When we have near-human level deep learning models for things humans can do, what does this level of performance look like when applied to things we can’t do at all? This is a nice example of this.