Anthropic researchers make progress unpacking AI's 'black box' - Fortune

Anthropic has pioneered a new method for revealing how Claude and other LLMs 'think,' hinting at why the models hallucinate and can be jailbroken.