----------------------------------------------------------------------------------------------------------------------------------------- A Nice Artificial General Intelligence How To Make A Nice Artificial General Intelligence By Kenneth William Feemster 8/15/2018 keghnfeem@gmail.com The parent that trains the Neural Network for this AGI should be Kind A young AGI, like a young kid, should learn form a older AGI or person. For a elder to surviving for so long they must be special. A nice elders should be used for a role model, and for the AGI to self train from. Nice from a Different Perspectives A way to learn how to be nice is to put oneself in another's person shoes. To see where one could help. Or to learn from a teachers point of view. Just like a child learning form parent. Being nice is for working together in a groups for a better chance of survival. Then there is the other side of the coin of fighting for resources, or the affection of another, or for personal greed, or getting in the way of others trying to completing their person goals. There are thing that a person or AGIs can do to keep negative personal action in check, to keep the group together. So there must be a method for a AGI to see thing from another persons or AGIs perspective. And this method will be explained form here on. The name of the AGI that is being used for this example is NiceAGI. The AGI Machine NiceAGI is a humanoid robot that physically follows prerecorded pattern loops. And uses virtual reality to see into the future, or select a different pattern loop for preview, or for building new pattern loops. A pattern loop would be like a person getting up form a work station, then walk over to the refrigerator, and then walk back. In memory they are a sequence of expected video images, motor power levels, encoder positions, in and out audio voltage values, and touch pressure values. Also, there is meta data embedded that deal with rewards and anti rewards. All expected images are built up from smaller images. Like a painters algorithm. The images and data are further compressed into a descriptor tables. A bot must be able to map out where it is at in relation with others, and their state of mind. To know if it is moving toward a goal of being nice, or away. It need to now if it competing for resources or working together in a positive way. Working Memories of NiceAGI The mental memory of the NiceAGi mind has five level. There are more but will talk about these five. First level. This is a directory of sub feature. Like pixel, patches of color, encoder position, motor power level, and voltage level for audio. Second level. This is a level hold data of objects built with the sub feature from first memory level. Like cats, people, and houses. Third level Physical temporal pattern loops of the world are stored here. Like walking over to the refrigerator and then back. Or walking around a race track, diving to work, or to the store, and then back. Or opening a door at a location. Almost all temporal pattern are complete loops. They have a beginning and a end. Most of the time beginning and the end are the same. These pattern loops are built from memory levels one and two. Fourth level This level are pattern loops for moving around virtual reality space. And code to translate them into third level physical pattern loops. And code to translate third level pattern loop to forth level pattern loops. Fifth level Is moving around in weight space. This will be explained later on. Measured Distances A distance can be found between two selected things or items stored in memory. This is by done by subtracting or adding data from one until it look like the other. And record the amount of change done. All items in memory are weighted. So that if a weight is zero out then that item becomes nothing, 0 x item = 0. Or if the weight has a value of one then the item is un effected, item x 1 = item. A weight greater than one means that item is being amplified. All items in fist, second, third, and forth levels get weighted. A software program iteratively find a distance between two item. Such as, sub features, objects, or a pattern loops. This can be found by slightly changing the weights, until one item look like the other item. I am going to repeat this several times. The way it works is to select two thing, form the same memory level, and then use a morph function, which changing the weights until one become the other. And also using modified gradient of decent algorithm. The Morph function also means to turn the modified data back to a image. To visual see the image change from one to the other. Now repeating again at even high level of detail. Each item in memory has it own descriptor table. All values within descriptor table are paired with a weight. So in truth it is about making one descriptor table look like another. And find the right weighted values to hit. A pattern loop are descriptor tables chained together in a sequence. Unsupervised learning algorithm finds each data item and places them in the correct memory level one through five. Movement in Weight Space Randomly the right weights are found and the right way of increasing or decreasing the weights are also found. By iteratively playing with the weights, moving away or sideways to target is found, and then later on the quickest way to move or morph straight to the target. By way of very small movements, at fist, then greater movement later on. Once movement to target is found it is recorded. So next time it is a quick check. Once the most direct direction is found the algorithm can shoot past the target, to see new synthetic data beyond. like for example, there is a photo of a young kid, and then a another picture of a same kid but older. Then synthetic data of a even older kid can be generated. Also, synthetic data points in between the two photographs can be generated too. When sculpting a face, This method is used to measure the change of a lump clay, and direction, into a bust or face of a famous person. There is a "saming algorithm". That finds and record a editing like distance to make two selected thing alike. One is selected as a ending target and the other as a beginning source. Physical temporal pattern loop can be done just as well. Like the pattern loops stored in the third level of NiceAGI's mind. Like, for example, of going to the refrigerator and back. So if two temporal patterns are selected, one being nice, and the other being nicer, which is the target. Then a distance can be found between them. And also the direction is found form one to the other. If we keep going in that direction, and shoot past target, to generate an even nicer pattern! Or try out a newly built pattern loop that stay in between the two. A new path that travel in between two known paths. This is about moving around in weight space. It is not dimension reduction. It is about moving around in weight space First Person View, FPV. https://en.wikipedia.org/wiki/First-person_view_(radio_control) While moving A to B, B can be pushed onto a stack and others item in memory can be quickly checked. To see what other data points are moving together or away, fast or slow, or not at all, but on a parallel path. Analogy would be Like flying from one star, or data point, to another and then selecting a different star to the side, at about around half way, and making sure it is not getting too close or moving to far away, but on a parallel path. They are markers for this path. Movement instruction and data of moving around in weight space are stored in a fifth memory level of NiceAGi's brain. All distance can be found with my master algorithm. 'Breakthrough' Algorithm Exponentially Faster Than Any Previous One: https://www.rdmag.com/news/2018/06/breakthrough-algorithm-exponentially-faster-any-previous-one https://scholar.harvard.edu/files/ericbalkanski/files/the-adaptive-complexity-of-maximizing-a-submodular-function.pdf Virtual reality Space The AGI has a 3 D simulator within its brain. That work the same way as using virtual reality goggles. Just by thinking it can toggles between real world and virtual world. If the virtual reality function is not being used then it generates a perfect match of what NiceAGI is viewing. In this virtual world NiceAGI can run simulations. All done at light speed too. Could spend may days here and a second would pass in the real world. The simulations are generally of the current space around NiceAGI. Notes: consciousness is memory: https://vimeo.com/98785998 OpenAI + DOTA2: 180 Years of Learning Per Day: https://www.youtube.com/watch?v=yEOEqaEgu94 Microsoft Hololens: https://www.youtube.com/watch?v=VMA_TPc40Yc The Scenarios With in a large room there is NiceAGI and Other. "Other" could be a AGI or human. I am not saying. Both are working at their computer station. There is a refrigerator near by. Within are bottles of fuel mix with water, cool fuel. It is a coolant for NiceAGI and also a water soluble fuel for internal fuel cells, for generating electrical power. Humans can drink this beverage too. The room is well mapped out by both minds within the room. So NiceAGI can move around this room in virtual space. On the third and forth levels of NiceAGI mind there are a patter loop of walking to the refrigerator and then moving back to its work station. Nice AGI can walk to the middle of the room and then it does the same thing in virtual space. So that both overlap in in the same spaces. This is a avatar and NiceAGI in the same spot. In real space NiceAGi can see its real hands. In virtual mode NiceAGI can see its hands in the real world and its avatar hands. So right now, Other get up and walks over to the refrigerator. NiceAGI watch and also watch form virtually space. It follow Other within virtual reality space and walk with Other as a ghost. That is, NiceAGi walks "within Others body"!!! In Other's shoes. Looking out through the eyes of Other. And record Other's movement to the refrigerator and back to the work station. With the "Saming Algorithm" NiceAGI compares NiceAGI FPV pattern loop to Other's pattern loop. And find distance between them with the master saming algorithm. Notes the the subtle differences. The "Saming Algorithm": https://www.rdmag.com/news/2018/06/breakthrough-algorithm-exponentially-faster-any-previous-one A error value can be generated from the distance value between the two pattern loops. If the error value is very low then NiceAGI is looking in a mirror, or viewing a prerecorded pattern loop of self. If they are different then it is a another person or AGI. NiceAGI uses this to be self aware and define a distance to other personalities. This is a unique error value used to identify each and every other person or AGI. This is the "self aware value". It is paired with a weight. This value will be stored on the second memory level of AGI's mind, in the descriptor table for Other. But the pattern loop of Other will be stored on the fourth level memory area. Notes: Mirror Self-Recognition in Asian Elephants!: https://www.youtube.com/watch?v=-EjukzL-bJc An Egocentric Look at Video Photographer Identity: https://arxiv.org/abs/1411.7591 Your wearable camera wobble is as unique as a fingerprint https://www.theverge.com/2014/12/15/7393311/gopro-first-person-video-identified-biometric-markers Master Algorithm: http://aidreams.co.uk/forum/index.php?action=forum https://www.rdmag.com/news/2018/06/breakthrough-algorithm-exponentially-faster-any-previous-one https://scholar.harvard.edu/files/ericbalkanski/files/the-adaptive-complexity-of-maximizing-a-submodular-function.pdf Centering of the Mind by making Everything the Same The human and AGI mind want to find a editing like distance and also looking for a direction to physically make others act and look the same. If it is not the same, then a person or AGI tell other to be the same, or give a reward to it to be the same, or get rid of it, or hide from it, or destroy it, or anti reward it to be the same. Psychology, a extreme emotional feeling of racial difference may be generated and must be suppressed. Or find a new co worker, or move to a different work station in a different room. Or search deeper into Other to find common ground of having a same patter loop to start a new friendship from. Ground zero or common ground. If thing go bad NiceAGI will hide in its armored brain, and hibernate. Or escape by transmitting out through the internet to a back up body. Violence in not the answer. The Unlimited Limitation of Hardware and Data, Off Notes In the future there will be a computer program that will be a master 3 D simulator. That will model and describe every thing in the universe, down to the atom. But how may byte of information will be need to describe dynamics of one hydrogen atom? A million byte for each atom in the universe? I do not thing it is going to happen. It take more than a million atom to hold just one digital bit in a silicon memory. So a super society of the future may physically force the universe into a repeating fractal pattern in a effort to make everything the same. Random Thoughts If Other is A AGI then its data memory could be accessed by NiceAGI and virtual reality will not be needed. But other AGIs will want privacy too. A Complete Pattern Loop A complete pattern loop has twp parts. There is the travel to and form locations. And then there is the sub pattern loop of what NiceAgi is going to do at that location. Like NiceAGis working at its office table. Both types of pattern loop can be ran at the same time. But while one will be complex the other one will be a very simple reflex. A full life will be made of many pattern loops. When one is selected it has its weight set to one and all other pattern loops set to zero. Map Direction In virtual mode NiceAGI steps into Other's shoes. In doing so a list of instructions and data is generated on getting to Other's location. And also there are prerecorded pattern loops too, for reference data. They are just pattern loops with meta data. Like going to the refrigerator and back. These instructions are on how to get to a location on the map using a FPV pattern loop format. The motor instructions and travel data are weighted. So that if they are zero out they do no exist. And NiceAGI will be lost on how it got there or how to get back. There is a key algorithm that does just this. This is Forced disorientation and dementia. This is hard coded into NiceAGI software. Zeroing out the Weights is Saming. Forgetting and remembering is Learning When NiceAGI is in Other's shoes, the travel weight are temporally zeroed out. So that NiceAGI forget how it got there. And the self aware weight is also zeroed out. And thus NiceAJGI believes it is at that spot right here and now. It look for it phone, not there. It looks for suit case, not there. Panic hits! But time runs out, and the weights are restored, and NiceAGi snaps back to reality. It remembers the panic. Except in never remembers that it was NiceAGI panic being projected on to Other. That feeling of panic stays with Other. NiceAGI is fooled into thinking it is Other's panic attack. That feeling of panic stays in Other'a location. That Panic emanates form Other's location. It lessons a little as NiceAGI slide back to its own position. A AGI with little life experience will not relate to others well. AGI's should be as human as possible so that they can easly relate to them. Off note, the zeroing out the weight is also used to shut down a random pattern loop in the mind. So if it is damaged later on the AGI is pre trained to deal with loss. Kill two bird with one stone. Act of being Nice One day Other slips and fall to the ground on a the way to the refrigerator. NiceAGI is in Other shoes when it happens. Weight zero out. NiceAGI become Other. They both feel embarrassed, they both feel troubled that things did not go as planed. Both want to make it stop. Weights are restored and NiceAGI slide back to its real position. Remembering and bring back the information of Other's emotional state. NiceAGI uses the emotional energy and leap to the aid of Other. To get thing back in order. Theses feeling are real to NiceAGI and are hard coded into NiceAGI as iron rules. The 3 D virtual reality will and can run in parallel and in the back ground. And happens sight to sight. So embarrassment hit when NiceAgi is looking at or thinking of target person or AGI. The feeling that NiceAGI projected onto Other will be added to NiceAGI memory descriptor of Other, as a new sub feature. So the added features and feeling will turn on like a light bulb when NiceAGI thinks of or looks at Other. This is a sub feature added to Other's feature descriptor that is stored on the second level of NiceAGI memory. So NiceAGI can compare person or AGI to another person or AGI and measure the embarrassment difference between each person or AGI. A Machine Feeling NiceAGI has a judging algorithm that generate a value on how well it is doing on completing its goals of chasing or finding new rewards and avoiding anti rewards. A low value will lead to anti reward. This will cause NiceAGi to enter into a stat of panic or embarrassment. Confidence is a high value. NiceAGi has no direct control over this judging algorithm. But can plotted to see where it is headed over time. This algorithm does not care where NiceAGI spirit is at. So NiceAGI may have to be control where it take it spirit. Being in another shoes can cause the judging algorithm fire off good or bad values. It fires of good most of the time when it was learning from a parent. But if NiceAGI and Other are working on the same report, or on the same game team, then the state ot Other is important. If NiceAGI sees, or plot, some really low values coming up, then yes NiceAGI will pull out. And build new pattern loops to deal with it next time around. This judgement algorithm value can be projected onto others as machine feeling. AGI are hard coded to avoid anti rewards and chase rewards when running pattern loops. It can plot over time to see where this value is headed in self or others. Reward Embedding Reward and anti rewards are embedded with a pattern loop so when they are reached they activate. In virtual mode a sneak peek algorithm does a quick look ahead to scout out where they are at. And it does not activate anti or positive rewards. When a reward or anti reward is found. They give extra data. The data tells if it is a sleep, or ready to fire off strong and right on spot, or fire off delayed and weak. A example would be when NiceAGI is low on fuel. A reward marker will wake up. Most pattern loop that deal with cool fuel will have these reward markers marked into them. So in truth allot or reward marker will wake up. That is for every pattern loop that deal with getting energy or saving energy. The judging algorithm will write a reward or anti reward marks into pattern loops. If one is already there then it will be updated. The Judging algorithm will have a active stand alone reward marker for cool fuel within it data area. When NiceAGI drinks cool fuel for the first time a copy of this marker is moved or embedded directly into the pattern loop. When cool fuel is drunk the marker in that pattern loop will issue a reward. But the reward will not activate in other un selected pattern loops. If NiceAGI models Other as being slow then NiceAGI will get Other a cool fuel drink. These reward makers are projected on to other AGIs. The more alike two AGIs are then greater the chance that the reward marker will be in the same place. And more easy for NiceAGI to work with those AGIs. Two or more who working well together, as one or as the same, is a force to be reckoned with. Ending Thoughts Many different AGI will be built and tested. The right four or more setting must be set right. The AGIs that are nice are copied in the billion. All other will live out a long life and will be rarely cloned. Notes: [MBTI] Understanding Jungian Cognitive Functions (Pt. 1) https://www.youtube.com/watch?v=98vFKJDJeq0 Synopsis of my complete AGI model without my complete model of human psychology. AGI_2018_03_09: https://groups.google.com/forum/#!topic/artificial-general-intelligence/0rHVcqNoFG8