It's true, masked-language-modelling has been hugely successful in natural language processing. So a priori I wouldn't have guessed it would work so poorly for

It's true, masked-language-modelling has been hugely successful in natural language processing. So a priori I wouldn't have guessed it would work so poorly for humans and that phonics-based teaching is so much better. But the evidence is clear. Just because MLM can work doesn't mean it does work in all contexts. We get to change the software (and hardware although that has been less relevant IMHO) behind ML models as much as we want, for humans we need to adapt the training methods.