And the learning rate of a AGI on the net can be on the order of magnitudes much greater than that of a child, due to many concurrent users accessing and teaching. Given a solid framework, there is no real limit of the number of interactions these bots could handle. A decent system once designed should be able to learn most basic english in 6 months. The training time there is not the limiting factor at all.
Learning the English language (or whatever language) seems to really be the blocking point for most every AGI project I have heard of, and I would like to see a lot more interst and work done on that, even possibly before the overall structure of the rest of the AGI is concerned with, or at least a much mroe significant portion. Along with this is the representation problem, of course, of storing this information as its learned. The easiest and most usable / flexible AGI will be accesible to the populous using english, be it spoken or typed in, and we need to have that in place, to have any bot respond or act in a reasonable manner. James Ratcliff Matt Mahoney <[EMAIL PROTECTED]> wrote: --- Benjamin Goertzel wrote: > > But children do learn language by listening to adult level language > (although > > not optimally). > > Note that this listening occurs within a perceivable/manipulable > context, shared by the speaker and the listener. Kids don't learn > language by listening to an ungrounded stream of adult yakking! > > That would be a verrrrrrry slow learning process, e.g. think of a > child learning language solely by listening to adult-level talk radio > delivered with little emotion and no sound effects... Of course listening to speech is not sufficient, but it is a big part of learning language, especially learning phonemes and word segmentation rules before the first 12 months, then later grammar. Of course learning also involves speaking, pairing words with nonverbal stimuli, and so on. By "not optimally" I meant that parents teach their children language faster by using single words and simple phrases rather than adult level sentences. > > And statistical language models do learn a model equivalent > > to that of a young child after training on adult level writing. > > What do you mean by "equivalent to that of a young child"; I don't get > it.... How the heck do you know the internal language model of a > young child???? I find this claim not very credible... I mean that a statistical model such as a trigram model can distinguish between likely and unlikely phrases of up to about 6 words, but not adult level sentences. This is similar to the language level of a small child, but of course it is a rough comparison. -- Matt Mahoney, [EMAIL PROTECTED] ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?&; _______________________________________ James Ratcliff - http://falazar.com Looking for something... --------------------------------- Ahhh...imagining that irresistible "new car" smell? Check outnew cars at Yahoo! Autos. ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
