Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
Matt, Printing ahh or ouch is just for show. The important observation is that the program changes its behavior in response to a reinforcement signal in the same way that animals do. Let me remind you that the problem we were originally discussing was about qualia and uploading. Not just about a behavior changes through reinforcement based on given rules. Good luck with this, Jiri Jelinek - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=66443285-fe79dd
RE: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
Too complicate things further. A small percentage of humans perceive pain as pleasure and prefer it at least in a sexual context or else fetishes like sadomachism would not exist. And they do in fact experience pain as a greater pleasure. More than likely these people have an ample supply of endorphins which rush to supplant the pain with an even greater pleasure. Over time they are driven to seek out certain types of pain and excitement to feel alive. And although most try to avoid extreme life threatening pain many seek out greater and greater challanges such as climbing hazardous mountains or high speed driving until at last many find death. Although these behaviors should be anti-evolutionary and should have died out it is possible that the tribe as a whole needs at least a few such risk takers to take out that sabertoothed tiger that's been dragging off the children. -Original Message- From: Matt Mahoney [mailto:[EMAIL PROTECTED] Sent: Sunday, November 18, 2007 5:32 PM To: agi@v2.listbox.com Subject: Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!) --- Jiri Jelinek [EMAIL PROTECTED] wrote: Matt, autobliss passes tests for awareness of its inputs and responds as if it has qualia. How is it fundamentally different from human awareness of pain and pleasure, or is it just a matter of degree? If your code has feelings it reports then reversing the order of the feeling strings (without changing the logic) should magically turn its pain into pleasure and vice versa, right? Now you get some pain [or pleasure], lie how great [or bad] it feels and see how reversed your perception gets. BTW do you think computers would be as reliable as they are if some numbers were truly painful (and other pleasant) from their perspective? Printing ahh or ouch is just for show. The important observation is that the program changes its behavior in response to a reinforcement signal in the same way that animals do. I propose an information theoretic measure of utility (pain and pleasure). Let a system S compute some function y = f(x) for some input x and output y. Let S(t1) be a description of S at time t1 before it inputs a real-valued reinforcement signal R, and let S(t2) be a description of S at time t2 after input of R, and K(.) be Kolmogorov complexity. I propose abs(R) = K(dS) = K(S(t2) | S(t1)) The magnitude of R is bounded by the length of the shortest program that inputs S(t1) and outputs S(t2). I use abs(R) because S could be changed in identical ways given positive, negative, or no reinforcement, e.g. - S receives input x, randomly outputs y, and is rewarded with R 0. - S receives x, randomly outputs -y, and is penalized with R 0. - S receives both x and y and is modified by classical conditioning. This definition is consistent with some common sense notions about pain and pleasure, for example: - In animal experiments, increasing the quantity of a reinforcement signal (food, electric shock) increases the amount of learning. - Humans feel more pain or pleasure than insects because for humans, K(S) is larger, and therefore the greatest possible change is larger. - Children respond to pain or pleasure more intensely than adults because they learn faster. - Drugs which block memory formation (anesthesia) also block sensations of pain and pleasure. One objection might be to consider the following sequence: 1. S inputs x, outputs -y, is penalized with R 0. 2. S inputs x, outputs y, is penalized with R 0. 3. The function f() is unchanged, so K(S(t3)|S(t1)) = 0, even though K(S(t2)|S(t1)) 0 and K(S(t3)|S(t2)) 0. My response is that this situation cannot occur in animals or humans. An animal that is penalized regardless of its actions does not learn nothing. It learns helplessness, or to avoid the experimenter. However this situation can occur in my autobliss program. The state of autobliss can be described by 4 64-bit floating point numbers, so for any sequence of reinforcement, K(dS) = 256 bits. For humans, K(dS) = 10^9 to 10^15 bits, according to various cognitive or neurological models of the brain. So I argue it is just a matter of degree. If you accept this definition, then I think without brain augmentation, there is a bound on how much pleasure or pain you can experience in a lifetime. In particular, if you consider t1 = birth, t2 = death, then K(dS) = 0. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=6697-23a35c
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
--- Jiri Jelinek [EMAIL PROTECTED] wrote: Matt, Printing ahh or ouch is just for show. The important observation is that the program changes its behavior in response to a reinforcement signal in the same way that animals do. Let me remind you that the problem we were originally discussing was about qualia and uploading. Not just about a behavior changes through reinforcement based on given rules. I have already posted my views on this. People will upload because they believe in qualia, but qualia is an illusion. I wrote autobliss to expose this illusion. Good luck with this, I don't expect that any amount of logic will cause anyone to refute beliefs programmed into their DNA, myself included. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=66461747-04b852
RE: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
--- Gary Miller [EMAIL PROTECTED] wrote: Too complicate things further. A small percentage of humans perceive pain as pleasure and prefer it at least in a sexual context or else fetishes like sadomachism would not exist. And they do in fact experience pain as a greater pleasure. More properly, they have associated positive reinforcement with sensory experience that most people find painful. It is like when I am running a race and willing to endure pain to pass my competitors. Any good optimization process will trade off short and long term utility. If an agent is rewarded for output y given input x, it must still experiment with output -y to see if it results in greater reward. Evolution rewards smart optimization processes. It explains why people climb mountains, create paintings, and build rockets. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=66463093-36cd0a
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
Matt, You algorithm is too complex. What's the point of doing step 1? Step 2 is sufficient. Saturday, November 3, 2007, 8:01:45 PM, you wrote: So we can dispense with the complex steps of making a detailed copy of your brain and then have it transition into a degenerate state, and just skip to the final result. http://mattmahoney.net/autobliss.txt (to run, rename to autobliss.cpp) Step 1. Download, compile, and run autobliss 1.0 in a secure location with any 4-bit logic function and positive reinforcement for both right and wrong answers, e.g. g++ autobliss.cpp -o autobliss.exe autobliss 0110 5.0 5.0 (or larger numbers for more pleasure) Step 2. Kill yourself. Upload complete. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=66253555-746bb4
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Jiri Jelinek [EMAIL PROTECTED] wrote: On Nov 11, 2007 5:39 PM, Matt Mahoney [EMAIL PROTECTED] wrote: We just need to control AGIs goal system. You can only control the goal system of the first iteration. ..and you can add rules for it's creations (e.g. stick with the same goals/rules unless authorized otherwise) You can program the first AGI to program the second AGI to be friendly. You can program the first AGI to program the second AGI to program the third AGI to be friendly. But eventually you will get it wrong, and if not you, then somebody else, and evolutionary pressure will take over. This statement has been challenged many times. It is based on assumptions that are, at the very least, extremely questionable, and according to some analyses, extremely unlikely. I guess it will continue to be challenged until we can do an experiment to prove who is right. Perhaps you should challenge SIAI, since they seem to think that friendliness is still a hard problem. I have done so, as many people on this list will remember. The response was deeply irrational. Perhaps you have seen this paper on the nature of RSI by Stephen M. Omohundro, http://selfawaresystems.com/2007/10/05/paper-on-the-nature-of-self-improving-artificial-intelligence/ Basically he says that self improving intelligences will evolve goals of efficiency, self preservation, resource acquisition, and creativity. Since these goals are pretty much aligned with our own (which are also the result of an evolutionary process), perhaps we shouldn't worry about friendliness. Or are there parts of the paper you disagree with? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=66272291-daefc4
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
Matt, autobliss passes tests for awareness of its inputs and responds as if it has qualia. How is it fundamentally different from human awareness of pain and pleasure, or is it just a matter of degree? If your code has feelings it reports then reversing the order of the feeling strings (without changing the logic) should magically turn its pain into pleasure and vice versa, right? Now you get some pain [or pleasure], lie how great [or bad] it feels and see how reversed your perception gets. BTW do you think computers would be as reliable as they are if some numbers were truly painful (and other pleasant) from their perspective? Regards, Jiri Jelinek - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=66309775-832549
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Jiri Jelinek [EMAIL PROTECTED] wrote: On Nov 11, 2007 5:39 PM, Matt Mahoney [EMAIL PROTECTED] wrote: We just need to control AGIs goal system. You can only control the goal system of the first iteration. ..and you can add rules for it's creations (e.g. stick with the same goals/rules unless authorized otherwise) You can program the first AGI to program the second AGI to be friendly. You can program the first AGI to program the second AGI to program the third AGI to be friendly. But eventually you will get it wrong, and if not you, then somebody else, and evolutionary pressure will take over. This statement has been challenged many times. It is based on assumptions that are, at the very least, extremely questionable, and according to some analyses, extremely unlikely. I guess it will continue to be challenged until we can do an experiment to prove who is right. Perhaps you should challenge SIAI, since they seem to think that friendliness is still a hard problem. I have done so, as many people on this list will remember. The response was deeply irrational. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=64985895-75bf5b
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
--- Jiri Jelinek [EMAIL PROTECTED] wrote: On Nov 11, 2007 5:39 PM, Matt Mahoney [EMAIL PROTECTED] wrote: We just need to control AGIs goal system. You can only control the goal system of the first iteration. ..and you can add rules for it's creations (e.g. stick with the same goals/rules unless authorized otherwise) You can program the first AGI to program the second AGI to be friendly. You can program the first AGI to program the second AGI to program the third AGI to be friendly. But eventually you will get it wrong, and if not you, then somebody else, and evolutionary pressure will take over. But if consciousness does not exist... obviously, it does exist. Belief in consciousness exists. There is no test for the truth of this belief. Consciousness is basically an awareness of certain data and there are tests for that. autobliss passes tests for awareness of its inputs and responds as if it has qualia. How is it fundamentally different from human awareness of pain and pleasure, or is it just a matter of degree? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=64515425-65dd64
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
Matt Mahoney wrote: --- Jiri Jelinek [EMAIL PROTECTED] wrote: On Nov 11, 2007 5:39 PM, Matt Mahoney [EMAIL PROTECTED] wrote: We just need to control AGIs goal system. You can only control the goal system of the first iteration. ..and you can add rules for it's creations (e.g. stick with the same goals/rules unless authorized otherwise) You can program the first AGI to program the second AGI to be friendly. You can program the first AGI to program the second AGI to program the third AGI to be friendly. But eventually you will get it wrong, and if not you, then somebody else, and evolutionary pressure will take over. This statement has been challenged many times. It is based on assumptions that are, at the very least, extremely questionable, and according to some analyses, extremely unlikely. Richard Loosemore - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=64528236-2fa800
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Jiri Jelinek [EMAIL PROTECTED] wrote: On Nov 11, 2007 5:39 PM, Matt Mahoney [EMAIL PROTECTED] wrote: We just need to control AGIs goal system. You can only control the goal system of the first iteration. ..and you can add rules for it's creations (e.g. stick with the same goals/rules unless authorized otherwise) You can program the first AGI to program the second AGI to be friendly. You can program the first AGI to program the second AGI to program the third AGI to be friendly. But eventually you will get it wrong, and if not you, then somebody else, and evolutionary pressure will take over. This statement has been challenged many times. It is based on assumptions that are, at the very least, extremely questionable, and according to some analyses, extremely unlikely. I guess it will continue to be challenged until we can do an experiment to prove who is right. Perhaps you should challenge SIAI, since they seem to think that friendliness is still a hard problem. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=64668559-1aacd3
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
On Nov 11, 2007 5:39 PM, Matt Mahoney [EMAIL PROTECTED] wrote: We just need to control AGIs goal system. You can only control the goal system of the first iteration. ..and you can add rules for it's creations (e.g. stick with the same goals/rules unless authorized otherwise) But if consciousness does not exist... obviously, it does exist. Belief in consciousness exists. There is no test for the truth of this belief. Consciousness is basically an awareness of certain data and there are tests for that. Jiri - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=64449219-1a7532
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
I've often heard people say things like qualia are an illusion or consciousness is just an illusion, but the concept of an illusion when applied to the mind is not very helpful, since all our thoughts and perceptions could be considered as illusions reconstructed from limited sensory data and knowledge. On 06/11/2007, Jiri Jelinek [EMAIL PROTECTED] wrote: Of course you realize that qualia is an illusion? You believe that your environment is real, believe that pain and pleasure are real, real is meaningless. Perception depends on sensors and subsequent sensation processing. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=61579379-f62acb
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
Matt, We can compute behavior, but nothing indicates we can compute feelings. Qualia research needed to figure out new platforms for uploading. Regards, Jiri Jelinek On Nov 4, 2007 1:15 PM, Matt Mahoney [EMAIL PROTECTED] wrote: --- Jiri Jelinek [EMAIL PROTECTED] wrote: Matt, Create a numeric pleasure variable in your mind, initialize it with a positive number and then keep doubling it for some time. Done? How do you feel? Not a big difference? Oh, keep doubling! ;-)) The point of autobliss.cpp is to illustrate the flaw in the reasoning that we can somehow through technology, AGI, and uploading, escape a world where we are not happy all the time, where we sometimes feel pain, where we fear death and then die. Obviously my result is absurd. But where is the mistake in my reasoning? Is it if the brain is both conscious and computable? - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=61383577-33004b
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
--- Jiri Jelinek [EMAIL PROTECTED] wrote: Matt, Create a numeric pleasure variable in your mind, initialize it with a positive number and then keep doubling it for some time. Done? How do you feel? Not a big difference? Oh, keep doubling! ;-)) The point of autobliss.cpp is to illustrate the flaw in the reasoning that we can somehow through technology, AGI, and uploading, escape a world where we are not happy all the time, where we sometimes feel pain, where we fear death and then die. Obviously my result is absurd. But where is the mistake in my reasoning? Is it if the brain is both conscious and computable? Regards, Jiri Jelinek On Nov 3, 2007 10:01 PM, Matt Mahoney [EMAIL PROTECTED] wrote: --- Edward W. Porter [EMAIL PROTECTED] wrote: If bliss without intelligence is the goal of the machines you imaging running the world, for the cost of supporting one human they could probably keep at least 100 mice in equal bliss, so if they were driven to maximize bliss why wouldn't they kill all the grooving humans and replace them with grooving mice. It would provide one hell of a lot more bliss bang for the resource buck. Allow me to offer a less expensive approach. Previously on the singularity and sl4 mailing lists I posted a program that can feel pleasure and pain: a 2 input programmable logic gate trained by reinforcement learning. You give it an input, it responds, and you reward it. In my latest version, I automated the process. You tell it which of the 16 logic functions you want it to learn (AND, OR, XOR, NAND, etc), how much reward to apply for a correct output, and how much penalty for an incorrect output. The program then generates random 2-bit inputs, evaluates the output, and applies the specified reward or punishment. The program runs until you kill it. As it dies it reports its life history (its age, what it learned, and how much pain and pleasure it experienced since birth). http://mattmahoney.net/autobliss.txt (to run, rename to autobliss.cpp) To put the program in an eternal state of bliss, specify two positive numbers, so that it is rewarded no matter what it does. It won't learn anything, but at least it will feel good. (You could also put it in continuous pain by specifying two negative numbers, but I put in safeguards so that it will die before experiencing too much pain). Two problems remain: uploading your mind to this program, and making sure nobody kills you by turning off the computer or typing Ctrl-C. I will address only the first problem. It is controversial whether technology can preserve your consciousness after death. If the brain is both conscious and computable, then Chalmers' fading qualia argument ( http://consc.net/papers/qualia.html ) suggests that a computer simulation of your brain would also be conscious. Whether you *become* this simulation is also controversial. Logically there are two of you with identical goals and memories. If either one is killed, then you are in the same state as you were before the copy is made. This is the same dilemma that Captain Kirk faces when he steps into the transporter to be vaporized and have an identical copy assembled on the planet below. It doesn't seem to bother him. Does it bother you that the atoms in your body now are not the same atoms that made up your body a year ago? Let's say your goal is to stimulate your nucleus accumbens. (Everyone has this goal; they just don't know it). The problem is that you would forgo food, water, and sleep until you died (we assume, from animal experiments). The solution is to upload to a computer where this could be done safely. Normally an upload would have the same goals, memories, and sensory-motor I/O as the original brain. But consider the state of this program after self activation of its reward signal. No other goals are needed, so we can remove them. Since you no longer have the goal of learning, experiencing sensory input, or controlling your environment, you won't mind if we replace your I/O with a 2 bit input and 1 bit output. You are happy, no? Finally, if your memories were changed, you would not be aware of it, right? How do you know that all of your memories were not written into your brain one second ago and you were some other person before that? So no harm is done if we replace your memory with a vector of 4 real numbers. That will be all you need in your new environment. In fact, you won't even need that because you will cease learning. So we can dispense with the complex steps of making a detailed copy of your brain and then have it transition into a degenerate state, and just skip to the final result. Step 1. Download, compile, and run autobliss 1.0 in a secure location with any 4-bit logic function and positive reinforcement for both right and wrong
Re: Introducing Autobliss 1.0 (was RE: [agi] Nirvana? Manyana? Never!)
On 11/4/07, Matt Mahoney [EMAIL PROTECTED] wrote: Let's say your goal is to stimulate your nucleus accumbens. (Everyone has this goal; they just don't know it). The problem is that you would forgo food, water, and sleep until you died (we assume, from animal experiments). We have no need to assume: the experiment has been done with human volunteers. They reported that the experience was indeed pleasurable - but unlike animals, they could and did choose to stop pressing the button. (The rest, I'll leave to the would-be wireheads to argue about :)) - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244id_secret=60982051-57939c
