Test every model that comes out, but be selective about what you put in production.

This is the line the press picked up from El Ojo, and it is the shortest version of how I think about model choice.

Testing everything is cheap and tells you what is possible. Putting something into production is a commitment to its failure modes, its cost and its behaviour when the inputs are strange. Those are different decisions and they should not be made at the same moment. The full talk makes the longer case.