DeepMind scientist: the intelligent agent is the world model, which coincidentally aligns with Ilya's prediction from 2 years ago.

DeepMind scientist Jon Richens' paper presented at ICML 2025 points out that any agent capable of generalizing to multi-step goal-directed tasks must have already learned a predictive model of its environment. This model can be extracted from the agent's policy; to improve the agent's performance, or to complete more complex goal tasks, a more accurate world model must be learned.