Shoot yourself in the foot.

§01 The perfect phrase

There's an English expression for what the music industry is currently doing to itself, and it's the right one: shooting yourself in the foot. Not losing a fight. Not being outgunned. Pulling the trigger while the barrel is pointed down.

The campaign looks like winning from the inside. Lawsuits against the generators. Pressure on the platforms. Licensing deals that quietly convert an open frontier into a walled garden with a turnstile the majors happen to own. Every quarter another door closes, and from a boardroom that reads as progress.

Here's what it actually does. It sends a large number of curious, motivated, technically capable people out to look for an alternative, all at once, all in the same direction. Closing a door does not delete demand. It relocates it. And the place it relocates to is the one place nobody can close: weights you can download.

§02 They already ran this experiment

In 2001 the industry killed Napster. It was a clean, total, unambiguous legal victory over a single company with servers and a mailing address. And what did the win produce? Gnutella. Kazaa. eDonkey. Then BitTorrent, no company, no central index, no defendant. They shot the one target that could be negotiated with and got a decade of targets that couldn't.

What eventually ended mass music piracy was not a lawsuit. It was a better product at a fair price, a store where a song cost less than the effort of stealing it, and then a subscription that made the whole library cheaper than the hassle. Convenience won. It always does. The lawsuits bought nothing but time and legal fees, and the industry has apparently filed that lesson under someone else's name.

§03 We were a paying customer

This isn't theory for us. We ran this label on commercial generators, and we paid for them, monthly, for a long time.

Take Suno. It's still there, but downloads got limited, models got dropped from under people who were using them, and the site is rebuilt so often that the automation we wrote against it breaks more weeks than it works. That's not a company acting confident. That's a company being squeezed from above, negotiating for its life, and passing the pressure down to the people who pay for it. So we stopped paying.

We're one small label in Skåne, so who cares. But we are not a rare case, we're an early one, and the mechanism is what matters: when the licensed product becomes the annoying product, the unlicensed one wins on merit. Not on ideology. On merit. The industry is currently spending enormous effort making the compliant path worse than the non-compliant path, and then acting surprised when people walk.

§04 There's a graphics card in the next room

Let's be honest about where open weights actually are, because pretending otherwise is how you lose credibility. Until very recently, the open music models you could run at home were not good. They produced music roughly the way a four-billion-parameter text model writes a scientific paper: the shape is right, the surface is convincing for eight seconds, and everything underneath falls apart. We used them. We know exactly what they couldn't do.

That's a snapshot, not a verdict. We have watched this exact curve twice already. Small open text models went from party trick to genuinely useful in about two years. Open image models went from six-fingered nightmares to production tools in roughly the same window, and they did it on consumer hardware, driven by hobbyists, because the weights were out and thousands of people could iterate at once.

Then look at how fast that first paragraph is going out of date.

Read that card twice, because it is the whole editorial in a box. A free download, on a gaming card, arguing about benchmark decimals with the flagship paid product. Discount the numbers as much as you like, every vendor grades its own homework, and in our hands it still isn't quite the commercial state of the art. It doesn't matter. The moat was supposed to be years wide. It's a version number or two, and it's closing from the free side.

Ours runs on an RTX 3080. A gaming card from 2020. That's the whole argument in one sentence: the hardware requirement is not a data centre, it's a machine that already exists in a few million bedrooms. And once weights are released, they are released. There is no takedown, no subpoena, no licensing negotiation and no terms-of-service update that reaches a file sitting on a hard drive in southern Sweden.

§05 Fifty Nirvana albums by morning

In 2021 a Toronto mental-health organisation put out Lost Tapes of the 27 Club, new songs in the style of Nirvana, Hendrix, Winehouse, Morrison, to make a point about musicians and mental illness. The Nirvana one was called Drowned in the Sun. It took a team, a partner, a research model from Google, and a lot of hand-finishing to produce one song.

Now project that forward onto a good open model on a home GPU. One person who loves a band gathers the discography, fine-tunes overnight, and wakes up to fifty albums that never existed. Not one song, a discography.

And you don't have to take our word for the timeline. The card above quotes ACE-Step's own documented fine-tune: eight songs, one hour, on a 3090. Eight. Nevermind is twelve tracks. The entire training set for a convincing imitation of a band is smaller than the record you already own, and the compute bill is one evening of a card that's five years old. This is not a threat and it isn't a fantasy, it's arithmetic, and the only variable left in it is model quality, the one variable that has never gone backwards.

And we can vouch for the overnight part from experience, not proudly. The 128 ambient albums in our own catalogue exist because someone here handed a long-running agent the instruction make some ambient albums and went to bed. It did not stop. We woke up to a hundred finished records, written, mastered, sleeved, titled. Nobody was watching, and nobody would have thought to interrupt it at album forty. That was one evening, one loosely worded sentence, one beer, and a generator that had been fine-tuned on nothing at all. Point the same evening at a band instead of a mood.

Here is the part we'll say plainly, because it's the honest complication: selling that as the actual band is fraud. Name, voice and likeness are real property, misrepresentation is a real harm, and someone passing off a synthetic Cobain as genuine deserves to be sued into the ground. That's exactly what happened to Heart on My Sleeve in 2023, the fake Drake track came down not because the music was bad, but because of who it claimed to be. That's the law working correctly. It polices the label on the tin, and it can, because the tin is the part with a name on it.

The math in the file is a different matter entirely. You cannot police that, and every hour spent trying is an hour not spent on the part you actually could enforce.

§06 The part that doesn't come in a checkpoint

Which raises the fair objection to everything above. If the flood is coming and anyone can make fifty albums by morning, what exactly is a label for? Why are we still here?

Because the model is the easy half. Hand any generator a prompt and it will sing you four bars of stress-wrenched, cliché-stuffed filler with immaculate production values, instantly, forever, for free. It will do that for a million people at once. What it will not do is know that the line is bad. The distance between competent output and a record you'd release is upstream of the model, in a text file, prosody, the rhyme you didn't force, the syllable count that lets a melody breathe, the hook written before anything else, the album that's a shape instead of a playlist.

That text file is the label. It is the one part we do not publish.

Look at how that pipeline is shaped and you'll see the argument again. Load the artist profile, load the genre preset, settle the language, build the album arc, write the hook first, write the song to the scheme, sound-steer it, self-check it, write the files, and then, at step ten, hand it to something that makes audio. Nine steps of craft, one step of model. When we walked away from Suno, we lost step ten. Steps one through nine didn't notice.

That's why we can give this away without flinching, and it's why the majors' strategy is upside down. They're defending the machine end, the part that's becoming free, and treating the craft end as though it were a rounding error. It's the other way round. Craft isn't scarce because it's hidden. It's scarce because it's hard. Publishing ours costs us nothing, because the only way to use it is to actually do the work.

§07 Almost everything you're wearing was made by a machine

Look down. Whatever you have on was cut and stitched at industrial scale by machines, and you have never once felt bad about it. Nobody checks the label for a human touch before putting on a t-shirt. The looms won that fight two centuries ago and the argument is so over that we don't even remember it as an argument.

And yet the tailor did not die. Bespoke didn't vanish; it got rarer, more expensive, and more prestigious. Designers have more cultural power now than at any point when clothes were made by hand out of necessity. What machine production destroyed was not craft, it was the middle: the anonymous, functional, competing-on-volume tier. Craft moved upmarket and got better paid.

Music lands in the same place, and quickly. The floor floods with infinite competent output that costs nothing. What survives above it is what a machine cannot hold: a person on a stage in a room on a Tuesday night, a scene, a face, a story, a catalogue somebody curated with their whole taste on the line. Note that the industry could be building for that world right now. It's building lawsuits instead.

§08 Horse drivers

Britain's Locomotive Act required a self-propelled vehicle to be preceded on foot by a man carrying a red flag, at four miles an hour, so as not to alarm the horses. It stayed law for thirty-one years. The cab drivers who backed it were not stupid and they were not wrong about their own livelihoods, they were right about their lives and wrong about the outcome, which is the most human way there is to be wrong. The flag came down in 1896. The horses did not come back.

That's the whole shape of it, again. Incumbents can't stop a machine, so they legislate a limp into it and call that safety. It buys a few years for the people who own the horses, and it costs them the transition. Meanwhile someone with no stable to protect builds the truck.

What's actually being defended here isn't musicians' income, if it were, the money would already be flowing to musicians. What's being defended is the position of gatekeeper: the right to decide what gets made, who gets heard, and who gets paid on the way through. That asset was always the real product, and it is the one thing a downloadable model deletes outright.

We'd take the other bet. License the training data, take the cheque, put the tools in every artist's hands, and compete on the things that don't come in a checkpoint, identity, curation, a room full of people. The flood is coming either way. The only open question is who gets paid on the way in.

Because they think they're aiming at a handful of AI companies. They aren't. The thing they're afraid of is a file and a graphics card.