MIT researchers can now track AI’s decisions back to single neurons

Would be nice to read some comments on that from practitioners. How far are we with being able to interpret neural networks at the scale used in industry? And what if the inputs are not images?