This my personal opinion and may be entirely worthless. The quality of answers I read in all of the examples posted in that article read like the questions were

This my personal opinion and may be entirely worthless. The quality of answers I read in all of the examples posted in that article read like the questions were routed to an offshore boiler room where the answers were crafted by humans. Like some modern day Mechanical Turk. Especially in the 6 eggs example, there is a complete discontinuity of thought across the answers, isn't this within a single session with the AI? To me it looks like different brains answered each question/challenge and seemed to have a bias toward not offending the human asking the questions. Also, in this example, the first answer of 2 is correct: broke 2 (6-2 = 4), fried 2 (4-2 = 2) then ate 2, which most commonly implies it was the fried eggs that were eaten (2-0 = 2)