On Wed, Feb 13, 2013 at 6:39 PM, John Clark <johnkcl...@gmail.com> wrote: > On Tue, Feb 12, 2013 Telmo Menezes <te...@telmomenezes.com> wrote: > >> > So far, nobody has been able to figure out a learning algorithm as >> > generic as the one our brains contains. > > > The developers of Watson have come very close to doing exactly that.
What I mean by a generic learning algorithm is one that can program itself using very high-level feedback signals, something equivalent to pleasure/pain. That allows the same algorithm to learn how to drive a car and learn to speak new languages, and even learn new ways to learn. I think it's possible, but I'm not convinced Watson is it. > >> >> > there is definitely room for generalists. > > > Then why don't family doctors recommend that their patient see a generalists > when they run into a particular problem they can't handle? Because you don't want brain surgery performed by an amateur, and also to avoid law suits. But if I had to chose just one doctor for the rest of my life, I would chose a generalist. I'm not saying that specialists are not valuable, just that we've gone too far in fetishising them. Or, put another way, there aren't as many "brain surgeon"-level fields that require maniacal focus for competence as people seem to think. But everyone wants to believe that of their own field because specialisation is currently viewed as high status. > >> > Einstein might have been a great scientist in any field. > > > Perhaps Einstein could have been great in ANY field, but he most certainly > could not have been great in EVERY field. Agreed, mainly because of lack of time. Immortal Einstein would probably get bored of theoretical physics at some point and explore something else. > >> > Watson and Deep Blue cannot change their minds. > > > The great thing about computers is that every time they run a new program > they quite literally CHANGE THEIR MINDS. In a sense, but not in the sense I was alluding to. > >>> >>> >> Deep Blue beat the world human chess champion and it required a >>> >> supercomputer to do so, but that was 16 years ago and Moore's law marches >>> >> on; >> >> >> > Sort of. > > > There is no "sort of" about it, Moore's law marches on. In 1994 I bought one > of the most powerful PC's in the world, it had a one core microprocessor > running at 5 *10^7 cycles per second with 8*10^6 bytes of solid state memory > and a 2*10^8 byte hard drive and cost me $4000 in expensive 1994 dollars; > Today I am using a 4 core microprocessor running at 3.4 *10^9 cycles per > second with 1.6 *10^10 bytes of solid state memory and a 2*10^12 byte hard > drive and it cost me $2000 in in much cheaper 2012 dollars. The Moore's "law" marches on if you allow for multi-cores after a certain date and not before a certain date. > >> > Now it's progressing due to multi-core architectures, which one could >> > consider cheating > > > If I grew up on a farm and was retarded I might consider that cheating too, > but I didn't and I'm not so I don't. Funnily enough you're still vulnerable to basic formal fallacies, as the sentence above illustrates. > >> > because algorithm parallelisation is frequently non-trivial. > > > Few things worth doing are trivial, You wanted a touché but settled for a cliché? > but fortunately for us most physical > processes are inherently parallel as are most algorithms that are of > interest such as video and audio processing, playing chess, making quantum > mechanical calculations, understanding speech, language translation, weather > forecasting, car driving, Higgs particle hunting, and the sort of thinking > Watson did on Jeopardy. Yes, and in Nature they run on inherently parallel hardware. With von Neumann class computers we are stitching together sequential machines and trying to make them operate in a parallel way. This leads to very though problems like race conditions and deadlocks. One possible way out is the use of purely functional languages like Haskel, but implementing I/O in a purely functional way is also tough. Also, the CAP theorem imposes a theoretical limit on the capabilities of distributed computers: http://en.wikipedia.org/wiki/CAP_theorem Most of these difficulties can be surmounted, but at a cost. The higher the number of cores, the higher the cost. To make Moore's law work, people apply simple arithmetics (like adding the number of transistors) and ignore all these problems. > >> >> >I believe you're underestimating the complexity of a good chess program > > > A chess program good enough to beat the best human player could be run on > very primitive 1997 hardware, therefore I am not underestimating the > complexity of a good chess program. QED. It cannot be achieved with a few tweaks on a completely different program like Watson, which was what you were implying. It's not like the developers of Watson just said: "hey, we've created a very intelligent system, let's just throw some grand-master level chess playing capabilities in there". > >> > can Watson, for example, introspect on the chess game and update his >> > view of the world accordingly? Can he read a new text and figure out how to >> > play better? > > > Yes, Watson can and does learn from his mistakes Can he read a text about learning strategies and update his own learning strategy accordingly? > >> > could you ask Watson to go and learn by himself? > > > Yes, Watson spent many many hours organizing the vast amount of information > it contained and figuring out what it did wrong when it provided incorrect > answers in the past and trying new ways to improve performance. As a result > even the programers of Watson had no way of knowing what that machine would > do next; when Watson was asked a question they had to just watch and wait to > see what sort of response he would give just like everybody else. The only > way to know what Watson would do is to just watch him and see. That is true of many machine learning algorithms and it's not what I meant. > >>> >> If a person did half of what Watson did you would not hesitate for one >>> >> second in calling him intelligent, but Watson is made of silicon not >>> >> carbon >>> >> so you don't. >> >> >> >> > Nor for another second in considering him/her profoundly autistic. > > > Gregory Perelman is a mathematical genius who made the most important > advance in pure mathematics in the last 10 years, Perelman is also autistic. > Perelman is certainly not a genius about every aspect of human endeavor, he > recently turned down a $1,000,000 prize for proving the Poincare Conjecture > even though he's almost homeless. Perelman has his faults but would you > really want to say he is not intelligent? I'm making a distinction between generic and domain-specific intelligence. I agree that Watson and Deep Blue are intelligent, I disagree that they have generic intelligence. I assume a lot about Gregory Perelman just because he's a human being (albeit a very unusual one). The reason that claiming victory too early annoys me is that I actually care about the goal of AGI and believe it can be achieved. > > Another example is Richard Borcherds, he is also a mathematician and he won > the Field's Medal, in prestige it is the mathematical equivalent to the > Nobel Prize. Borcherds admits that he has been officially diagnosed with > having Asperger's syndrome, a condition closely related to autism. Yes, and there are many other examples. Turns out that autism can really make you focused. So what? Telmo. > > John K Clark > > > -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to everything-list+unsubscr...@googlegroups.com. > To post to this group, send email to firstname.lastname@example.org. > Visit this group at http://groups.google.com/group/everything-list?hl=en. > For more options, visit https://groups.google.com/groups/opt_out. > > -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to everything-list+unsubscr...@googlegroups.com. To post to this group, send email to email@example.com. Visit this group at http://groups.google.com/group/everything-list?hl=en. For more options, visit https://groups.google.com/groups/opt_out.