Ooh that’s a neat fact. I wonder how much training it would take for it to be able to make those images. I’m not sure the available models are capable of that right now. They have trouble with English text and I’m sure that they’re trained mostly on that. There’s an added wrinkle to Arabic in that GPT writes from left to right. From my understanding models like BERT, ROBERTA and BART don’t have that issue as readily, since they “read” bi-directionally. Pi does, as well. I don’t know enough about the back-end of DALLE 3 to know if it would naturally have extra issues with Arabic but my suspicion is that it would not, as I imagine it uses a diffusion model, which works by refining random noise.
In that case, it’s fine. I’m surprised nobody exploits it.
Ooh that’s a neat fact. I wonder how much training it would take for it to be able to make those images. I’m not sure the available models are capable of that right now. They have trouble with English text and I’m sure that they’re trained mostly on that. There’s an added wrinkle to Arabic in that GPT writes from left to right. From my understanding models like BERT, ROBERTA and BART don’t have that issue as readily, since they “read” bi-directionally. Pi does, as well. I don’t know enough about the back-end of DALLE 3 to know if it would naturally have extra issues with Arabic but my suspicion is that it would not, as I imagine it uses a diffusion model, which works by refining random noise.