Anthropic's warning over AI self-improvement has a hidden message — accelerating development requires more compute before companies ever risk losing control of frontier AI models
… METR , for example, found last year that the length of task an AI can finish with 50% reliability has been doubling roughly every seven months. …