Anthropic has finally lifted the veil on how large language models actually “think”