Thread by @fchollet

It would be fascinating to instrument deep learning developer workflows, so as to run the numbers on the productivity increase that comes from using tools that reduce cognitive load vs. tools that lay traps for you to fall into

The first thing that users that recently switched to Keras mention to me is the productivity boost. In the time it would have taken them to debug their way through the implementation of one idea (`zero_grad()` anyone?), they can try out 2, 3 ideas.

It matters because trying more ideas (with fewer mistakes) means you will converge faster towards better ideas (thus winning competitions more often or increasing your paper acceptance rate).

I'm thinking Kaggle kernels or Colab would be a way to gather hard data on this...