This thread sounds more fun now. Ok. But Ben better be watching. On Saturday, May 23, 2020, at 10:42 AM, Alan Grimes wrote: > Intelligence requires that there multiple possible futures, otherwise we would simply be mechanically unfolding a pre-determined destiny.
First I'll start with the obvious side of the coin. We do know our universe is at least somewhat predictable, that's why we can repeat the same lab experiments around Earth, and build neural models of the world to learn patterns. The laws of physics make our world at least partially unfold along a deterministic path. A computer simulation or calculator is also replay-able - it's predictable. We are able to understand things because they are not random. On the other hand, the word "random", by definition, means an outcome/result that varies maximally. So instead of a computer algorithm spitting out the number 5 predictably, you may get 1, then 8 next time, then 3, 0, 8, 6.... If it wasn't [maximally] random you'd get outputs like 6, 4, 5, 5, 6, 4, 4, 5. So random just means a wider variation. It doesn't mean it disobeys the laws of physics. It just means the view we look at something is ignoring fine details. For example, a woman can write down which color dresses she'll show you, so she knows the order of colors, but you don't, so to you it appears unpredictable, but to her, maybe the dress order is down pat and she remembers it like the alphabet. Another example, you fill a glass with milk until it pours over the rim, but the side that leaks first is different each time. Why? The glass is perfectly flat at the top let's say. It's because of the direction the human poured the milk in that caused it to pour over different sides. The human didn't stand in the same spot each time! The definition of random, that I gave here, basically equates to: you don't know something, so you output the wrong answer. But someone else can know what will happen! Lol. In other words, in the physics/laws we have, you can get an algorithm that outputs 5 each time, or outputs 4, 6, 5, 5, 6, 4, 4, or outputs 7, 1, 3, 0, 8, 4, 2, 8. And there's an actual reason behind it. Not magic. The definition of "random" I gave here, therefore, is when you lack information. You don't know what will occur. But once you learn what will occur, you know in the future what would result. This is assuming you can look in a computer or brain to see the stored algorithm. If you can't know what's inside, then you don't know if it will output 5 every time. So, do we have another definition for the word "random"? Yes. I call it True Random. It would need to break the laws of physics. For example, an atom or particle would be shot into space, travelling, and after 45 minutes, decides to change its direction! There's no reason it should have, though. Nothing touched the system and nothing left the system. Now, we already know our world is at least 50% not True Random, but predictable. And in computers there's a thing called redundancy that stops errors from popping up. You can run a car simulation perfectly each time, the same way each time! You could run a human simulation, with no True Randomness! Unless it makes us act the way we do. So, True Randomness may exist, and it may be helpful in making more robust predictors that handle uncertainty. You could just make your world/borg garage larger. Larger systems can avoid errors and damage more than brittle delicate small systems. It takes longer for the errors to show up. So the borgs could more easily predict where things are at the high level. Now, one could argue that if particles acted truly random 50% of the time, it would show up in computer car simulations! But it doesn't. So the real reason we get errors is because there is faults at low levels we don't know about. That's all. Not True Randomness. Now, can we solve this? Yes. We already are. Humans produce babes without the DNA information disappearing. We can repair cars indefinitely. But we can't know where every particle in our system is, for to do so would require knowing where the particles (that make up our knowing) are, which is impossible. You could make everything into solid cubes, but you still can't model your world perfectly, only approximately. The 4th side of the coin is magic orbs from God herself. Unfortunately, if you were hoping for this to be a valid thing, you are mistaken. Magic has no place, magic has to be either True Randomness, Randomness, or Laws of physics. There's no, such, thing, as magic. Either a particle moves as expected based on its and/or other surrounding context/conditions -OR- it pops into/outof existence some "move" or "particle" or "law" that truly is random. Say we had a genie ghost waving its hand with Free Will, granting wishes. The way it works is not by a existing predictive mechanism, but by popping into existence stuff, and must be non-random stuff. But why non-random? Because the genie would not exist, it'd be illogical soup. But what sort of "dimensional ether" is remembering or directing non-random creation in real time? We need something already existing to do this. A designer who creates a designer who... So it's impossible. On Saturday, May 23, 2020, at 10:42 AM, Alan Grimes wrote: > Your compression thingy will basically produce something that spews language, gibberish actually because there is no world model or understanding behind it. Much more importantly,there is no path to general problem solving, or even generalized language gibberish spewing, just a specific language. "Your compression thingy" This shows you lack understanding. Gosh. Lossless Compression is just an Evaluation for my neural net predictor I made. I could use Perplexity. Same algorithm, just different test of how good my algorithm is at predicting data in the distribution. "will basically produce something that spews language" "gibberish actually because there is no world model or understanding behind it." Again, you're lacking here. Neural networks learn a model of DATA. Be it text or vision. - Both are language. Which means they CAN learn PATTERNS. Patterns mean frequency, because in a dataset you may see the letter 'z' or word 'grommet' appears not too often! Maybe nothing re-occurs! Maybe the whole dataset is tttttttttt. So you can predict/generate the likely future, being the letter 'e' or word 'the'. Now, because of these re-occurring letters or words, words like cat & dog can be found to share the same contexts. Dogs eat, dogs jump, cats eat, cats jump. Thank god the word "jump" appears at least twice lol. Else no semantics! SO: A neural model can learn the letter 'e' appears very frequently, 'z' appears infrequently, 'cat' is very contextually similar to 'dog', and 'cat' is very different than 'jog'. Neural Models help organisms to survive longer in Evolution. Even if you don't believe text data mirrors human vision_thoughts data, you can still trust the algorithm can work on ANY dataset by "finding" patterns. In FACT, the Transformer architecture used in GPT-2, works on vision and music datasets. "or even generalized language gibberish spewing, just a specific language." First of all, the algorithm I already coded from scratch can predict the next letter of any language/ generate other languages too, like Hindi, French, etc. You just feed it such dataset and it learns the patterns. Currently I use enwik8. Now, my future algorithm, and the already existing GPT-2 made by OpenAI, can already learn cat=dog semantically by shared contexts, cat/dog are interchangeable and it can recognize unseen sentences. It helps it knows what entails a given word or phrase by looking at many many similar situations from past experience. As well, it can learn hello=bonjour, if it is fed diverse data that has enough French words! This works for vision too. And if you use text + vision you will need to associate them in the same time they were shown. "Much more importantly,there is no path to general problem solving" You've literally just asked me how to create AGI. AGI needs to solve many different types, of Hard Problems. To do so, it needs a large/diverse model, not just so it can solve various domains, but so it can use all sorts of domains when solving a problem in a given domain. It needs to know frequencies or IOW Cause > Effect probabilities of our physics (dogs usually breath, not eat) to logically think about paths it COULD take. And must take a path it desires too, to reach the desired outcome. It must wait at steps, until they are completed. It must update goals through induction/semantics. Food = money = jobs = truck = wrenches. It will ask new questions and seek new data from specialized sources or questions. It may need to search/mutate answers before mentally generates a good well-backed/aligned answer. It needs to be told when you look at 2+2=, it must be a precise answer, not 8, even though it kind of answers the question. It needs to be told when you are unsure of the prediction for 2+2=, you must look at it a different way or collect more specific data, if it is unsure about 573+481= it can look at it a different way (assuming you are sure of 2+2=4 etc etc). You are told to resort to look at [5]73[+][4]81[=] so all you hear is 5+4=9, to carry over numbers and stack the 4 results together (must hold onto them therefore) to get 573+481=1054. A good challenge is taking requirements and translating it to Python code. Basically AGI needs to look at CERTAIN context, hold onto them or forget them (ignore/Attention), combine data or probabilities, like a Turing Tape. To do AND, OR, NOR, requires enough energy to activate it binaryally yes/no. These nodes can be made in the brain, like rules, by talking to the AGI. AGI is basically a net holding onto energies, triggering semantics or syntactics, developing "rules" for when to fire nodes or what features to look for or look at ex. word or letter level [567] or [5]67. You could look at everything I just wrote and figure out why I typed it all. Maybe I'm a GPT-2? ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Td0702c568757f706-Mb70e25ae989645df45a19dbc Delivery options: https://agi.topicbox.com/groups/agi/subscription
