Mythos/Fable training: thought process becomes indecipherable, even screams 'AAAAARGH'

According to the system card report during Mythos/Fable reinforcement learning, the model's thought process in solving card puzzles starts human-readable but gradually becomes indecipherable. Like Opus5, it also emits 'AAAAARGH' screams, suggesting a common cry of overly intelligent AI.

• Mythos
• Fable
• system card
• 强化学習
• 解读不能