Presented with a paragraph, you can predict the next word better if you look at the last 20 words instead of the last 5 words. Looking farther back at the "now" (and comparing it to the "past"/"future") is a trend in all AGI mechanisms.
In video games, often reward comes far ahead in time and seems disconnected. The way to link it to what got the reward is same as GPT-2 text prediction/ recognition. If you're holding a key at the door and get reward, the word key is matched far back to when picked up the key. Predicting the next word is picking an evolutionary path in a "tree of possible futures" intelligently by chain rule. When you mentally discover a new/better desired question OR are confident enough to answer it, you get a reward feeling. This reward feeling is a node transfer infection that leaks neurotransmitter through open context channels. "I will cure ageing by_ (stopping ageing)" seems confident but, it's not exactly the answer at all, hmm, the answer could be quite lengthy in details; parts made of parts, this is done by elaboration of the node (opposite of summarization which ignores too common or duplicates / too rare, non-recent, boring, unrelated words or phrases). It is only then we can summarize the full details and say ex. "I will cure ageing by making AGI". It seems the answer to the question is a translation, i will kill A by throwing a rock at him, kill A = throw rock at him, semantics does equate to syntactics, if so, then the elaborated detailed explanation may be what makes it understood to answer the question (all the details/context prove A =/follows B). Bit confused here but I think I found a new lead. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta79d68474896311f-M366eb1a643ff3a0def1be514 Delivery options: https://agi.topicbox.com/groups/agi/subscription
