The tool that reads your inner ear

Juan opens Gemini’s new Create Music feature, powered by a model called Lyria 3, and runs it like a lab experiment: is it natively multimodal, or does it fake it ▸ 19:48? A simple prompt (a diss track about their game, “los demás son patéticos”) works well enough ▸ 19:58, and feeding it an image produces an acceptable song, but the deeper probes give it away: a rhythm he taps out, a Vivaldi score with the name stripped, both return something generic ▸ 24:10. His conclusion: it isn’t native multimodal, it converts the image to a text description and only then to audio ▸ 22:46.

The failure points at the prize. The definitive music tool is the one that lets you describe a sound and get it, the way directors and bandleaders already work, Michael Jackson humming a bassline until the guitarist matches it, Freddie Mercury pushing “faster, tr-tr-tr” ▸ 25:08. They hear it in their heads and transfer it; the tool that captures that inner ear directly is the endgame, and whoever has the vivid internal representation, the photographic visual memory for images, the composer’s ear for sound, will produce enormously when the interface finally reads it ▸ 26:40.

la herramienta definitiva lee lo que ya suena en tu cabeza →

The reason he isn’t dismissive is a pattern he’s learned to trust, the flip side of specialization: Google’s first swing at a category gets mocked, their early video model “couldn’t reach Sora’s ankles”, and then two versions later it dominates ▸ 24:01. Google explores, lags, learns the direction, “ya sé por dónde”, and closes the gap ▸ 24:26. So Lyria is a basic first version he expects to “break it, like it always does”, six months out, then a year ▸ 24:26. Today’s toy is the ankle; the leg is coming…