You could definitely do better than that with image recognition for terminal guidance. But I would assume those published accuracy numbers are very conservative anyway..

I think for LLM like Open AI, it wouldn't be about hitting the target but target selection. Target selection is probably the most likely thing that won't be accurate