You will get that in my upcoming guide but for now try this explanation (2 parts to it):
ROOT FORCE: I'll trust yous already know GPT-2 and the even cooler Blender. My discovery to improve Blender is: These AIs collect lots of diverse/general data (explores), but lots of it doesn't answer it's main question(s) or "domain of choice". Your AGI should have a domain like cancer research questions for example (preferably survival), forced (as they do in Blender), which causes it to talk most the day about cancer and therefore [generate data about cancer from the context around it] and also [collect cancer data off the internet] (like a cancer dataset). Exploitation. This is akin to feeding your AGI more data, because it helps you answer your question of focus better. EVOLVING THE ROOT: Your AGI should begin life with just a few words Forced (as they call it in Blender). Like food, sex, survival. For example "I will get maximum food by ". But it has to specialize more to better answer them. It won't get food if it doesn't become immortal, and food/reproduction is just trying to do that anyway. I mean it needs to invent new questions, not how do I get food/sex but how do I get money or job or immortality or AGI. The way it happens is food/sex is recognized semantically (by shared contexts) as money. Food/sex=money. Now it Forces the word/phrase "money"! It has a new focus/goal. It wants data around this context "money" now! Now it can further specialize by recognizing money node as job node by shared contexts. What this is is updating like a checkpoint where to pay attention to. It is translation, like semantics, you use similar experiences to answer your question, BUT, it's changing the question and changing what data it collects, from a specific domain source. If you look at robot RL where they learn how to walk etc, they update/tweak their favorite moves that give them most acceleration/prediction accuracy. At some point it doesn't gain much at each new tweak and may even get stuck in local optima. Same for what I presented above. When humans try real world experiments (lab, or even at home: living is an experiment!), we do it just to collect data! From a specific focused domain we desire. Our bodies are a feedback loop, we evolve where we collect data from, we update the test, we may change from cancer lab to AGI lab. Well, the brain can do the same thing. You don't need a body or world. Just data. Lots of data. You can find new discoveries in data. Once you do, you can update your Forced Goal dialog and search there to collect more new hidden data! Robots that learn to walk using say, Transformers, is nothing compared to something like GPT-2, because text/images model/describe Earth much more expressively than walking does. So stay away from robots learning to walk.... And the real lab tests mentioned, same, stay away, it is not they key to AGI, it is not efficient or as powerful or flexible. I can plan to go to Mars in my brain, faster, safely, increase my tool size, shape, etc all in my brain images/text. I don't need lab to make discoveries, just big data and data extraction. Our universe/world only has a few patterns, so enough data can capture the whole world. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T604ad04bc1ba220c-Md38f878ceb52b4a02f031dfe Delivery options: https://agi.topicbox.com/groups/agi/subscription
