The Bitter Lesson says general methods win. So which data model is least "human-sophisticated" for AI?

I think it comes down to two metrics: flexibility and structure.