Superintelligence is written by Nick Bostrom. This is about construction of a general artificial intelligence and the reality that this is likely going to be an existential threat to humanity when it happens. Bostrom seems to be convinced that this will happen although the timeframe may well be 100 years as constructing a general artificial intelligence is still quite a difficult task. Personally, I have been trying to construct an argument using Gödel's incompleteness theorems to argue that a general artificial intelligence is really not constructable. Nonetheless, Bostrom describes a number of different paths that are currently being pursued that will get us to the same place.
The book is a tour de force of ideas and vocabulary. I've not learned so many new words since reading George Will's columns. If I had actually bought myself a copy rather than borrow it from the local library, I would have ended up circling some sentence every 2 or 3 pages with the note that it would be a good undergraduate thesis topic.
There are just a couple of spots where it feels like the subject matter has been massaged to appear more academic than it is (for example the analogy with horses), but overall, this shows that there is a significant amount of thinking that has gone into many of these topics and Bostrom does a superb job of summarizing the thinking and the issues.
Surprisingly, for a four year old book on the subject of AI, it doesn't really seem out of date. I am sure that it will start to appear aged in another 5 or 10 years, but I don't think that the real problems (the control problem, perverse instantiation, etc.) will be any different then than now.
I would highly recommend this to anyone who is interested in the field of computer science and especially to anyone who thinks that the current automated assistants (e.g. Alexa, Siri, "hey google", etc.) are the first step to Nirvana.
No comments:
Post a Comment