Emergent introspective awareness in large language models