[no_multitudes]

> internet training data is not where frontier capabilities come from

In that case, it should be no problem for the labs to train their new models without using public data, right?