Posted 2/29/2024, 4:32:46 AM
Google Upgrades Music AI with 300,000 User Preferences, Foreshadowing Industry Disruption
- MusicLM is Google's text-to-music AI that was recently upgraded using a technique called reinforcement learning from human feedback (RLHF)
- RLHF involves training an AI on human preferences, allowing it to generate outputs users are more likely to enjoy
- Google gathered over 300,000 user music preferences through an online MusicLM demo to train their reward model
- The upgraded MusicLM, now called MusicRL, quantitatively outperforms the original model in direct comparisons and mean opinion scores
- Musicians should learn about AI as advances like MusicRL foreshadow major industry disruption in the next 1-3 years