Another person already wrote this itt, but the cost in time and money of building something that “works” is very low. That cost increases every time you need to fix something. It’s why I’ve switched to deepseek flash instead of the bigger, better coding models. For the same price as a million tokens of Claude for example I can get a hundred million tokens on flash. It takes three times as long and needs a little prodding but it’s thirty times cheaper.

Even if you’re not dealing with ballooning costs when it comes to upkeep, the present environment around ai programming is filled with perverse incentives that reduce the chance something is open sourced. Why preserve the old work and try to carry it into the future when you can just rewrite it? Why create standardized libraries and approaches to specific problems when it’s only gonna be used once? Why make it open when you don’t have time to deal with requests and bug reports you don’t care about? Why use gpl when you don’t intend to make it public anyway?

Those were all arguments against gpl and open sourcing parts of the Unix codebase too, btw.

Harnesses or agents try to address some of this by sharing improvements but that’s three layers removed from directly turning fossil fuels into cpu cycles to make the same program a thousand times slightly different for a thousand different people.

I tend to see the ai programming defense in the same way you described yourself and the “ai is bullshit” crowd. People defending this stuff generally either aren’t looking where they’re walking or haven’t had to use it long.

As a decently prolific user of ai programming, it’s turned computer code into disposable plastic wrappers. Even if they don’t persist and make a giant garbage raft in the ocean or calcify our pineal glands we still wasted a bunch of energy on them.