Comment by unleaded
1 day ago
Something that's confused/annoyed me about the AI boom is that it's like we've learned to run before we learned to walk. For example, there are countless websites where you can generate a sophisticated, photorealistic image of anything you like, but there is no tool I know of that you can ask "give me a 16x16 PNG icon of an apple" and get exactly that. I know why—Neural networks excel at fixed size, organic data, but I don't think that makes it any less ridiculous. It also means that AI website generators are forced to generate assets with code when ordinary people would just use images/sound files (yes, I have really seen websites using webaudio synths for sound effects).
Hopefully the boom will slow down and we'll all slowly move away from Holy Shit Hype things and implement more boring, practical things. (although I feel like the world has shunned boring practical things for quite a while before)
I just asked ChatGPT-5 to "give me a 16x16 PNG icon of an apple" and it did exactly that. It looks good, too.
Not that I don't recognize the inherent limits of LLMs, but there are as many edge cases covered as are found in the training sets. (More or less.)
well i just asked it the same thing and it gave me a 1MB 1024x1024 png with fringed edges & sensor noise that measures out to a 17x21 pixel image. https://files.catbox.moe/1q4jtp.png
In the time it would take to keep retrying until it makes one that fits, then reshaping it to fit into 16x16 nicely I could have just drawn one myself.
As you seem to understand, creating something that generally fits a description is the walking for AI. Following exact directions is the running. It may just feel reversed because of the path of other technology.