Putting a ceiling on how large a single update can be. It stops one unusual batch from wrecking a model that has been training for weeks. Cheap insurance, and standard practice in large training runs.
Technical view
Rescales the gradient when its global norm exceeds a threshold, preserving direction while bounding magnitude. Distinguish from element-wise value clipping, which distorts direction. It is a routine defence against loss spikes and divergence, particularly in recurrent networks and large transformer runs, and the clipping frequency is a useful diagnostic signal that something is wrong upstream.
Related
Part of AI Terminology Index, section: How the Machinery Works.