He means developers should commit work in progress directly to the master branch. The work in progress would be disabled in production deployments by a "feature flag" boolean configuration setting. The exception to this rule would be very large changes which must be developed in a separate branch, and have been carefully planned with a corresponding RFC document.

In general I don't like this approach. Branches are troublesome when you hire mediocre developers who don't communicate and coordinate between each other. Merge conflicts are often a result of laziness, poor architecture and planning, and weak communication skills. In my experience, smart developers do not create problems with feature branches.