Interesting, thanks. He brings up some good points which I partly agree with, but he seems to be only considering highly exact recreations, which I would agree are unfeasible. We don't need anything near exactness for success though.

This leads to a first tentative argument against reconstruction based on external data: we are acquiring potentially personality-affecting information at a fairly high rate during our waking life, yet not revealing information at the same high rate. The ratio seems to be at least 1000:1.

True, but much of the point of our large sensory cortices is to compress all the incoming sensory information down into a tiny abstract symbolic stream appropriate for efficient prediction computations.

A central point would be the inner voice: our minds are constantly generating internal output sentences, only a fraction of which are externally verbalized. The information content of the inner voice is probably some of the most crucial defining information for reconstructing thoughts, and it is very compact.

That's my short reply on short notice. I'll update on Anders points and post a longer reply link here later.