On Sun, Apr 7, 2013 at 6:31 PM, Piaget Modeler <[email protected]>wrote:
> Please expound. Some more examples and explanation As I have explained before, I keep a certain distance from domains peripheral to AGI such as neuroscience, linguistics and psychology, because a) they hardly ever bother with the engineering challenge of an (infinitely?) complex world, rather they still debate how do you store an apple and how do you signify a bicycle b) they hardly ever take into account the extended context of their work which again will have to be reverse-engineered and engineered, for example semiosis as something involving two non-identical parties at least, and mind design that would also have to find a compressed generative expression such as DNA, etc. We have also seen a lot of AGI work being solipsistic, I will never tire suggesting that work involving societies of agents, societies of mind etc will be far more fruitful than working on a single input-process-output loop, because the defining factor in most environments we care for is agency, unpredictable external and possibly peer, nuanced non-identical agency. In fact in my design the internal society and the external ones are a single continuum, it is my design goal to make sure "two minds are better than one" and "half a mind is not a crippled mind". Specifically, and I have written this before, semiosis will work differently depending how you engineered your "reality engine", your mind: mind <----> intention <----> sign <----> intention <----> mind Let me rephrase the above diagram. 1) I think I know how the world is 2) I would rather counterparty B work with me to improve it 3) hope these words will do it 4) counterparty B luckily is available to read the sign 5) mind B will think it over and either sign back or actuate, You may be a bit skeptical about the second "intention" above, surely I can sign "help!" and get to your brain even if your intention was to simply sleep for 8 hours. Well, if we want to be pedantic we can always assign you an intention after my screaming, perhaps the intention "ignore", "shut up" or "I am the kind of person that helps people, so action". Or we could consider rewiring my communication scheme, I do not consider it either fixed or final. Quite obviously most AGIers sweat the mind part while PM decided to look closer at signs, for whatever reason (troubled by grounding, perhaps?). Now, the reality engine can be constructed in loads of different ways, symbolic or connectionist, with a little, some or a lot of build in biases etc. It goes without saying that one of the biases in a social agent is to look for signs. Again signing could be attempted in many different ways, build a language from scratch or share a few building blocks or, impossibly for human agents, share and replicate whole cognitive units between agents "here is my bicycle recognizer, here's my Chinese speech synthesis module, now let's move on". Interestingly, I have certain thoughts (mostly prejudices) for, let's say, Romanians. Of course it is an extremely fluid category that objectively will be hard to tie down to any particular individual as they ones I do know are all different, and certainly I am quite clueless about the geographical borders of Romania and even if I were clued up it would be indefensible to say that someone from the Romanian side of a hill is an instance of a concept that is different from the concepts available on the Hungarian side of the hill. The reason I am bringing this up is to both a) more or less reject Platonic eternal concepts (except in the trivial sense that everything expressed or impossible to express is a Platonic concept) and b) suggest that there is no Romanian concept in my reality engine, except to the degree I have integrated the linguistic engine into the reality engine (Wittgenstein), and I can think all I want about the concept of a Romanian (human individual) without really bothering with a reality check, just as I could be pontificating about Klingons and dodos, especially blue-beaked dodos. Hopefully you agree with me that the concept of blue-beaked dodos exists even though you have no idea how they may look and frankly someone will have to disprove their actual existence (Popper). You may also agree that without fuzzy concepts like Romanians in my mind I would live a very poor intellectual life, perhaps one consisting exclusively of objects of immediate gratification and survival, it is hard to argue that a can of beer is a weaker concept than a quantum particle, and the hypothesis that beer is a collection of quantum particles adds very little to the beer. While beer may help a lot when dealing with quantum particles lol. So, where does this all leave me - it depends on the entire cognitive architecture end-to-end. a) you created a "mute" reality engine based on a statistical technique, and it can do impressively many things in its environment. Now it observes a peer apparently and would like to harness the synergy potential, if only you could sign your intentions. Given that another entity has had a different "life" therefore created different statistical distributions, decision trees etc you probably shouldn't rush into some heavy handed manipulation like pushing its buttons they way you would push yours. Inevitably you would start with pantomime/body language/audiovisual aids , which could take a very long time until, incrementally, you would be able to compress it to "me Tarzan, you Jane, me left, you right, we kill antelope" and perhaps you would take it from there. In some ancient postings I inquired about possible cognitive primitives, things a human-like intellect should know a priori, and "kill" is one of them (regardless of how long it takes for a child to lean about death and all the "your mom went on a very long trip" stuff). b) you created a reality engine with a heavy dose of symbols built in, it knows thousands of handcrafted categories like "dangerous snake" and "ripe fruit" and is even experimenting with the world creating (inside the reality engine) an avalanche of faux-concepts like the concept "Romanian" for individuals that hold passports with that indication, and tries to generalize their traits. Then any new Romanian passport holder becomes an instance of the type Romanian, and you allocate some resources to this data survey and analysis of yours, depending on the relevance of this type to your survival/gratification. Then you meet another reality engine that knows about snakes and fruits, but among its peers it is mainly running a different experiment, right-handed vs left-handed individuals. You want to share your interim findings and retrieve your counterparty's, but you have no clue what each other's work entail, even left-handed is unknown to you as a term/sign. How are you going to proceed? You probably have to bring the mystery sign down all the way to your reality engine, and allocate some resources (depending on how much you love/trust/fear your peer) for its elucidation. I know that the subtleties of these descriptions and examples will go unnoticed by most, it is not a matter of education or IQ, rather a matter of tuning in. If you tuned in you may agree that semiosis is largely a non-problem. How do you do your basic semiosis? You are hardwired to(and if you are a machine you should be hardwired to) quickly acquire/agree upon an everyday vocabulary. How do you acquire new signs, hopefully more advance ones? You swallow them whole, try to make sense out of them or at least to get value out of them, and ideally in good time you spit them out. Of course I am not an expert into Peircean semiotics or most of the other branches of science that are cited in the bibliography you shared, but reading that they shied away from investigating language development and instead looked into how to share or ground a few words, and that in 20 or 60 years of research, that is semiosis enough for me! And as much as I applaud PM's attempt to categorize different AGI approaches, I think that the world needs my taxonomy which is more focused on how agent societies could develop language and how the workload inside a brain or inside a society is distributed elastically and robustly (you don't want slow learners, liars or plain "idiots" slowing down your society or brain to a breaking point, do you? I guess the world is not ready for my taxonomy yet so I haven't bothered producing it ;) But it is in the works and I urge interested parties to explore all pathways, develop all taxa and phyla, even the ones the do not like. PS a bit about grounding: I don't think I have all the answers but I do have two answers a) an embodied autonomously surviving agent is de facto grounded b) all agents are grounded in "their own way", grounding resembling "life" and thus forever too subtle, controversial and possibly impossible to define. 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
