I am in favour of all kinds of Products on top of the current AGI codebases, as opposed to Ben who has OKed only some kinds of Products. But I see no place for a Minimum Viable Product as scalability is they key unknown of search space algorithms, if I may call it that. Perhaps Siri is a brilliant NLP assistant or Siri 3 will be a brilliant NLP assistant, mastering 1000 words, but as I have pointed out before, by no means being the first one to say so, full-blown language depends on a constant goertzelification, constantly redefining words and occasionally creating new ones, trying to produce a tight fitting film over a much more expansive reality, a bit like the rubber case for your smartphone. Reaching out to the millions of words and meanings and uses and deciding which ones will do the trick or which ones to modify to do the trick, well, even in terms of parallel programming and complexity metrics is a bit of a nightmare. I do believe however that the products will show the way, just like I expect robotics with their cumbersome dancing routines and other human-inspired but limited and primitive repertoires to slowly but inexorably advance towards MVP-like states and beyond. Especially by focusing on the *transitions* between repertories or domains etc, for example imagine how well a chatbot would do if it could appropriately and seamlessly transition between certain dialogue modes like "me", "wikipedia", "joke", "arithmetic", "common sense" - it would probably pass the Turing test already.
I was going to jump the gun and say I know what is not an MVP for AGI, "machine vision would not be an MVP", but then I had to remind myself that in my very own analysis a human level vision system would probably need to ask itself questions such as "could it be that there is an irregularly shaped black and white table partially obstructing a black and white cat?" in order to do visual object recognition. I cannot prove that the compositionality inherent in machine vision is of the same order as the compositionality/productivity of natural language or that of "physical existence" (meaning the compositionality offered by the material world itself once you actually interact with it, with its tremendous potential for emergence/surprise, such as getting an electric shock by touching the millionth piece of metal you encountered while all the previous touches gave no such shock). I think it would be OK if machine vision is an order magnitude below the complexity of language too, it would still be an MVP and quite a formidable one! Although my terminology is a bit more standard than the acrobatics of our usual suspects on this list, I am perfectly aware the intuitions will not be clear to the majority of AGIers, and certainly they are intuitions that relate weakly to a lot of academic but "unreal" AGI like Hutter's. I will try to see if watching Hofstadter can equip us with a few updated, powerful shared terms/metaphors, but at the end of the day the intuitions are for those who understand them and are willing and able to do three things: build, build, build! AT ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
