Volume 2, Issue 9
Can We MCP? Asking for a friend (Epidemic)
This is still a first, as far as I know. Epidemic Sound released a connector for their MCP server in Claude, allowing catalog searches, previews, download links, and version adaptations on demand. I recorded some stupidly simple jazz stingers (ii-V-Is) for social media recently, so I connected Claude to Epidemic and tried findin’ some of what I was writin’. I had a good experience, and I think I would have gotten usable tracks if I were going to use them for a project. Interestingly, the connector suggested that it could trim a stinger from a full-length track, if that would suit my liking, when it ran short of suggestions. That seems like evidence of the adaptation. So overall, it seems great. Does it work day to day? I’m not as sure. I’m going to guess that really experienced music supervisors will work faster around it, unless they are looking for something hard to describe. I say this because that is my experience. There are many things—writing is one of them—that I just can’t optimize with AI at this time. It takes longer, yields worse results, and doesn’t come up with weak, antiquarian headline puns. Or phrases like that.
Austria Leads with C2PA
The Austrian Parliament will now be using C2PA Content Credentials for all new video-on-demand content. It is using Big Blue Marble, who run a C2PA-conformant product. That’s a bit of jargon for you, but in C2PA parlance, being “conformant” means being on the conformance list. You get there by building a product that adheres to the specifications and other requirements and then passing through the conformance program. That takes effort, and while the list of companies that have achieved conformance is still relatively small, it is growing, and the growth rate is increasing. That is a sign that the market sees opportunities and is willing to invest in them.
The partnership is a pattern you are going to see over and over again. Companies that need a provenance solution are reaching out to the rapidly expanding list of conformant providers. These providers run the gamut from niche players to well-funded insiders quietly building big platforms. For many companies, it will not be worth the cost of integrating the raw technology and running it themselves (and gaining conformance) if there is a cost-effective, turnkey solution.
So congratulations to Big Blue Marble. And to the Austrian Parliament, we say herzlichen Glückwunsch!
Bruh, this is FIRE!
Meanwhile, over on social media, people are connecting GPT-6 Astra (or other AI assistants) with their DAW and losing their shit over it. I think these are fantastic tools, but I’m not sure what all the hype is about. I’ve been using AI to assist with mix and recording functions for quite a while now. The closer integrations mean that there is more direct work between the AI and the DAW, so it’s more hands-off for the producer. The biggest advantage there is that the AI assistant can directly measure things like meter levels, leading to better insight and control over the environment.
But it’s just another tool, people. A very good one, but if you watch through some of these, it’s not perfect either. And it can easily make bad decisions. And let’s not confuse sounding great with sounding like you want it to sound, which might be another version of great. Which makes me sound negative about it, but I’m not.
When you throw too much unguided AI at any problem, there are going to be issues. The biggest of them is the lack of focus. When you start tweaking a system under AI control, you tend to get a lot of unintended drift and occasional loss of context unless you’ve anchored it in a specification, or process, or other durable artifact. So if it’s a complicated mix, changing it may be difficult. And occasionally you realize that you could have started from scratch with an AI assistant instead of a live agent, and then you might still be in the same place, but it would have taken less time and you’d actually understand how it all worked.
But if it were me, and it might be soon, I would spend a few hours defining exactly what I want out of a mix in general and then augment that with specific requests for the tracks I’m working on. A specification for the AI agent to work within. That could be a prompt or a reference document. But something that ensures that the AI doesn’t do something stupid or that I don’t want. There I can document my preferences for laying out tracks, labeling things, plugin usage, bus strategy, and so on. These videos are all sensational, but they don’t necessarily represent a real workflow. My advice is to build and maintain a real workflow using the agent, rather than just running the agent.
It’s also worth noting that AI-mixing itself doesn’t trip the AI content labeling recommendations recently put forward by the industry as long as it doesn’t introduce or operate on any generative content. I am not a lawyer, but most mixing operations are safe. Stem separation is one area to watch out for. Some stem separation uses generative hallucination to create a stem from samples of the original rather than extracting the existing performance from the track. Those are the ones to bring to your lawyers.
If it’s Tuesday, this must be … another AI rights legal update?
There have been some important events regarding the legalities and ethics of using unlicensed materials to train generative AI models. First, there is an important piece in Zinstrel entitled UMG Is Targeting Both Ends of the AI Music Pipeline. Obviously this piece is about UMG, but part of the story is the litigation throughline. UMG is suing DistroKid and in a bonus move for us litigation watchers, they have joined with Sony to sue Suno again. Neither you nor Suno can catch a break. You, because you put up with my alliteration, and Suno because even though they trained new models with licensed material, UMG and Sony, who didn’t license any material, believe that the new models also trained on some music that was generated by the old models. And those models allegedly used the original catalog recordings from UMG and Sony. So the new models are poisoned by the outputs of the old models, which were poisoned by the inputs of the original catalog tracks. Are you still with me? In the end, those new models are still allegedly infringing on UMG and Sony’s rights, because the value of those catalogs was passed on in the inputs from those old models made with the very same catalogs.
And DistroKid? Well, the details are pretty damning. They accuse DistroKid of a lot of things. Monetizing unlicensed and manipulated tracks. Gaming the system by saying they are for human artists but allowing AI music to flourish. And perhaps the most pernicious—when YouTube or another platform triggers an infringement warning, DistroKid will acknowledge it but keep distributing the track to all other platforms.
But back to UMG. They made a huge move last month when they revealed they had obtained 24+ patents across many patent families with an additional 50+ in process. These patents cover the entire AI production process, from data gathering to watermarking and tracking to generative methods to licensing and monetization. I need to do a bit of a mea culpa here, because I should have written about this last month. I didn’t because I had deadline pressure and didn’t fully understand the implications of the news and hence was hesitant. On the other hand, I’m not really a news outlet, and I frequently don’t cover big stories if they are kind of boring and mainstream. But next time this happens, I’ll at least mention it.
Anyway, this news also included partnership announcements with Udio and GRAI where they are licensing the aforementioned patents from UMG (er, actually a joint venture with Liquidax Capital named “Music IP Holdings”. Hey, at least they aren’t spending their money on brand consultants, right?) And all of this brings us back to Marcus Lawrence on Zinstrel.
Marcus does a much better job of connecting the dots by talking about how UMG appears to be releasing a framework for AI music and attempting to enforce it at the same time in a not-so-subtle attempt to show the industry how to do it right by their rules. We are going to have to watch that closely.
Beatoven Can’t Compete
Near the end of that story is the really disappointing news that Beatoven.ai has given up on consumer services and is working on licensing their IP. According to co-founder Mansoor Rahimat Khan,
…our original approach of ethically licensing music and building foundational models from scratch faced strong headwinds as heavily funded competitors emerged, building models through data scraping…
So there you go. It feels like some real, present-day impact where a company trying to do the right thing is forced out of the market by a few players with a fantastic arbitrage hack. That’s what it feels like to me. And it stings. Yes it does.
On a lighter note, I typed Udio into Google to get the right domain name, and Suno was the top paid listing. I realize this is probably par for the course for an aggressive, well-funded startup, but I still wonder how much they are paying for those hits. $5 per click? Maybe $10 or more?
The Other Side of the Coin
I want to be open to all viewpoints, so I follow the AI Music Creators Association. They advocate for AI artists and music. Their goal is “to ensure AI music creators have a voice, are treated fairly, and can thrive in the future of music.” I applaud this goal. I clearly don’t agree with all of their positions, but I absolutely believe that generative music is here to stay, and there will be plenty of AI artists that will thrive. If you want a quick take on their platform, check out their issues page. Whether you agree with their perspective or not, the eight issues outlined there are ones that we should be talking about and solve as a community.
Leave a Reply