Honestly, I'd assume that pretty much any of these flaws are more fixable than you'd expect. We probably won't hit upon the right way for a while, since large models are so expensive, but with some trickery on the architectural level and bigger models, they'll spit out better and better art until it's indistinguishable from human level. Most of your brain is doing the same kind of dumb learning these models are, just much better and faster. There are some fundamental architectural differences, especially relating to the higher-level functions we associate with "consciousness" (working and short term memory, attention, maybe the executive?), but for almost every task you stick with the big dumb blob of compute and it gets the job done better than your higher faculties ever could. Don't underestimate the extent to which your brain plows through difficulties by throwing neurons at the problem until it goes away. Our biggest models are still a fraction of the size and a smaller fraction of the complexity of the bits of brain which perform the same function. I won't rule out anything until we see models of that scale.