I was discussing this with a (visual) artist this morning. She was pushing the idea that there’s lots of space to explore with small, narrowly-trained models that do a specific thing.
The “LLMs for everything” approach is really an artifact of hundreds of billions of dollars flying around. The Facebook/Google/MS corporate gang require the answer to be “come to Facebook/Google/MS for all your needs”. Every solution that doesn’t scrape the entire Internet has been left in the dust, because it doesn’t put all power in the hands of megacorporations.
The problem (“problem”) is that to make a narrowly-trained model which does exactly what you need, you have to understand enough about AI models to create it! Shoving prompts into ChatGPT simply will not get you there. You have to approach the problem the same way we tell people to approach IF tools: it’s programming, and you’re going to have to learn it, and you’re also going to have to hold onto your artistic vision and figure out how to mold the tool to your needs.
…but we as a community aren’t building those skills, because the Facebook/Google/MS axis has such repellent business practices that they’ve turned most of us off.
(I very much include myself there.)
Anyhow, to get back to the original suggestion: a contest is only helpful if there are enough interesting experiments in the air to fill out a contest slate.
Possibly the upcoming IFComp will answer that question.