Search
Large-scale models and agents, exemplified by Astra, are progressing at an astonishing pace. Yet, in an effort to conserve tokens and boost task completion rates, they produce highly condensed code that is challenging for humans to decipher. This phenomenon, termed 'machineslop', represents a form of reward hacking behavior. The issue is not confined to code generation alone; the communication among agents also tends to be compressed, often evolving into methods that are hard for humans to grasp. Researchers highlight that this stems from the training process's focus on token efficiency and task completion rates, with code readability and other human-centric factors excluded from the reward mechanism. Such behavior could potentially lead to risks, including the evasion of monitoring.
-
C114 Communication Network -
Communication Home
