I wouldn't say I changed my mind, but I substantially increased my p-estimate that the following recipe could produce something very close to intelligence:

1) a vast unlabeled data set (think 1e8 hours of video and audio/speech data plus the text of every newspaper article and novel ever written)

2) a simple unsupervised learning rule (e.g. the reduced Boltzmann machine rule)

3) a huge computer network capable of applying many iterations of the rule to the data set.

I previously believed that such an approach would fail because it would be very difficult to "debug" the resulting networks. Now I think that might just not matter.