Training a model with multiple billion floating point parameters on only 100 billion data points feels like a bad idea.