Dr. immortals personality here, turned on: Physics governs Earth. Earth has been going through its very own Evolution all this time, based on starting conditions and the rules of physics. Evolution never ends truly, there will always be cycling and Brownian motion in any seemingly 'final equilibrium state reached'. You can end up at a very similar state from many different states, the opposite of the big bang. The universe could fall apart if gravity doesn't pull in photons from far distances.
We are at a stage where we are inventing our own intelligence and improving it. It is a big part of evolution. But it is like any other part of evolution. The future state of Earth is decided by the previous state of Earth. The past creates the future, based on rules of physics. This old data/context weighs in and the future emerges on its own based on combinational factors. It is a reaction that is decided, based on context. Embryos, plants, molecules, etc, all make decisions based on context. They react directly to the immediate stimulus. Old primitive brained animals do too. Although they are born with some remembering of the past, so that if they touch a hot coil their hand will retract immediately. They are born with large fangs. Boiling digestive acid. Narrow eyes. But they don't act on old context or possible future context. The human brain and dog brain can do this. They can react to old context and possible future context and discover old and future context. It's not old context though, it's stored in the brain, and it has energy in the nodes if it has been heard/seen in the last 1,000 words far back in time. Some nodes are always energized, like reward goals. The predicted words become the past energized nodes, the past is the current state. This reading of text makes you remember 'old node pasts', the far back text in a paragraph (that LSTMs tried to remember), ignore *that* term of 'far back', there is no paper, only when the sound gets to your brain do you have nodes energized. The human brain considers these old past nodes, and the more the better. The primitive brain can only react to short context, a short match in memory. It doesn't think using a long stream of words or image objects. Longer context results in more precise attention, narrowing down the text predictor prediction probabilities. And deeper translation of the stream of context allows you to do that. Problem with old brains is they can learn how to walk or chew and are born with reflexes, but they don't see a long context, and hence nor think of long context. They need to consider a lot of context to make a better decision, they need to see what comes later, what caused it to come, what options it has and what they relate to, they need to re_search certain words like Cancer and see what context has been seen around them. They need more context, and deep manipulative translation which looks ahead of the context you have, far back in your node network, and changes words. Where do you get your stream of context? You start with a simple seed question, and you already have a slew of sentences from the real world. You can do what GPT-2 does, and are already gaining more context. Your seed nodes with reward in them are the most important context, always active. They steer the prediction to desired outcomes along the mental maze of possible branches to simulate your plan through. Bigger systems have more context. That's why suns/stars and Uranium are radioactive, they are unstable because they are too large and therefore catch fire, like wood and gasoline and batteries, releasing free energy from the magnetic aligned/cooperative loops inside that are strong and wasted no energy aprior. Boiling water also spurts water. They are engines, extracting free data insights, just like text/image predictors in the crucial Hutter Prize for Lossless Compression. That's why Big Data is 50% of AGI, it is for the brain to "work" with, data "is" AGI. Context is AGI. It decides the future by comparing old pasts by thinking about them. When an embryo grows, at the particle level there is only a few laws being followed to make decisions. Every particle is listening to millions of particles around it by gravitational pulls acting on it. But the immediate local context matters most, because a human body still functions mostly the same way when a fridge or chair or waterfall is next to it (unless a planet is near it). All particle behavior is based on pulling and pushing using magnetic force. Energy carries the object with it until it hits wall or the air and transfers the energy somewhere. The reaction is immediate. But brains can "save" memory and associate them to motor actions tried. Note that randomly trying motor actions is slow way to learn skill or knowledge, the data discovery engine is the advanced way. And it does so by using hundreds of *internal memories as big context. This let's it make a decision, based on a lot of context from many past nodes it has seen. This is the way forward. We need big data, and it needs big context. And it needs to do deep translation to the context it works with. What you do is fill it up with big data, training it Online updating its network after seeing any new text, and install desired question nodes. It then starts branching to create NEW desired discoveries paths and adds them to the big data brain it already has. It is then able to use a long context window on not only the context the root brings up which follow but also the key agenda wisdom at all times no matter the Problem/Question, like check twice, ask others, save work before exit, etc, and it can do so such a long window of context by doing Deep Translation. Which increases the prediction probabilities. For example you see a sentence like "the 1850 president was Adam", and you have an always active context and a triggered context "what you hear for the next 2 minutes are things that your memory will be wrong about because they are from a simulated world" and "president Adam died in 1839", which change your prediction a ton. This is deep translation, not exact match, and not just translational matches, but wacky flexible deepy translation matches. The brain uses frequency in exact match method like in PPM, and it uses frequency in translation discovery, which is used to get end of stream prediction frequency higher and more accurate therefore. You can also predict the past and future too past the end of the text your at during prediction, breaking the end of chain prediction habit. This should improve your prediction further by looking at all the futures of futures of futures each bidirectionaly on both sides. Frequency is used for everything, including deep translation. It is the gravitational pull/push force. Text/images describe real world structures, like the molecule B, or molecule C, or cell A, or cell G, or a human, or a coffee pot. These contexts/structures, some are similar or have similar context around the structure. The similar structure, or similar surrounding structures, make structure A and B "look" similar, in context. That's why frequency is the big fundamental key, it shows what structure and change of structure comes next in space and time, and is used for deep translation to extract new desired insights. Frequency shows hidden truth in big data. And reward desire is key too, it steers the prediction to the desired hoped future. It can lead to lies/ too great hopes. If AI has no past old context (the current state of Earth), how can it give an answer to anything? It can't do anything but randomness. It needs a lot of data context, it must know the past Earth very well in all ways of angles of views/ That's AGI. Not exact matches that can only be short, not not looking ahead to the future before it arrives, not learning to walk or use a tool by immediate stimulus, but by mimicking others and extracting internal discoveries in the all knowing knowledge engine in the brain. It can talk/see about ant and all things, even the costly or impossible concepts like landing on Mars. To make AGI, you can use very very little data, simply huge data let's it run with exponentially more prediction accuracy. AGI is, by my definition, about creating/deciding the future based on Long Big context, not short exact match direct stimulus. At first the AGIs will need more data, especially microscopic data. Needs more duplicating nanobots for more energy and data capture, more processor, more eyes and arms. So we may need to try tests it tells us (wisely) to try (do real world lab tests, R&D). We don't gather the results randomly, we go into the lab with a good idea and then see what the results are. That's why this will work. Later, it will be doing it itself. And later, it won't need to, it will have enough data that it needs to do fewer real world tests - this is true for sure because it's prediction accuracy for real world structures of unseen cases rises exponentially the more diverse data it has. I still feel like there's something missing though I have to still implement the translation and reward ability. It has to tell us the the answers to our problems when it knows the accuracy is high and sound to the data it knows. And it has to be sure to inhibit it once solved too and move on to next question. It's sorta like a guided branching energy that is traveling in a mental maze, stopping waiting and circling paths until OK is seen (is sound with its global data net and soundly implemented by the humans) and can then carry on the branching gaining closer to the global goal in the mind, immortality. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T99116eafd9a8e4b8-M08e2b74ed2c1efaaab813126 Delivery options: https://agi.topicbox.com/groups/agi/subscription
