When these things start solving many more long standing problems, and start introducing more novel problems, will people finally admit that the "next token predictor" is not the gotcha they think it is?
It's not a gotcha. It's incredible what these things can do despite being next token predictors from a weird dataset. That's at the heart of the "bitter lesson", and you don't have to believe in magic to see it.