[gwern]

Obvious observation: this seems like something that a GPT-4-V-like multimodal model could probably do, which would make it much more scalable, and probably more tractable for interviewing too due to the faster speed.