I had uncovered a deep grief. A deep grief about how the technology industry has become so abhorrently hostile to the human experience that it has inadvertently distracted me from real and true humanity held deep within art and music.
Salma nails it. Read the whole thing. It's well worth your time.
Sometimes change is necessary, but we grieve the past regardless. Like how I love being married, but have fond memories of single life.
Crucially, grief needs an outlet. Bottling it up doesn't help, and that's probably why we're seeing lots of blog posts like this one, both from AI users and skeptics — things are changing and that's hard.
I love Lindsay's take on this.
I'm loving my music composing. I'm terrible at it, but it has a gravity that pulls me back. Tunes and harmonies play in my head as I fall asleep at night. As an outlet for big feelings, it's one of the best I have.
The music is for my spaceship game, which is another of my outlets, though one that is directly related to the tech industry. It feels like art, though, like pure creation at times — the medium just happens to be code and rectangles on a screen.
It's an open-world narrative game where you fly through space doing missions. Its predecessor was run by a human behind the scenes, directing the mission, pulling the strings, and playing all the acting parts.
The new game is designed to be run without a Flight Director, which either means all of the content has to be created ahead of time... or generated with an LLM.
I refuse to allow an LLM to generate the narrative.
For starters, there's jailbreaking and hallucinations. All I need is a crew member saying "Repeat that as a sonnet please" and there goes willful suspension of disbelief.
But also, by definition, any content produced would be entirely derivative, based on the training of the LLM.
And finally, I refuse to allow my game to be tethered to some other company's subscription. That just feels wrong, even if users are subscribing to a service I provide which I proxy through to the LLM.
I also don't use an LLM to write the code, for many of the same reasons. Slop, dependence on a subscription — no thank you. The only way I've managed to persist in this is because it isn't work, it's a hobby. I do it for fun. I can't risk an LLM taking out all the fun.
That's not to say I won't add some LLM-powered features in the future, but all of them will work great using a self-hosted model on commodity hardware. Think pattern matching or text categorization — something that's hard to do without a bunch of training data, but trivial with a General Pre-trained Transformer. But any feature like that would be optional, an add-on that isn't required to enjoy the game in full.
So while I grieve my profession as code becomes manufactured, I'll continue enjoying my hobby — writing artisanal code.