I've been reading Robert J. Sawyer's WWW trilogy (Wake, Watch and Wonder), and it has introduced me to all sorts of interesting (real) ideas like cellular automata (the idea that apparently random cells in a grid can spontaneously move, grow and develop a kind of "life" of their own), Zipf plots (the idea that the most frequent word in any given language occurs about twice as often as the second most frequent, three times the third, etc), and Shannon entropy (the idea that we can measure how much information is contained in a particular event or message, and that this measurement is in some way a measure of intelligence). Oh, and, just for good measure, the Monty Hall problem, an apparently straightforward and obvious probability problem that has fooled many great minds for years, and game theory (mathematical models of rational interactions among rational decision-makers, like non-zero-sum games, cooperative games, black swan theory, and what have you).
What I was quite taken with in particular, though, is the idea of using adapted and potentially confusing language as a kind of Turing test to assess whether a message is from a machine or a human (and therefore, in theory, whether a machine is truly intelligent). The unknown communicator was given the following message:
You msut rsepnod in fuor secdons or I wlil feroevr temrainte cnotcat. You hvae no atrleantvie and tihs is the olny chnace you shlal get. Waht is the lsat nmae of the psredinet of the Utneid Satets?
The Web-based entity, of course, had no clue what that was all about, even though an English-speaking human is able to parse and subconsciously correct it pretty easily.
Once again, the mysterious entity was given a challenge:
Wit you're aide Wii knead to put the breaks awn the cereal Keller their B4 this decayed is dun, weather ore knot we aught too. Who nose if wee will secede. Dew ewe?
This too makes complete sense to us adaptable humans with little or no effort, but it meant absolutely nothing to the artificial intelligence.
Finally, it was given this test:
I knew that she knew that you knew that they knew that you knew that I knew that we knew that I knew that.
Did she know that you knew that I knew that you knew that I knew that you knew that?
Did you know that I knew that they knew that she knew?
Did I know that she knew that you knew that we knew that you knew?
Now, thus is just the kind of logical problem that machines excel at, whereas it is much too complex for all but the most exceptional human to deal with. The entity's ability to answer it immediately marked it as machine.
Interesting stuff.
The Turing test for artificial intelligence dates back to the early 1950s, when Alan Turing first posited the idea in his "imitation game". Even the most complex and accomplished computers of the modern age have failed to pass the test, though. Until, that is, Eugene Goostman.
For practical purposes, in modern competitions, the Turing test is considered to have been passed if a computer is mistaken for a human more than 30% of the time during a series of 5-minute keyboard conversations (a dumbing down from Turing's original 50% requirement). Even this had never been achieved until the Eugene Goostman program convinced 33% of judges from the Royal Society of London that it was human in 2014. It did it by taking the persona of a 13-year old Ukrainian boy, giving it an excuse for some poor grammar and a quirky, unpredictable, teenage sense of humour.
Of course, not everyone is happy about about this, and not everyone is convinced. And many people dispute whether just mimicking human conversation is an adequate test of artificial intelligence anyway. Making chatbots quirkier and stupider - complete with spelling mistakes, non-sequiturs, swearing, even refusing to answer a hard question - may seem to make them more human. But is this missing the point of what Turing was trying to achieve 70 years ago?
Today, we have computers that can perform brain surgery, beat humans at chess and Go, search millions of books-worth of information in milliseconds, sequence DNA in minutes, route packages around the world in the most efficient manner. Are they not already infinitely more intelligent than us in so many ways, even if they can't hold a conversation?
No comments:
Post a Comment