On Sat, Feb 5, 2011 at 12:27 PM, Colin Hales <c.ha...@pgrad.unimelb.edu.au> wrote:
> I think perhaps the key to this can be seen in your requirement... > > " Doing this is equivalent to constructing a human level AI, since the > simulation could be given information and would respond just as a human > would given the same information." > > I would say this is not a circumstance that exemplified human level > intellect. Consider a human encounter with something totally unknown but > human and AI. Who is there to provide 'information'? If the machine is like > a human it shouldn't need someone there to spoon feed it answers. We let the > AGI loose to encounter something neither human nor AGI has encountered > before. That is a real AGI. The AGI can't "be given" the answers. You may be > able to provide a software model of how to handle novelty. This requires a > designers to say, in software, "everything you don't know is to be known > like this ....". This, however, is not AGI (human). It is merely AI. It may > suffice for a planetary rover with a roughly known domain of unknowns of a > certain kind. But when it encounters a cylon that rips it widgets off it > won't be able to characterize it like a human does. Such behaviour is not an > instance of a human encounter with the unknown. I am considering a special type of AI, an upload of a human brain. A "bottom up" AI, if you like. SPICE allows you to simulate a complex circuit by adding together simpler components. You can construct a simulated amplifier out of simulated transistors, resistors, capacitors etc., then input a simulated signal, and observe the simulated output. You construct an uploaded brain out of simulated neurons, input a simulated signal to simulated sense organs, and observe the simulated output to simulated muscles. The input signal could be a question and the output signal couldbe verbal output in response to the question. If the SPICE model is a good one its output would be the same as the output of a real circuit given the same input. If the brain upload model is a good one its response to questions would be the same as the responses of a biological brain. For example, you could tell it the result of experiments, it would come up with a hypothesis, propose further experiments for you to do, then modify the hypothesis depending on the result of those experiments. There is no specific novelty-handling model: the upload is merely an accurate model of brain behaviour, and the novelty-handling emerges from this. The analogy is that the SPICE software does not have a specific model for what to when the input is sine wave, what to do when the input is a square wave, and so on, but rather the appropriate output is produced for any input given just the models of the components and their connections. > Humans literally encounter the unknown in our qualia - an intracranial > phenomenon. Qualia are the observation. We don't encounter the unknown in > the single or collective behaviour of our peripheral nerve activity. Instead > we get a unified assembly of perceptual fields erected intra-cranially from > the peripheral feeds, within which the actual distal world is faithfully > represented well enough to do science. > > These perceptual fields are not always perfect. The perceptual fields can be > fooled. You can perhaps say that a software-black-box-scientist could guess > (Bayesian stabs in the dark). But those stabs in the dark are guesses at (a) > how the peripheral siganlling measurement activity will behave, or perhaps > (b) a guess at the contents of a human-model-derived software representation > of the external world. Neither (a) or (b) can be guaranteed identical to the > human qualia version of the the external distal world _in a situation of > encounter with radical novelty (that a human AI designer has never > encountered). This observational needs of a scientist are a rather useful > way to cut through to the core of these issues. > > The existence or nature of 'qualia' that give rise to (by grounding in > observation) empirical laws, are not predicted by any of those empirical > laws. However, the CONTENTS of qualia _are_ predicted. The system > presupposes their existence in an observer. The only way to get qualia, > then, is to do what a brain actually does, not what a model (empirical laws) > of the brain does. Even if we 100% locate and model the atomic-level > correlates of consciousness (qualia), that MODEL of correlates is NOT the > thing that makes qualia. It's just a model of it. We are not a model of a > thing. We are, literally, something else, the actual natural world that > merely appears to behave (in our qualia) like a model. In the case of the > brilliant model "electromagnetism", our brains have electromagnetism > predicted by the model. My PhD thesis is all about it. Humans get to BE > whatever it is that behaves electromagnetically. Conflating these two things > (BEing and APPEARing) is the bottom line of the > COMP is true belief. > > Note that none of this discussion need even broach issue of the natural > world AS computation (of or not of the kind of computation happening in a > digital computer made of the natural world). This latter issue might be > useful to understand the 'why' of qualia origins. But it changes nothing in > a decision process leading to real AGI. > > BTW my obsession with AGI originates in the practical need to make a > 'generic AI making machine'. One AGI makes all domain specific AI. I don't accept that computers cannot have the same qualia as brains, but even if that is true, it is possible to do science without direct observation. A written description of the results of an experiment is enough. -- Stathis Papaioannou -- You received this message because you are subscribed to the Google Groups "Everything List" group. To post to this group, send email to firstname.lastname@example.org. To unsubscribe from this group, send email to everything-list+unsubscr...@googlegroups.com. For more options, visit this group at http://groups.google.com/group/everything-list?hl=en.