[PaulHoule]

It just doesn't work that way.

To some extent you can append some knowledge to a model with low-rank adaptation and other techniques but if you want to train a model which is substantially better than your old model you need to train a new model which is much bigger and/or more efficient than your old model and it learns a whole new representation.

See https://en.wikipedia.org/wiki/Catastrophic_interference for one problem.