I haven't heard the stochastic parrot fallacy (though I have heard the phrase before). I also don't believe there are hard theoretical barriers. All I believe is that what we have right now is not enough yet. (I also believe autoregressive models may not be capable of AGI.)