You're starting a new RNN language model and want gradient norm clipping from the first step.

Nobody on the team has trained this architecture before, so there's no threshold to copy.