> 1. No biases. Following LLaMA, PaLM, and others, we exclude all bias terms from our architecture in order to improve training stability.

What does this mean? What is a "bias term"?