Actually you're right, that was my mistake. Current-gen autoregressive token predictor "hallucinations" probably can't be fixed, but yeah, the way forward will involve new architectures and I would place my bets that something like world models will be involved to ground generative output.