points by NothingAboutAny 2 days ago | 0 comments
I'm not so sure, I've made an autobattler and even recently using Astra to come up with new mechanics or cards they're usually always bad as if it doesn't "get it" even if I feed it all the current cards in game and say "like this, these are good because of x and y."
When brainstorming, I usually get it to generate 20 ideas and then I come up with something entirely new based on bad stuff it responded with. I've never been able to actually use its ideas verbatim.
I'm not sure if this is a limitation of the tech or whether it will eventually have "taste" in such nebulous domains as "what is fun." They still can't write funny jokes despite trillions of dollars of training because even though there's plenty of funny material and discussions on what funny is, I guess it's still kind of a human thing.
Currently, if you want better results, you have to prime the model with what I like to think of as a Context Mood Board consisting of material that you select to express a gestalt. Turning reasoning off also helps since it forces it to rely on what's already in the context.
This is a lot of manual work to set up but you can save it as something that gets loaded when needed. For open weight models you can create a LoRA based on a dataset that represents what you consider to be good taste.
The point is that it is possible to do even though it's currently not available from the major providers. They're more focused on maintaining their lead in capability, taste takes a backseat. At some point it will become more important and then we'll probably see either a curated selection of styles to choose from or they'll just keep them hidden and automatically route you based on your request and your user profile.
If you want a preview of the difference this can make try out some image generation models in ComfyUI and compare the results from the same prompt using the original base models versus the fine tunes and LoRAs created by the community. It's incredible what a difference it can make in the output and what you're seeing is the design sense or artistic taste of the person that did the fine tuning and the artists whose material was selected as input.
Looking even further ahead, there's a huge amount of undiscovered talent out there in people who lack the technical ability to produce art but have a very good sense of taste. Somehow getting access to their preferences would advance things even more. There's still a lot of training to be done so we're hardly at the limit of what is possible for instilling taste or artistic sense in LLMs.