demis gave hint. what is hint for agi? only train with 1901 data and let ai think 1905 einshtein. and train that thought process for every major nobel price winning articles to train. train how to "think" like scholar. maybe try hybrid approach. set fundemental laws like physics first.