The live nature of modifying and viewing software has always been a special thing for me. The instant feedback and the direct results of your changes is an excited and fun way to explore a problem space. There is a magical moment when things begin to work. A magical moment that is particularly elusive in machine learning. When training machine learning models instead of solving a problem with a set of instructions, you define a set of constraints and a space to explore. It's a different paradigm and one that our current programming tools break down for.

Working at Weights and Biases I'm working to bring more of that joyful interactive experience. Choosing the correct visualization for debugging machine learning model is a challenging task. Machine Learning Model are training on very wide sets of data and over a long period of time. Instead of debugging a single point in time common in most software debugging problems, you are debugging on result sets with millions of data points across days of training.