3 ms·
Maybe the way to go is to do pre-training on copyrighted data, then to thoroughly shake things up so that hopefully only some useful abstract structure of world
by quonn 3y ago
Maybe the way to go is to do pre-training on copyrighted data, then to thoroughly shake things up so that hopefully only some useful abstract structure of world knowledge remains and then train that on carefully selected licensed data.
- disgruntledphd2 3y agoIf the models weren't just doing massively complicated interpolation then this would probably work. Honestly the only way to deal with this is to change the training data and retrain everything (probably at the cost of performance).