
This is a true story: in 2007, Ringtones were a billion dollar business.
The history of this industry is full of technological advances that – at least first – include a few (seemingly) inherently difficult things at the margins, and those rough edges give rise in turn to a sometimes large-ish service industry of companies that want to make that difficult thing simpler or easier, and make a few bucks doing it.
Sometimes – for a while, at least – there’s real money to be made in businesses like this. But soon those edges get sanded off; devices get a more powerful, their interfaces and designs settle out or standardize as user needs are better understood. The tools get easier to use and all the things that were difficult before become easier, sometimes trivial, for the uninitiated to navigate; pretty soon the market starts shrinking, and soon after that the market vanishes.
Earlier this week, a company called DeepSeek released new language model, in large and small sizes, that’s as good as either the big players’ and is claimed to have cost just six million dollars to make.
OpenAI’s current best model reportedly cost north of a hundred million dollars to train. Google’s Gemini, at least half that again.
Earlier this week, a company called DeepSeek released new language model in large and small sizes, that’s as good as either the big players’ and is claimed to have cost just six million dollars to make. If you’ve got a newish computer you can play with the DeepSeek model right now for free. The download is about the size of a DVD. Remember those?
Then, just a day later, a new – and this time fully-open-source model, open as in curated and consentfully-obtained training data, auditable and openly-licensed code, model weights, all of it – called Sky-T1 has hit the scene, with training costs claimed to be under five hundred dollars. You can run it on anything you’d call a decent gaming rig; even one of the higher-spec Mac Minis looks like plenty.
I say again: in 2007, ringtones – ringtones! Not even a whole song, just a few second of tinny audio carved out of a song! – were a one point one billion dollar business. They were ten percent of the entire music industry! And five years later that business model didn’t exist. Today you’d have a hard time convincing anyone under thirty that it had ever existed, the whole idea sounds ridiculous. I mean, does this chart make any sense at all to you?

One distinction in this whole exercise, obviously, is that people actually wanted custom ringtones. It was a market that existed because a technology was meeting a demand. Humans who wanted their interactions with a device to be a bit more pleasant paid money for something and received it. Artists got paid, if (as usual) likely not enough. The AI market doesn’t look anything like that, mostly because there isn’t an obvious way to use a ringtone as a weapon against the working class. But it’s a pleasant change of pace to see efforts to weaponize commodification against the working class and the forces of commodification themselves fighting each other for once.
Ringtones still have a place in the world; we don’t all always have our phones on silent. But that place isn’t a business, much less the industry it briefly was; the technology matured, the market evaporated and today custom ringtones are one bullet point on a long list of boring standard features, so universal they’re not even mentioned on the boxes of the cheapest phones you can buy.
There’s a place in the world for automatic pattern recognition and repetition, and mechanical translation, too. But that place also isn’t going to be an industry or a business; the technology is maturing fast, the market is evaporating just as quickly, and I think I know where the AI/ML space is going. These tools will be one bullet point on a list of features soon, one you build for a few hundred bucks with some carefully selected training data and a few dozen of last year’s graphics cards, because that’s all it takes.
Hiring really good librarians to curate your training data is going to be the hard part, good curators are expensive.