[cs702]

I don't think it's been tried at scale, with large models.

It remains to be seen if it works better than conventional training schemes.