Shane Legg has some thoughts on the Summit. I was able to talk to him for about 15 minutes at a party before the Summit. Unfortunately I missed his spiel about which AGI approaches are most promising.

On the way out, I picked up Shane’s doctoral thesis, Machine Super Intelligence. I saw in Anders’ summary of the event that he bought it too. I’ve just made it through the first 36 pages and already it has presented several important points that I thought were unique. For instance, it has the first summary of tests of machine intelligence in the literature, as it points out. Some of the math is beyond me, but I look forward to the rest of the thesis.

In the conclusion of the first chapter, Shane writes, “Although this chapter provides only a short treatment of the complex topic of intelligence, for a work on artificial intelligence to devote more than a couple paragraphs to the topic is rare. We believe that this is a mistake: if artificial intelligence research is ever to produce systems with real intelligence, questions of what intelligence actually means and how to measure it in machines need to be taken seriously. At present practically nobody is doing this. The reason, it appears, is that the definition and measurement of intelligence are viewed as being too difficult.”