathematical patterns might persist in seemingly disparate structures. So it may be, say, that it's easier to lazy-evaluate small systems (like simulations in RAM) because it's hard to change large amounts of information at once, and it's hard to change large amounts of information at once because that's how our universe works.
In a way, not taking this correlation into account when speculating about whether the universe is a simulation begs the question.
But their daemon is BitTorrent with blockchain and hookers! And blackjack! Who could possibly refuse?
Seriously, though, I wonder why they didn't just reupload the stuff to YouTube. If the videos are CC, it should be perfectly legal.
I suspect it's the other way around: simulations look like the universe because the simulations are confined to the universe.
we might as well ask what reason there would be for the completely automated AIs to have even the machine-owning meatbags around. What purpose do the Titans, err, the .1% serve once one of them invents an Omnius?
I'd imagine the most likely explanation to be that the statistical probability of event x happening in the future given that y has happened in the past is not the same as the probability of y happening in the future given that x happened in the past; but at very small scales, the difference is too small to see at any small time delta.
If so, try speeding up the video or yes, look at a higher level.
I think, and my thoughts cross the barrier into the synapses of the machine, just as the good doctor intended. But what I cannot shake, and what hints at things to come, is that thoughts cross back. In my dreams, the sensibility of the machine invades the periphery of my consciousness: dark, rigid, cold, alien. Evolution is at work here, but just what is evolving remains to be seen.
—Commissioner Pravin Lal,
“Man and Machine”
It's lethal in rats. Fatal familial insomnia is also pretty horrible, although one might argue that it's the prion that makes it lethal, not the sleep deprivation.
The way it seems to be for games AI is generally:
..."
- Minimax type AI has spectacularly good micro, but sucks at macro. E.g. chess AIs, or see minimax used on RTSes - quote: " RTMM plays perfect short term micro-scale game, but plays a very bad high-level (long term) strategy
- UCT type AI has somewhere between consistently poor and consistently average play on both macro and micro, see e.g. Go programs prior to AlphaGo.
- Neural net AI has good macro but suck at micro. See AlphaGo: as long as it could do a death of a thousand cuts type play against Sedol, it won; but when Sedol forced it into a tactical trap in game 4, it failed badly. And from the TD-Gammon article on Wikipedia: "TD-Gammon's strengths and weaknesses were the opposite of symbolic artificial intelligence programs and most computer software in general: it was good at matters that require an intuitive "feel", but bad at systematic analysis." A problem with neural net AI is that it has to be designed to fit the problem. AlphaGo used a UCT hybrid with convolutional neural networks while TD-Gammon used temporal difference learning.
Humans are also much better at figuring out a game without being told how it works, or at constructing models for situations where trial and error is out of the question; this probably contributes to why we're not seeing fire-and-forget AI for say, governance or management. There's no rulebook for that kind of "game", and a neural net AI can't train itself by self-play unless it knows what the rules are to begin with.
Then why did Minecraft succeed and SL fail? Was it just too easy to make cocks in SL? :)
An AI that can only optimize is not a very general AI. Optimize what? That's a matter of ethics, and that's just as hard, if not moreso, as the instrumental intelligence part.
This is called the AI effect.
Quoting the article, "as soon as AI successfully solves a problem, the problem is no longer a part of AI."
Or even with less of a focus on retribution. It may feel good to treat the prisoner like scum, but it's not a good strategy.