I've been expecting work like this as a next step. LLMs are a powerful part of the equation, but awareness isn't a threshold that is achieved by having a ton of data or processing power. The animal kingdom has plenty of examples. Awareness is a loop. It arises from the relationship between runtime and experience.
What I expect to be shocking, is when a CTM has multiple LLMs as part of its architecture, dynamic RAGs, and retraining of some of its smaller models.
More relevant work: https://github.com/em-llm/EM-LLM-model
The fun part about hole io is not just eating the most, but also moving skillfully, avoiding bigger opponents and taking advantage of every corner to grow the fastest. The feeling of going from a tiny hole to becoming a "boss" swallowing the whole map is a very satisfying experience.