In October 2025, Andrej Karpathy said on Dwarkesh Patel's podcast that current large models "have no continual learning. You can't tell it one thing and expect it to remember," predicting it will probably take another ten years to fix these cognitive defects. Two months later, in his annual review, he noted that the leap in model capabilities in 2025 mainly came from Reinforcement Learning with Verifiable Rewards (RLVR), but memory, multimodal perception, continual learning, and the ability to operate computers remain obvious shortcomings. He stated, "We have prototypes, but we don't yet have agents that can be used as colleagues." Continual Learning, also called Lifelong Learning, refers to a deployed model's ability to continuously absorb new tasks, knowledge, and experiences without forgetting prior learning. The core obstacle is catastrophic forgetting, where knowledge stored in billions of weights is lost during fine-tuning. This difficulty has made continual learning the hardest bone on the road to "AI colleagues." The path is no longer a single direction but several diverging routes advancing simultaneously. In the past year, academia and industry have developed distinct technical routes, including attaching memory outside the model, rewriting weights, re-pre-training the model, and newer ideas that redefine the act of "learning." The article expands on these directions, clarifying what each route bets on and where it gets stuck.