An incredible skill that I have witnessed, especially at OpenAI, is the ability to make “yolo runs” work. The traditional advice in academic research is, “change one thing at a time.” This approach forces you to understand the effect of each component in your model, and

→ View original post on X — @_jasonwei,