On Thursday, May 28, 2020, at 11:38 AM, Alan Grimes wrote: > Let's think about this in terms of a problem stack Ok
On Thursday, May 28, 2020, at 11:38 AM, Alan Grimes wrote: > We place at the bottom of the problem stack: "Well I decided to deviate from my daily routine because Reasons". (We use Reasons to cut off a much lengthier discussion of causality...) The next line is "I started doing X but ran into Problems so I started working on ...." Um, this is just an example I'm guessing? Ok. I get it. On Thursday, May 28, 2020, at 11:38 AM, Alan Grimes wrote: > So what is your problem stack that got you to working on language compressors and why is it the best path to AGI? So you want some history of the direction I've taken? Ah I see. Well I've wanted immortality etc for years and am sure AGI is the next big step in Evolution. So I'll skip to AGI since we're all interested in that anyway. But first: The only other way in my eyes to immortality is Cryonics, and I have ideas how to improve it, but I have not the equipment. We already do Cryonics and it already does keep your global form. That goes to show it's a valuable path that is "easy". Again, I don't have the lab else I'd TRY, LOTS. One test is different sized nitrogen containers compared, if its bigger then the heat won't bounce back into the body and should freeze faster, we want it to leave and not come back. We need an accurate evaluation for better freezing too. I could spend time on this if others helped try novel tests I could come up with. So, AGI. I've been working 24/7 for 5 years, never a break. I jump out of bed and get right to the computer after eating. At some point you realize all AI networks do is learn patterns in data. They learn features/ representations. They learn a "model"/ hierarchy. Features build larger longer features. I already told you Alan, it's not a "compressor I'm working on", that's the evaluation used to test the network/predictor. I could use Perplexity, or the Turing Test. What it is, is a network, a web/ hierarchy. My algorithm / the theory of AI is understanding new data using previous experiences. It can predict what the next word is if shown a novel question. And it is not "language" as you call it either, all data such as vision or music works the same way right to the technical bits and guts, you can predict what comes next or do Translation. So what I'm working on is not compression nor language, but a data network. Note the net itself compressing is a totally different compression, that helps the network model learn patterns, things it doesn't have to store again. The evaluation-compression on the other hand tests the Predictor's ability in knowing what comes next in a feature/ sentence. So your question is really and absolutely and only "So what is your problem stack that got you to working on neural networks and why is it the best path to AGI?". My work explains / builds on others. I not only explain exactly what GPT-2 and PPLM and Blender is doing underlying the algorithm, but how to bring us much closer to AGI. THE WRONG WAY: I started off very very naively (but with good intentions!!). I was drawing roomba robots that would learn to crawl by randomly trying to move and tweaking actions that worked. A recursive update to "how to walk faster". I stopped, I realized AGI needs a lot of "considerations of things IOT style". Lots of context, big data. If it wants to walk faster, it may need to return home and not cross the finish line, build a rocket, and figure out how to steal a warehouse truck. I needed a thinker, not a walker robot body. The brain comes first, not the body...brain is where the erm "magic" happens. Brain is brain, not motors. You'll see where the body output comes into play though hold on. Output is a sensless thing without feedback, it's a loop that updates your domain attention. THE MODEL BASE: I was learning a lot how the model prediction / translation works. It can robustly recognize sentences or images or music despite similar objects/ words being in there and in different positions. There's more to this, as I showed in my videos. It allows it to understand cause and effect, to recognize unseen features as well. A model understands what can happen or what is what, and by how much so (%). MODEL DRIVER/ DESIRES/ STRONG ATTENTION DOMAINS FORCED: I learnt later how to force the model to talk/ask about certain things, like a GPT-2 that utters all day "I'm immortal by?" "I can become immortal if we?" "I will survive by?" AN EVOLVING GOAL, A NEW DATA SEEKER: Although I knew this I didn't see it as clear as recently. AGI needs to intake our current data, and pay attention to certain areas and generate (or collect) desired data from specific sources like certain websites, people, or questions to itself. It updates where to do this too, like a robot learning how to walk faster, it modifies it's goals evolving to a more narrow specialized domain ex. survival > food > money > job > boss > smiles > plastic surgery > plastic tool production, and starts generating data from those domains. It must nest back and answer the question once it gathers data from similar domains, I mean it's gathering data it needs to answer the real question in the end. Finding a path is important. To become immortal I need AI, AI needs nanobots, nanobots need energy, the goals/domains are similar and update. And AGI must recognize desired outcome milestones along the way or hold onto partially supported data by collecting loads of more partially likely/true data My work is simple. Unifies lots. I know a lot of things from nanobots to the future to activation functions. I've made that book and multiple movies. I build large note files of all my thoughts/ discoveries. AGI, needs a lot of data and a lot of attention when making a good decision. My design has 3 attention types: permanent (reward agenda steering) which can leak to related domain features to force it to talk on its own about sub-goals, Temporary energy activation (recently heard words likely to appear again (similar words are boosted as well, needs little data to work)), and long term memory storage of what usually follows some context during prediction or translation or byte pair encoding Segmentation. My design creates long term memories by frequent accesses and wires together features heard in close time. That's natural learning. My design also explains we prune not used nodes to forget them and Pooling as well during Mixing predictions in the net, among other things. AGI need a oddly abstract reasoning and fetch ability. It must do logical AND OR NOR, fetch or await some letter in a book, track time every hour checks, ask people questions instead of thinking about doing so (trigger real motor speech), must hold onto thoughts if decides it just must do so, carry over numbers mentally, and recognize the correct answer in the following by matching or entailing sentences of things: Those witches who were spotted on the house left in a hurry to see the monk in the cave near the canyon and there was the pot of gold they left and when they returned back they knew where to go if they wanted it back. They knew the keeper now owned it and if they waited too long then he would forever own it. Who owns what? Answers: Witches own monk/witches own canyon/monk owns gold/monk owns house/monk owns cave/cave owns pot/there was pot/he owns it AGI can translate, summarize, extend, segment data features. Extend and summarize are the same thing, you elaborate more if say more less important filer words, else say just the most desired, recent, frequent, and related word to summarize the more likely thing(s). Translation is recognition, it is similar sentence, and is stable just between summarize and elaborate and stays the sameish length when spoken. The brain is a data seeker...it can predict well....it finds desired answers to questions without coding/ implementing anything, all by large data evidence to find likely candidate leads that should occur in probablistical likelihood.....sometimes it does find a really good answer and it does everything as well - God, it finds a hole in the ground and can't even backtrack, a local but not global optima. I love Generative models. Because we need new, desired data to long standing root underlying questions like food, survival, sex, and shelter. We need AGI to run through desired paths in the net and know how to do so sensically. It must babble correctly, and babble about the desired domains ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T224cb80d0cc8b0e7-M9be9615547a48609d9c0826d Delivery options: https://agi.topicbox.com/groups/agi/subscription
