I keep seeing people say AI image generators are getting smarter but that's not the right word for it

What I notice is folks praise models like DALL-E 3 for "understanding" prompts better when it's really just more training data and better tagging. I work with Stable Diffusion almost every day at my shop in Austin and the jump from version 1.5 to XL wasn't about intelligence, it was about bigger datasets and more compute. If AI actually understood what a "warm sunset" meant it wouldn't give you orange blobs half the time. Does it matter if we call it smarter vs. better trained, or is that just splitting hairs?

1 comments

1 Comment

stellam891mo ago

Well, you've hit on something that bugs me too. I remember when my grandson tried to explain to me that his video game characters were "learning" from him, and I said no, the programmers just wrote code that makes them react a certain way based on your inputs. It's the same with these image generators - they're just really good pattern matchers, not thinking beings. My oldest daughter works with data tagging for a living and she says it's all just about feeding the machine more and more labeled examples until it spits back something that looks right.