At EuroPLoP last week I learned of a popular Open Source gaming framework (I forget which one) had instigated a blanket ban on AI contributions. It appears that AI meant there were more submissions coming in and the maintainers were struggling to understand and review them.
I hadn’t had time to think through the implication of this before the Financial Times carried a big piece on the same issue across the Open Source world (paywalled: Who cleans up after the vibe coding party?). It actually turns out this problem could undermine many of the OS libraries we’ve come to depend on.
Next a friend was telling me about his dozen AI coded mini-products, one of them has eight or so prototype. None are finished, none in the market, all sound like good ideas and potential products.
All these examples are perfect illustrations of Jevons paradox in AI: when things get more efficient (cheaper or easier) then more is consumed. AI coding agents make coding cheaper so we do more of it.
But, as we have known for years: build it and they will come rarely, if ever, works.
Since LLM coding agents appeared the number of submissions to app stores has rocketed and the number of available apps increased faster than before. Yet, the number of app downloads has not gone up. Say’s Law does not hold, supply does not create demand, just because there are more apps written doesn’t create more downloads.
One can reason that the same pot of customer money is now being spread across more apps so the average contributor makes less money – maybe that’s OK because the apps are so cheap to produce. Or maybe with so many more apps to choose from people stick with the popular ones and there is an even greater gap between the few “winners” and the “also rans.”
These kind of problems have always existed but by lowering the cost of change AI has made them more common. Previously, the slower pace of creation mean that more questions were asked and there was more time to catch such issues. Now things happen so fast there is no time to ask and new problems are appearing.
Understanding what customers will pay for is more important than ever. Can AI help there? Perhaps with market research but someone still needs to make a decision.
So much of our world is concerned with managing, controlling and rationing limited resources – indeed the “allocation of scarce resources” is sometimes considered the underlying problem in economics. But increasingly we see problems of too much and we lack solutions for too much.
That is not to say we don’t have solutions. In the 1970s and 80s we had over production of some foods, there were wine and milk “lakes”, and butter and gain “mountains.” The EU had to start paying farmers not to produce.
Software product companies traditionally have an spend split of 33%/33%/33% (approximately): one third on technology engineering, one third on sales and marketing and one third on everything else, like finance, HR, property, etc. Many of those cost remain even if the cost of coding goes down. If there are more competitors then maybe more needs spending on marketing.
More companies that aren’t software companies will find they have software which needs to be supported in some fashion. More software demands more cyber security and until machines replace all the humans those that remain will need more training, and even AI written systems will require maintenance and updates. Then there is the cost of complexity from all those systems which nobody quite understand.
As to Open Source projects, I don’t know. Maybe it will be a shift back to proprietary products, or maybe the submitters need to be charged.
Jevons paradox is working exactly as expected. Resolving the problem in one place is creating problems elsewhere. Problems are truly generative.