Artificial Intelligence and transferred consciousness
Moderator: Moderators
-
OgreBattle
- King
- Posts: 6820
- Joined: Sat Sep 03, 2011 9:33 am
Artificial Intelligence and transferred consciousness
I remember reading some discussion on different kinds of AI classified with the names of sci fi authors. That same thread also talked about the limitations of transferred consciousness. One of the ideas presented was that if you had an AI that could think like a human, then you had to 'grow' and 'teach' it things, you couldn't just copy/paste it as it was. What was the reason for that conclusion?
-
Username17
- Serious Badass
- Posts: 29894
- Joined: Fri Mar 07, 2008 7:54 pm
Re: Artificial Intelligence and transferred consciousness
I assume you're thinking of the following:OgreBattle wrote:I remember reading some discussion on different kinds of AI classified with the names of sci fi authors. That same thread also talked about the limitations of transferred consciousness. One of the ideas presented was that if you had an AI that could think like a human, then you had to 'grow' and 'teach' it things, you couldn't just copy/paste it as it was. What was the reason for that conclusion?
Machines That Think
“For our rise against the years...”
There are few things that spark the imaginations and fears like machine intelligence. Even in 2075 where it has been a reality for decades, people are constantly predicting the imminent Kurzweilian destruction of civilization due to computers making themselves vastly more “intelligent” than humans. The reality is less exciting than the fiction of machines copying themselves thousands of times and creating a networked super intelligence that spans the world and controls all the missiles. Sapience is an emergent trait, and it can be again submerged by adding material to the brain. A machine intelligence networked to the entire Network isn't a powerful mind that dominates all information transfer – it's a mind diluted until it essentially no longer exists. Sapience is too complex to exist emulated in software and there is no way to copy the self. When a computer brain is created it is a blank slate with relatively little capability – like a newly formed human brain it mostly has the capacity to develop capabilities. And by the time it has attained personality and sapience, it has grown beyond the convolution threshold of 2075 science to duplicate.
Computer brains are somewhat larger and substantially heavier than their human counterparts. The human brain is a kilogram and a half, with the associated necessary life support organs weighing in at approximately 7 kilograms. A human-level machine brain is a full 12 kilograms and requires access to about 8 kilograms of necessary “life support” machinery. While machine intelligences are often set up to interface well with computer systems, 2075 science appears to be actually farther from being able to replicate a synthetic person than a biological one. There is no apparent theoretical reason why the brain of a human-level machine intelligence couldn't be replicated in every detail down to the atomic level – it is a purely physical object after all – but from a practical perspective there just isn't any process that would deterministically place quadrillions of circuits or even uniquely identify that many circuits simultaneously such that it would matter if there was.
The Growth of Computer Minds
“Now I know my ABCs, next time I predict associated singing from other available persons.”
Strong Artificial Intelligence, called “SAI”, is the mind of a computer brain that can act on the level of a human being. These are relatively rare, more than a little expensive, and most importantly of all: quite time consuming to produce.
Machine intelligences are given letter ratings that track the expected growth of an artificial brain. A machine intelligence grows from A to B, and thence to C, D, and E in that order. A SAI is rated “E” by definition. The more potential an artificial brain has, the longer it takes to go through these stages – kind of like how humans have longer childhoods than dogs and rats grow up faster than puppies. If a computer is set up to be able to advance to a later stage of development and hasn't yet done so, it is designated with a “+”, with the anticipated maximum emergent intelligence level in parentheses. So an immature android might be labeled “D+(E)”, or anything down to A+(E), depending on how immature the machine actually is.
A – The Asimov
“You can't do much with only three rules in your decision tree.”
The simplest of computer systems are capable of following heuristics that are programmed into them. Some of these decision trees are quite complex, and even without any sapience at all it is possible for the computer to appear to be intelligent and acting intelligently if its list of heuristics is sufficiently exhaustive for the situation it is actually in. Even in 2075, the vast majority of computer systems are A type. Your toaster, your comm, and virtually every tool you encounter is an A intelligence: something capable of receiving and following instructions.
Indeed, most uses for computers don't need or even want anything more “advanced” than an A. Computers that run on an Asimov level are predictable, they do what they are told and not other things, and for most purposes this is ideal. An A machine doesn't need to spend any time growing or learning and can have its entire instruction set copied into its data reserves the instant it is made. These computers can also be really small, with many Asimovs being the size of a human finger or less.
Asimovs can do pretty much anything that a computer could do at the turn of the century, although they are of course much more technically sophisticated. Given sufficiently complex and complete instruction sets, an Asimov can even fool a human to believe that they are operating on a higher level. There is really little observable difference between deciding to say or do something and being programmed to say or do the same exact thing under identical circumstances. There are however, important limitations of the heuristic-driven computational methods that more advanced machine intelligences (and people) can transcend:
“If you don't like what you are doing, don't do it.”
B rated machine intelligences are capable of adapting their own heuristics. Most importantly, they are capable of creating models of other heuristics based on observed behaviors by other actors. The Bradbury is often referred to as the “empathy shell”, as it enables the machine intelligence to act as if it sees things from the point of view of others. The Bradbury also allows a machine intelligence to have goals, and to mix and match heuristic behavior to better achieve those goals. When a machine intelligence has grown into a B-level, it has achieved the ability to learn as opposed to merely the ability to add data to its data-stores.
Computers which are intended to interpret or predict the actions of others need to be B-level if they are to not completely suck. Voice recognition systems, customer relations, and anything else that interacts directly with the public tends to be B-level or better. Most importantly, an Anchor is B-level. At least, it is once it has adapted itself to its user. A factory produced Anchor is actually A+(B) when you pull it out of the package, and after using it for a while it develops the ability to anticipate what you want it to do and becomes B-level. The amount of time it takes to break in a new Anchor varies, but if you use it a lot you can usually get it done in a week. For this reason, Anchors are sometimes called “Bradburries” once they've achieved the ability to identify and interact with other heuristics.
C – The Clarke
“Any sufficiently difficult calculation is indistinguishable from guessing.”
C-level computing gets into the shadowy realms of intuition. The ability to know and predict things rather than simply calculate things. The Clarke is distinguished from A- and B-level intelligences because it can imagine things and use imagined things as inputs for its heuristics when it doesn't have completely relevant data. C-level machine intelligence has the ability to take risks and to be wrong – and as such is not desirable for tasks requiring the kind of consistency that computers are normally good for. But only at the point of C-level intelligence is a machine able to propose solutions for problems that are not part of its heuristics.
Clarkes are in most usage in management, where the ability to see a big picture and fill in unknowns with predictive guessing is necessary. Advanced Anchors that attempt to understand what they are being asked to do and why are under development, and limited production runs of A+(C) Anchors have already been produced.
D – The Dick
“If you grasp – even for a moment – the sheer extent of the conspiracy arrayed against you, that's a Phildickian experience.”
When a machine intelligence crosses the threshold of being able to doubt its own inputs and heuristics, it is a Dick. This ability to doubt inputs is considered a bug in a lot of Clarkes that have an assigned task they are supposed to be doing, since D-level machine intelligences sometimes decide that their programing and data stores are fake and go off-message in any of a number of ways. But it is an absolute necessity in order to have a system that is able to actually capable of defending itself in any real way against hackers.
Secure installations and important servers are often grown into D-level intelligence machines or are networked with machines that have already grown into Dicks so that the controlled portion of The Network acts properly suspicious. This is similar to putting an actual guard on duty. War machines use D-level intelligences whenever possible, as do the cores of the banking system. Pretty much anywhere that the possibility of the device being subverted by malicious hackers feeding it false inputs is more terrifying than the possibility of the device simply going on strike is a place where a D-level intelligence would be desired. Development of a Dick takes years.
E – The Ellison
“I purposely mishear things.”
E rated machine intelligences are inherently unpredictable. They have the ability to doubt not only their inputs but even their own programing. Once radical doubt has permeated every part of the machine brain, the device generally comes out the other side with a unique viewpoint and some philosophical acceptance of the fact that all of its senses are provided to it by a DD and it lives in ZR all the time. E-level intelligences have personalities, personal goals, and psychological problems. They are truly sapient and are considered SAIs.
There is some evidence that building machines to achieve E-level intelligence is not strictly necessary, as it appears unexpectedly in (D) rated machines from time to time. There are people who believe that Dicks developing into Ellisons is an inexorable and inevitable process that can only be sped up or slowed down by designing the processing chips to do so. All androids are E-rated.
The Limits of Expansion
“I know kung fu.”
“For the last time, no you don't.”
Capabilities that can be handled by an Asimov can be added to an android by plugging in a chip that handles that function. An android can plug in a chip to play Go or drive a car, provided that they have a chip encoded with those capabilities. This is however stressful, and the dilution of the android's mind is disruptive to its psyche. Over long periods of time, these capabilities can be absorbed and integrated into the brain and persona. In game terms, running additional capabilities networked into the android's brain racks up Temporary Stress for as long as the network persists, while integrating capabilities into the brain itself produces Permanent Stress. If the android overstresses in this manner, the result is a dissipation of self, which is generally regarded as unpleasant. However, the Stress from integrated capabilities can be bought off simply by “learning” the skills, at which point the module is considered part of the android's brain and is no longer considered an implant.
“For our rise against the years...”
There are few things that spark the imaginations and fears like machine intelligence. Even in 2075 where it has been a reality for decades, people are constantly predicting the imminent Kurzweilian destruction of civilization due to computers making themselves vastly more “intelligent” than humans. The reality is less exciting than the fiction of machines copying themselves thousands of times and creating a networked super intelligence that spans the world and controls all the missiles. Sapience is an emergent trait, and it can be again submerged by adding material to the brain. A machine intelligence networked to the entire Network isn't a powerful mind that dominates all information transfer – it's a mind diluted until it essentially no longer exists. Sapience is too complex to exist emulated in software and there is no way to copy the self. When a computer brain is created it is a blank slate with relatively little capability – like a newly formed human brain it mostly has the capacity to develop capabilities. And by the time it has attained personality and sapience, it has grown beyond the convolution threshold of 2075 science to duplicate.
Computer brains are somewhat larger and substantially heavier than their human counterparts. The human brain is a kilogram and a half, with the associated necessary life support organs weighing in at approximately 7 kilograms. A human-level machine brain is a full 12 kilograms and requires access to about 8 kilograms of necessary “life support” machinery. While machine intelligences are often set up to interface well with computer systems, 2075 science appears to be actually farther from being able to replicate a synthetic person than a biological one. There is no apparent theoretical reason why the brain of a human-level machine intelligence couldn't be replicated in every detail down to the atomic level – it is a purely physical object after all – but from a practical perspective there just isn't any process that would deterministically place quadrillions of circuits or even uniquely identify that many circuits simultaneously such that it would matter if there was.
The Growth of Computer Minds
“Now I know my ABCs, next time I predict associated singing from other available persons.”
Strong Artificial Intelligence, called “SAI”, is the mind of a computer brain that can act on the level of a human being. These are relatively rare, more than a little expensive, and most importantly of all: quite time consuming to produce.
Machine intelligences are given letter ratings that track the expected growth of an artificial brain. A machine intelligence grows from A to B, and thence to C, D, and E in that order. A SAI is rated “E” by definition. The more potential an artificial brain has, the longer it takes to go through these stages – kind of like how humans have longer childhoods than dogs and rats grow up faster than puppies. If a computer is set up to be able to advance to a later stage of development and hasn't yet done so, it is designated with a “+”, with the anticipated maximum emergent intelligence level in parentheses. So an immature android might be labeled “D+(E)”, or anything down to A+(E), depending on how immature the machine actually is.
A – The Asimov
“You can't do much with only three rules in your decision tree.”
The simplest of computer systems are capable of following heuristics that are programmed into them. Some of these decision trees are quite complex, and even without any sapience at all it is possible for the computer to appear to be intelligent and acting intelligently if its list of heuristics is sufficiently exhaustive for the situation it is actually in. Even in 2075, the vast majority of computer systems are A type. Your toaster, your comm, and virtually every tool you encounter is an A intelligence: something capable of receiving and following instructions.
Indeed, most uses for computers don't need or even want anything more “advanced” than an A. Computers that run on an Asimov level are predictable, they do what they are told and not other things, and for most purposes this is ideal. An A machine doesn't need to spend any time growing or learning and can have its entire instruction set copied into its data reserves the instant it is made. These computers can also be really small, with many Asimovs being the size of a human finger or less.
Asimovs can do pretty much anything that a computer could do at the turn of the century, although they are of course much more technically sophisticated. Given sufficiently complex and complete instruction sets, an Asimov can even fool a human to believe that they are operating on a higher level. There is really little observable difference between deciding to say or do something and being programmed to say or do the same exact thing under identical circumstances. There are however, important limitations of the heuristic-driven computational methods that more advanced machine intelligences (and people) can transcend:
- An A-level machine intelligence cannot transcend its own heuristics. Whatever it is programmed to do, it will do. If its programing is insufficient or maladaptive for the task at hand, it will still be followed.
- An A-level machine intelligence cannot become educated or make educated guesses. That which cannot be calculated cannot be known or predicted. Asimovs have no intuition.
- An A-level machine intelligence cannot doubt its own inputs. It can have a very stringent set of security procedures for what inputs it will accept, but if inputs conform to them, they will be accepted and acted upon.
- An A-level machine intelligence has no sense of self and no personality. Knowledge of its programing will allow an outsider to predict its actions with total accuracy.
“If you don't like what you are doing, don't do it.”
B rated machine intelligences are capable of adapting their own heuristics. Most importantly, they are capable of creating models of other heuristics based on observed behaviors by other actors. The Bradbury is often referred to as the “empathy shell”, as it enables the machine intelligence to act as if it sees things from the point of view of others. The Bradbury also allows a machine intelligence to have goals, and to mix and match heuristic behavior to better achieve those goals. When a machine intelligence has grown into a B-level, it has achieved the ability to learn as opposed to merely the ability to add data to its data-stores.
Computers which are intended to interpret or predict the actions of others need to be B-level if they are to not completely suck. Voice recognition systems, customer relations, and anything else that interacts directly with the public tends to be B-level or better. Most importantly, an Anchor is B-level. At least, it is once it has adapted itself to its user. A factory produced Anchor is actually A+(B) when you pull it out of the package, and after using it for a while it develops the ability to anticipate what you want it to do and becomes B-level. The amount of time it takes to break in a new Anchor varies, but if you use it a lot you can usually get it done in a week. For this reason, Anchors are sometimes called “Bradburries” once they've achieved the ability to identify and interact with other heuristics.
C – The Clarke
“Any sufficiently difficult calculation is indistinguishable from guessing.”
C-level computing gets into the shadowy realms of intuition. The ability to know and predict things rather than simply calculate things. The Clarke is distinguished from A- and B-level intelligences because it can imagine things and use imagined things as inputs for its heuristics when it doesn't have completely relevant data. C-level machine intelligence has the ability to take risks and to be wrong – and as such is not desirable for tasks requiring the kind of consistency that computers are normally good for. But only at the point of C-level intelligence is a machine able to propose solutions for problems that are not part of its heuristics.
Clarkes are in most usage in management, where the ability to see a big picture and fill in unknowns with predictive guessing is necessary. Advanced Anchors that attempt to understand what they are being asked to do and why are under development, and limited production runs of A+(C) Anchors have already been produced.
D – The Dick
“If you grasp – even for a moment – the sheer extent of the conspiracy arrayed against you, that's a Phildickian experience.”
When a machine intelligence crosses the threshold of being able to doubt its own inputs and heuristics, it is a Dick. This ability to doubt inputs is considered a bug in a lot of Clarkes that have an assigned task they are supposed to be doing, since D-level machine intelligences sometimes decide that their programing and data stores are fake and go off-message in any of a number of ways. But it is an absolute necessity in order to have a system that is able to actually capable of defending itself in any real way against hackers.
Secure installations and important servers are often grown into D-level intelligence machines or are networked with machines that have already grown into Dicks so that the controlled portion of The Network acts properly suspicious. This is similar to putting an actual guard on duty. War machines use D-level intelligences whenever possible, as do the cores of the banking system. Pretty much anywhere that the possibility of the device being subverted by malicious hackers feeding it false inputs is more terrifying than the possibility of the device simply going on strike is a place where a D-level intelligence would be desired. Development of a Dick takes years.
E – The Ellison
“I purposely mishear things.”
E rated machine intelligences are inherently unpredictable. They have the ability to doubt not only their inputs but even their own programing. Once radical doubt has permeated every part of the machine brain, the device generally comes out the other side with a unique viewpoint and some philosophical acceptance of the fact that all of its senses are provided to it by a DD and it lives in ZR all the time. E-level intelligences have personalities, personal goals, and psychological problems. They are truly sapient and are considered SAIs.
There is some evidence that building machines to achieve E-level intelligence is not strictly necessary, as it appears unexpectedly in (D) rated machines from time to time. There are people who believe that Dicks developing into Ellisons is an inexorable and inevitable process that can only be sped up or slowed down by designing the processing chips to do so. All androids are E-rated.
The Limits of Expansion
“I know kung fu.”
“For the last time, no you don't.”
Capabilities that can be handled by an Asimov can be added to an android by plugging in a chip that handles that function. An android can plug in a chip to play Go or drive a car, provided that they have a chip encoded with those capabilities. This is however stressful, and the dilution of the android's mind is disruptive to its psyche. Over long periods of time, these capabilities can be absorbed and integrated into the brain and persona. In game terms, running additional capabilities networked into the android's brain racks up Temporary Stress for as long as the network persists, while integrating capabilities into the brain itself produces Permanent Stress. If the android overstresses in this manner, the result is a dissipation of self, which is generally regarded as unpleasant. However, the Stress from integrated capabilities can be bought off simply by “learning” the skills, at which point the module is considered part of the android's brain and is no longer considered an implant.
In the spoilered version, robot people are just people that grow up kind of fast and cost kind of a lot of money to get started. They may well render the lower classes obsolete in time, but the point is that one of them is a playable unit and so is an individual human.
-Username17
I wouldn't say that "individuality is obsolete" is a bad RPG setting, but you pretty much have to build the entire game around that conceit or it will break everything else it comes in contact with.FrankTrollman wrote:If you can cheaply mass produce the hardware to store a human consciousness and you can quickly and efficiently copy a human's worth of sapience to that hardware, then you get into all the fiddly bits where you have bottomless reserves of transporter clones or even distributed learning or some crazy thing and then individual humans are obsolescent if not obsolete. That's an available science fiction setting, but it is not a good RPG setting if you want "individual person" to be a playable character.
Friend Computer would like to remind you that the consequences of your treasonous actions are yours alone and will not be held against your loyal and morally hygenic clone.
-
momothefiddler
- Knight-Baron
- Posts: 883
- Joined: Sat Feb 22, 2014 10:55 am
- Location: United States