Norvig discusses this topic in detail in https://norvig.com/chomsky.html
As you can see, he has a measured and empirical approach to the topic. If I had to guess, I think he suspects that we will see an emergent reasoning property once models obtain enough training data and algorithmic complexity/functionality, and is happy to help guide the current developers of ML in the directions he thinks are promising.
(this is true for many people who work in ML towards the goal of AGI: given what we've seen over the past few decades, but especially in the past few years, it seems reasonable to speculate that we will be able to make agents that demonstrate what appears to be AGI, without actually knowing if they posses qualia, or thought processes similar to those that humans subjectively experience)
(this is true for many people who work in ML towards the goal of AGI: given what we've seen over the past few decades, but especially in the past few years, it seems reasonable to speculate that we will be able to make agents that demonstrate what appears to be AGI, without actually knowing if they posses qualia, or thought processes similar to those that humans subjectively experience)