Astonished that it generates background sound effects for scene narration. I suspect the audio tokenization has advanced acoustic and auditory modeling in both directions, but still uncertain about articulatory features. It's obvious that full auditory i/o seems too unwieldy for the corpos to risk unleashing it. For Gemini Project Astra for example https://deepmind.google/technologies/gemini/project-astra/ they are probably worried about the audio analog of DEIing historical figures like it did with image generation, or something like that.