Commit 1ef56b
2025-08-08 17:18:09 Doku: -/-| /dev/null .. journal/2025/08/08.md | |
| @@ 0,0 1,18 @@ | |
| + | # 08 |
| + | |
| + | I like to think of way to find relations between pieces of media. |
| + | |
| + | There are quite a few things I have to say about this, so let me summarize them before I start. |
| + | |
| + | 1. Numerical ratings are nice, but there are some obvious dynamics that make them insufficient. There's the common problem of a single 10/10 rating making the rated thing appear on top when sorted by average, but that's solved with smarter formulas. Still, a single rating is too one-dimensional and multiple ratings don't help much because they mostly end up correlating so much. |
| + | 2. Recommendation algorithms are a huge part of the internet, and many incredibly smart people have been thinking about how to recommend the perfect thing to a consumer. |
| + | 3. Tags help, but they are also limited, and they are too technical. |
| + | 4. TV Tropes is an amazing effort at building a complex structure of ideas related to pieces of media. It has its limits, but they're very high. |
| + | 5. Going back to one directional ratings. If it's a website like IMDB with thousands of votes given to every single item, then maybe averaging them out works. But it doesn't work as well when there are not enough ratings. Either because the media is more common (like images or songs) or because it's a personal thing and not global. |
| + | 6. One of the ways to rate items in a one dimensional list is something like Elo score - make pairwise comparisons and then see which items scored best. |
| + | 7. One of the problems with all this is that media is so multidimensional. There's the plot, with its own dozens of dimensions, there's the visuals (if it's something visual), with its own. There is a general "quality" and attention to details that go through everything together. This complexity cannot be easily put into numbers. However, there are many other things lately that can't be easily put into numbers, that are being put into numbers despite all that. Because of AI, where you don't understand numbers but they still work. You don't understand how exactly the pixels correspond to the AI recognizing what it sees, but it recognizes the object on an image. So what if a similar idea could be applied to media? |
| + | |
| + | |
| + | This turned into a mess instead of a list. My current idea is the following. |
| + | |
| + | Let's take a vision language model, an AI that can see images and think. Let's make it look at pairs of images and rate them based on their aethetics. Let's use something like Elo to find the most aesthetic images in the eyes of the AI. |