I think text/code genAI still have space to grow, but it's the same space fixed image generation models had two years ago. The usage grew, the costs diminished, the accuracy slightly increased, but the weirdness will stay, and gross mistakes are still there, at a similar rate.
This is not true in the slightest. Accuracy and weirdness are clearly getting better and better for both images, ai and articles.
Remember will smith rating spaghetti? If you ignore the trend line then that is utter blindness.
That being said you can make an argument that AI will flatline in the future. But your evidence for that does not lie in past progress as the past shows a trend line that is clearly upwards.
I said accuracy was sightly better . The improvement isn't that high though, not enough for me to update my images.
Weirdness is still there (when you want photorealism), and i think the rate of gross mistakes (at least for stablediffusion) is around the same ~ one in ten, since at least a year (i don't remember what i generated in 2024, but i think it was around the same too). And it still can't get historical generation right. I haven't seen once accurate clothing and architecture when you want to go historical scenes. It's great for fantasy though, which is my main usage, so i didn't wrote about it, but it definitely something that hasn't improved at all. If you want to argue its sightly less weird, i can listen to it (i think it's more desensitization more than anything, but you can have arguments). If you think the rate of gross mistake went from 1/10 to 1/12, maybe you're right and i'm too negative. But historical mistakes in scenes just never change, it is still extremely wrong.
Fixed images. As in static images. Of course i'm not talking about Sora or videos. I was talking about static images. Stablediffusion and other image generation peaked after two years. My original point was that static images peaked in 2024 and improvement have been minimal since then. "fixed images".
