Robotics data has a very different alpha decay problem than data used in large language models, according to a post shared by YZi Labs. YZi Labs posted on X.

Much of what LLMs learn can transfer across contexts, while robotics data is more conditional and depends heavily on the robot, the objects, and the environment.

As robots move into more complex real-world settings, previously collected data can lose relevance more quickly. For builders and investors, the focus shifts from how much data they have to how quickly they can close the loop between deployment, new data, and adaptation.