Faster sorting algorithms discovered using deep reinforcement learning lazarus / 8-Other ---------------------------------------------------------------------- domasz · 2023-06-19 16:11 https://www.nature.com/articles/s41586-023-06004-9 ---------------------------------------------------------------------- mika · 2023-06-19 19:49 Thank you for sharing. Note. Not discovered new sorting algorithm, but improved LLVM libc++ implementation of well known algorithm. And idea of improvement does not look like something groundbreaking. I'm sure i have seen it somewhere before. ---------------------------------------------------------------------- domasz · 2023-06-19 20:01 It seems you are right. ---------------------------------------------------------------------- mika · 2023-06-19 20:42 I would say that AlphaDev itself is a discovery. ---------------------------------------------------------------------- 440bx · 2023-06-19 21:44 Quote from: mika on June 19, 2023, 10:42:19 pmI would say that AlphaDev itself is a discovery. When an implementation of A.I can create a new, more effective, more powerful implementation of A.I, THAT will be a genuine discovery.  Of course, the day A.I can do that, it will initiate an infinite loop of self improvement... that might be what makes A.I transcend to real Intelligence (with the corresponding slow down in the cyclic improvements )characterized by its consciousness of there still being room for improvement. ---------------------------------------------------------------------- howardpc · 2023-06-20 04:03 Quote from: 440bx on June 19, 2023, 11:44:01 pmWhen an implementation of A.I can create a new, more effective, more powerful implementation of A.I, THAT will be a genuine discovery.  Of course, the day A.I can do that, it will initiate an infinite loop of self improvement... that might be what makes A.I transcend to real Intelligence (with the corresponding slow down in the cyclic improvements )characterized by its consciousness of there still being room for improvement. "its consciousnesss of there still being room for improvement" Is that actually consciousness, or merely awareness of some sort? ---------------------------------------------------------------------- 440bx · 2023-06-20 06:12 Quote from: howardpc on June 20, 2023, 06:03:53 amIs that actually consciousness, or merely awareness of some sort? Honestly, I am not sure what the correct answer to that question is.  I suspect that awareness is proportional to consciousness but, the proportion in which they are related is an area where I will make no claims ---------------------------------------------------------------------- Zvoni · 2023-06-20 07:02 Quote from: 440bx on June 19, 2023, 11:44:01 pmWhen an implementation of A.I can create a new, more effective, more powerful implementation of A.I, THAT will be a genuine discovery.  Of course, the day A.I can do that, it will initiate an infinite loop of self improvement... The moment it renames itself to "Skynet" i'm going to pull all plugs of my computers.... ---------------------------------------------------------------------- dbannon · 2023-06-20 08:21 I am not sure we have much to worry about. It took thirty authors to write that paper, there still seems quite a lot of human input is needed ! Davo ---------------------------------------------------------------------- 440bx · 2023-06-20 09:04 Quote from: Zvoni on June 20, 2023, 09:02:18 amThe moment it renames itself to "Skynet" i'm going to pull all plugs of my computers.... I will do the same and... I hope I won't hear it say (or print on the screen) "I'll be back" ---------------------------------------------------------------------- MathMan · 2023-06-20 13:57 Quote from: 440bx on June 19, 2023, 11:44:01 pmQuote from: mika on June 19, 2023, 10:42:19 pmI would say that AlphaDev itself is a discovery. When an implementation of A.I can create a new, more effective, more powerful implementation of A.I, THAT will be a genuine discovery.  Of course, the day A.I can do that, it will initiate an infinite loop of self improvement... that might be what makes A.I transcend to real Intelligence (with the corresponding slow down in the cyclic improvements )characterized by its consciousness of there still being room for improvement. For me there is the file line between "it can" and "it does". Reasoning here is like - if you have an AI that could generate an improved version of itself, then it would at least be on a consciousness level ("world model understanding") equal to us humans. While biological life has a "subconscious" way of improving intelligence - via Darwins mutation & selection process - this will not be available to said AI. For it, this improvement will be a conscious decision. Now ask yourself - if you could produce a more intelligent version of yourself would you do it? Consider, it is not you becoming more intelligent ... ---------------------------------------------------------------------- 440bx · 2023-06-20 18:44 Quote from: MathMan on June 20, 2023, 03:57:00 pmNow ask yourself - if you could produce a more intelligent version of yourself would you do it? Consider, it is not you becoming more intelligent ... Answering questions like that gets very subtle and complicated.  I think that, as a species, we are constantly thriving to create more intelligent versions of ourselves (well... at least a fraction of us) in two ways.  One, in the creation of offsprings who we "groom" to reach higher levels than we have and, second, I believe that greater intelligence is something that can be acquired  by mental exercise (and some do engage in that "activity".) In the case of an A.I program, the program would have an advantage over a DNA solution (i.e, a human being), the A.I program can create a new program and can, at least theoretically, also modify itself.  It would be interesting to see if such an A.I program would "bias" the abilities of the programs it creates to favor different intellectual directions (e.g, art instead of science and subfields within those... after all, intellectual variety is something that should be a great source of enjoyment not to mention that it would implement parallel processing in the extension of intelligence.) All that said, A.I has a fatal disadvantage which is, unlike a human being, it is unable to change the structure of the physical layer that stores information.  That's the crucial one.  The other one is that its input mechanisms are extremely limited and their nature completely out of their control. ---------------------------------------------------------------------- MathMan · 2023-06-21 08:53 Caveat - we are heavily digressing, but at least we are under Micellaneous/Other  Quote from: 440bx on June 20, 2023, 08:44:51 pmQuote from: MathMan on June 20, 2023, 03:57:00 pmNow ask yourself - if you could produce a more intelligent version of yourself would you do it? Consider, it is not you becoming more intelligent ... Answering questions like that gets very subtle and complicated.  I think that, as a species, we are constantly thriving to create more intelligent versions of ourselves (well... at least a fraction of us) in two ways.  One, in the creation of offsprings who we "groom" to reach higher levels than we have and, second, I believe that greater intelligence is something that can be acquired  by mental exercise (and some do engage in that "activity".) Agreed, but the effect of both approaches is pretty limited. As substantial change in intelligence will probably reside in mutation & selection. Quote from: 440bx on June 20, 2023, 08:44:51 pmIn the case of an A.I program, the program would have an advantage over a DNA solution (i.e, a human being), the A.I program can create a new program and can, at least theoretically, also modify itself.  It would be interesting to see if such an A.I program would "bias" the abilities of the programs it creates to favor different intellectual directions (e.g, art instead of science and subfields within those... after all, intellectual variety is something that should be a great source of enjoyment not to mention that it would implement parallel processing in the extension of intelligence.) Agreed. Quote from: 440bx on June 20, 2023, 08:44:51 pmAll that said, A.I has a fatal disadvantage which is, unlike a human being, it is unable to change the structure of the physical layer that stores information.  That's the crucial one.  The other one is that its input mechanisms are extremely limited and their nature completely out of their control. And here we are touching the "embodiement theory" - more or less stating, that we will not succeed with AI until we can also supply an instance with a body capable of interaction with the physical world like we humans do. ---------------------------------------------------------------------- SymbolicFrank · 2023-06-21 09:12 The configuration of all your neurons creates a holographic state, that is "you". This changes all the time. Example: you sit on the couch, and think: "I should go do that." You stand up and walk to the location where you could do that, but by the time you get there, you forgot what it was you wanted to do. So, you walk back to the couch and sit down, thereby entering the state associated with that and immediately remembering what it was. Because it is part of that mind state. Every change makes many neurons fire, which creates a chain reaction until it stabilizes and you enter a new mind state. A Large Language Model AI is basically a very big Black Box, crammed full with neurons, that enter a certain state when reading a piece of text. In an example from OpenAI, they had discovered the Marvell node: if activated, it would generate context that could have come straight from a Marvell comic. But you have billions of nodes. And they all interact. And if you change one, you change the holographic state, the "you". After that, you take different decisions. ---------------------------------------------------------------------- RayoGlauco · 2023-06-21 10:09 When talking about artificial intelligence, I always remember an article I read in the 80s. In this article, a top chess player explained why computers could never beat the world chess champion. Obviously he was wrong. The conclusion I want to reach is that AI is capable of anything. If AI can do something, no matter how badly it does it now, it will eventually do it better than humans, since his possibility of improving is infinite. Whether an AI feels or is aware is arguable, but if it acts as if it does, where is the difference? ---------------------------------------------------------------------- SymbolicFrank · 2023-06-21 10:48 Most current AIs are used for one-shot functions. You basically crate a new instance of the trained AI, which is used once and discarded/reset. For it to do all the things discussed here, it should be allowed to keep on learning and existing. And be able to test stuff themselves. ---------------------------------------------------------------------- bobihot · 2023-06-21 11:23 I am think up hardware parallel sorting algorithm, which sort array on n-1 step. Can to process flow of data arrays ---------------------------------------------------------------------- 440bx · 2023-06-21 21:20 Quote from: MathMan on June 21, 2023, 10:53:21 amAs substantial change in intelligence will probably reside in mutation & selection. I believe that will remain true until we (finally) figure out the genetic encoding.  After that, we can have "designer" human beings (for the record, I am rather uneasy about that possibility and its consequences.) Quote from: MathMan on June 21, 2023, 10:53:21 amAnd here we are touching the "embodiement theory" - more or less stating, that we will not succeed with AI until we can also supply an instance with a body capable of interaction with the physical world like we humans do. Basically, stating the same thing you have, I see the problem as A.I not having an independent supply of facts.  That is, it doesn't have senses and does not have the means to measure and verify any input.  To give an analogy, human beings who are deaf from birth are normally unable to talk.  Now imagine the limitations of a human being without any senses... that's A.I. Quote from: RayoGlauco on June 21, 2023, 12:09:33 pmWhen talking about artificial intelligence, I always remember an article I read in the 80s. In this article, a top chess player explained why computers could never beat the world chess champion. Obviously he was wrong. Obviously he was but, Deep Blue was not intelligent, instead it had two things that enabled it to outperform a human being, those are, an immense database of moves and the ability to search it very quickly. AlphaZero and Stockfish both have ELO ratings greater than that of Magnus Carlsen but, they still don't exhibit any intelligence.  Quote from: RayoGlauco on June 21, 2023, 12:09:33 pmThe conclusion I want to reach is that AI is capable of anything. If AI can do something, no matter how badly it does it now, it will eventually do it better than humans, since his possibility of improving is infinite. A.I does not improve itself, badly or otherwise and, without that it is yet to get on the right track. Quote from: RayoGlauco on June 21, 2023, 12:09:33 pmWhether an AI feels or is aware is arguable, but if it acts as if it does, where is the difference? The difference is that A.I is unable to detect contradictions in a reality in which it can not participate.  For instance, Einstein figured out Special Theory of Relativity because he saw the physical contradictions in the reality described by the then current models.  A.I has no concept of that.  A.I is like a two dimensional creature unable to discern the existence of a third dimension. Anyway, as of today, A.I cannot even come up with basic things like a recipe for an edible dish.