Charles:An AI doesn't need the concept of truth...except to communicate with people
No way. An AGI agent operating in the real world must, like all animals, check whether it can “believe its eyes” – did it really see/sense that thing it thinks it saw? Was that the noise of an animal approaching through the bushes, or just a breeze? Truth-checking is continuous in the real world. Except that it’s not so much “truth” as "*reality*-checking. The notion of all-or-nothing truth belongs to a logical/symbolic culture, which is playing around with symbols, checking whether the letter A is the letter A. Easy-peasy. Unidimensional. That suits the present generation of Chinese-room AGI-ers who never want to leave their room. But reality checking is a whole different ballgame, and will characterise the new multimedia culture of this millennium.. You don’t ask whether an image/picture of a person is a true picture. You ask how realistic it is – because as soon as you deal in images, you enter the complex, multidimensionality of the real world. Every picture for example has a certain POV and angle, looks at the subject in a certain light, applies a certain colouring, captures the subject at certain moments in time, places them in a certain frame and context, and so on. A priori, there are an infinite set of other possible pictures. This one may be a pretty realistic pic, but others will shed fresh light, offer a different perspective, capture different aspects of their personality and behaviour, set them in different contexts and elicit different associations, and so on. Truth is for an artificial, unidimensional world, realism for the real, multidimensional world. From: Charles Hixson Sent: Wednesday, February 20, 2013 12:56 AM To: AGI Subject: Re: [agi] Truth On 02/19/2013 11:03 AM, Piaget Modeler wrote: I'm sure this topic has been discussed before. Sorry for rehashing it if so. I have a specific question I'd like to answer. In designing a cognitive system, someone made a criticism that utterly confounded me. And got me thinking. The system receives sensory data sets from the world and transforms them into percept propositions which it asserts to its memory. Each percept proposition is activated when it is asserted. Infereneces are made from these percepts. These initial percepts and its inferences are called "Observables". All observables can be activated, but there is only a notion of activation. Next, the system can predict that these observables will recur at some point. But the prediction refers only to predicting the re-activation of observables. Then some one asked, where is the notion of TRUTH in your system. I was flabbergasted. Speechless. Then I asked well what is truth? I checked wikipedia. ( http://en.wikipedia.org/wiki/Truth ) It turns out that when someone says something is true, it means a very many things: a) It means that the statement is logically consistent (validity), b) that the statement corresponds, concurs, or conforms to reality (verity), c) that one is sure of the statement (certainty / confidence), d) that the statement is likely to occur rather than unlikely (Likelihood), and e) that we agree with the statement (agreement). So my questions are: (1) Is truth necessary or important to a cognitive system? (2) Which notion of truth should a cognitive system model? (3) How do we ascribe truth (values) to sensory input or inferences derived from sensory input? Your thoughts? ~PM. ------------------------------------------------------------------------------------------------------------------------------------------------ Confidential - This message is meant solely for the intended recipient. Please do not copy or forward this message without the consent of the sender. If you have received this message in error, please delete the message and notify the sender. AGI | Archives | Modify Your Subscription Truth is an illusion. It is the belief that what you believe to be most likely is, in fact, inevitable. An AI doesn't need the concept of truth...except to communicate with people. Internally it can operate off of graded degrees of probability, cost, benefit, etc. When communicating with people it needs to condense that so that when something has more than a certain amount of probability, and the benefit of asserting it is sufficiently large, and the cost of being wrong is sufficiently small, then it synopsizes this as proclaiming "truth". It's my belief that people operate in the same way, though this is disguised because different people use different constraints on things like "What is probable enough?". Also note that the cost and the benefit are figured on the basis of the cost/benefit to the entity proclaiming a truth rather than on those accepting it. So perhaps we would want a sufficiently capable AI to avoid talking about truth, and instead talk about what the probabilities are, and what costs and benefits can be expected. It's a bit harder to understand, but it strikes me as much safer. -- Charles HixsonAGI | Archives | Modify Your Subscription ------------------------------------------- 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
