Google's advanced music generation model and two new AI experiments

It's really interesting for me to see AI voice cloning taking off. A few years ago I worked on a paper called HuBERT (https://arxiv.org/abs/2106.07447) which has since really taken off for doing this kind of stuff (it's the speech representation that SoViTs uses, for example). At the time the main research focus was low-resource ASR; it was sort of a fluke that it works so well for voice conversion. It makes me feel confident that open-source ML will drive more value over the next five years than closed-source ML, despite all the hype around big tech, simply because random people building weird, unexpected stuff on top of foundation models will make things that no one could have predicted.

Some of this generative AI stuff is going to raise some interesting questions. Right now a lot of it is rough around the edges and distinctly inhuman. However, I do believe that even if we're not entirely on the right track, there's not much reason to believe a machine can't do human-like things as well as humans with enough computational power.

Here's my guess: I think that this stuff will continue to iterate until it's to the point where the amount of control people have over it as well as the quality of handling edge cases is sufficient to allow people with fairly unsophisticated tools to create things that look fairly high quality. And once this happens, people armed with amazing tools but no better taste than they had before, will flood the Internet with a bunch of tasteless garbage at rates never before seen. It hasn't happened yet, but I think this is partly an accessibility issue (not everyone has datacenter GPU access, there's still limitations in open models, and many hosted offerings are pretty limited compared to what local models can do right now in terms of tooling) but also it will make a big difference if we hit the point where it's really genuinely just not possible to reliably distinguish AI generated images and sound. Then, even if people wanted to ignore AI generated media, it would probably quickly become pretty difficult, and many people will probably not be honest if it limits their reach and potential.

It will just raise the bar of entry. Tolerance for generic garbage will decrease. Elevator muzak producers are in trouble. Real talent will still float to the top.

Michael Jackson was an excellent musician, who just happened to not play any instruments. He'd hum lines, beatbox rhythms to his band who'd then take his ideas and transform his ideas into instrumental music. Most people don't have access to the quality musicians MJ had, but there's a lot of creative people that lack the skills to create their artistic vision. Tech like this could be the unleashing of an age of new genres and musical creativity. I'm optimistic.

Hahahaha poor naive people that accepted to "collaborate" with Google, Google has taught us how their seemingly benign acts always end up as they started, right?

Nothing more trustworthy than a multi billion corporation trying to destroy the internet as we know it to get more billions out of ads convinces a bunch of washed up artists that their new project is benign and even are so good that they use the hashtag "Social Responsibility".

This is very totally not just a plan to try to get others on board, wonder what's the catch here? try to do a FOMO where only the first few get to get rights for their voice being used? playing a very long con getting voice exclusivity rights? or what exactly?

One thing I know is that whatever Google says is a lie.