Resolves YES if someone trains an AI system only on real-world data before information theory was invented (in the late 1940s) and the system is independently able to rediscover the key principles and theory around information theory with almost no domain-specific prompting, e.g. when instructed to "develop a principled mathematical theory of communication and information," or when given broad problems that we know require information theory to solve.

All non-synthetic data must be from before the late 1940s, but it's fine if training uses synthetic data so long as the methodology is credibly understood (based on accumulated wisdom from training retro-models) not to contaminate the model with information about the future.