A bigger Steamroller

I'm sure that Kurzweil will be making a lot of mileage out of this in subsequent books, but as an AI achievement it's not all that impressive.  Noam Chomsky describes it as "a bigger steamroller", and I think I agree.



My knowledge of TV game shows is pretty limited, so I'm unfamiliar with the details of how this game works. It seems to be a case of given an answer you have to guess the question - the reverse of a general knowledge quiz.

If anything this Jeopardy playing computer is a good illustration of what's wrong with what passes for AI, for quite a laundry list of reasons. The first and most obvious is the question of where the system's knowledge comes from. It's claimed that it "learns", and if it did this by watching old TV shows that would be much more exciting since this would mean that it had needed to solve many classically hard AI problems, but scratch the surface and I think you'll discover a whole bunch of men behind curtains manually entering text into databases in a highly selective way.  And that's just the beginning of the grievences.

Given a big database of past Jeopardy examples, and some other data sources like Wikipedia or Wordnet which can be used for additional semantics it would be possible to devise a classifier which matches the answers to the questions, or vice versa.  You can get as fancy as you like, employing boosted-neural-genetic-fuzzy-probabilistic classification methods of all conceivable types and flavours, and even run them in parallel with a winner takes all arbitration at the end which could itself be optimised, and I expect that's all that this system really is.  It doesn't need to have any self model or theory of mind, and in spite of superficial appearances it probably doesn't understand much human language either outside of the realm of Jeopardy questions.  Any situations involving construction or interpretation of a narrative, mental imagery, mental rotations, planning or perspective taking will also be off limits.

There's also the snake-oil salesman aspect, which isn't necessarily anything sinister but is a natural consequence of being presented with limited or selective information.  If you make gross simplifications - which is completely to be expected in the mainstream media - and vague claims like "It knows the English language" or "It has read Wikipedia" or "It learns new things as it goes" then it's easy for the naive viewer to conclude that this is a much bigger advance in technology than it really is by misinterpreting its limited scope and applying some anthromorphisms (I also expect some Singularitarians to leverage these misunderstandings for their own purposes).

Marvin Minsky's position on the workings of the mind is well known, but there is also an opposing view to this which comes from neuroscience.  People like Jeff Hawkins claim that much of the brain - specifically the cortex - consists of very homogeneous electro-mechanics, and that the labels assigned to functionally distinguishable areas are similar to the labels assigned to different muscles in an anatomical diagram.  The muscles do different things in different places, but fundamentally the mechanism is the same everywhere.  Who is right will eventually become much clearer once the connectome can be analysed.

There are some benefits to the above demonstration though.  It could be regarded primarily as a sort of public outreach, which might help to get young people initially interested in computer science or AI, which if Google ngrams can be believed has seen declining public interest since the late 1980s.  There is also the obviously self-serving benefit for IBM of getting some prime time TV advertising for their servers or software.

Here's some more video of the computer playing the game.