One of my experiments was to use the Wikipedia random page feature to help me come up with good questions. I write a question that is answered in the first paragraph. Usually it is on some very obscure topic, so the question is one that very few people could answer correctly. Then I enter the question into Google. About half of the time, the top link is the same web page, but there are many cases where it finds a different page that answers the question. About 10-20% of the time, Google fails to answer the question, but that is a much lower error rate than a human would make.
Another test is to search for unusual items like "blue banana", "dog smoking a pipe", or "flying clock". In my quick test, it came up with relevant images every time. On Mon, Oct 28, 2013 at 4:07 PM, Jim Bromer <[email protected]> wrote: > Matt said: > OK, how about this. Make up a list of 20 questions that you think an > AGI ought to be able to answer. Next, mark the ones you think that > Google will be able to answer. Then try it and see which ones it > actually does. Then post the results of your experiment. > ------------- > > > That is relevant to my plan to create a simple AGI program that would > be able to learn through text-based IO, but those are not examples of > an AGI program that is able to learn. I acknowledge that this idea of > pushing the goal posts further away for every advancement is nonsense, > but my whole point is that I believe that it is feasible to write a > -very simple- AGI program that is able to *genuinely learn* using a > -simple- context sensitive language (with some typeIV moments). > > I just tried a few questions on Google and the best answers just about > contained the words in my question exactly. That suggests that a > context-free association phrase-to-phrase followed by a simple > context-sensitive decomposition and a slight bias for more > authoritative links really can do some of the legwork if the database > is extensive enough. But that initial collection is not good enough > for AGI. You would want an AGI program to examine the links to see if > they are actually relevant and helpful. > > But I especially did not notice that Google was learning by having a > conversation with me. That has always been one of the goals of AI > from Turing on. I am not moving the goal posts, I seem to be > explaining to people how the game is supposed to be played. The only > real question is whether I will actually be able to compete in the > field. But I am starting to think that competition in this game is 1% > inspiration and 99% coding. > > Jim Bromer > > On Mon, Oct 28, 2013 at 2:29 PM, Matt Mahoney <[email protected]> wrote: >> On Mon, Oct 28, 2013 at 1:22 PM, Jim Bromer <[email protected]> wrote: >>> How many tablespoons in a cubic parsec is a quite amusing and >>> unexpected question to have been solved. But wait a minute... Isn't >>> that a calculator question? If you are trying to awe and scare me >>> with this Halloween factoid, it is not going to work. >> >> OK, how about this. Make up a list of 20 questions that you think an >> AGI ought to be able to answer. Next, mark the ones you think that >> Google will be able to answer. Then try it and see which ones it >> actually does. Then post the results of your experiment. >> >> -- >> -- Matt Mahoney, [email protected] >> >> >> ------------------------------------------- >> AGI >> Archives: https://www.listbox.com/member/archive/303/=now >> RSS Feed: https://www.listbox.com/member/archive/rss/303/24379807-f5817f28 >> Modify Your Subscription: https://www.listbox.com/member/?&; >> Powered by Listbox: http://www.listbox.com > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/3701026-786a0853 > Modify Your Subscription: https://www.listbox.com/member/?&; > Powered by Listbox: http://www.listbox.com -- -- Matt Mahoney, [email protected] ------------------------------------------- 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
