(1/2) By now you've probably heard about #Sora and the latest wave of #AI-driven text-to-#video tools. It's all quite spectacular, but far from perfect. If you take a closer look, it's still possible to detect synthetic content. Essential pro tips: Play everything in slow motion. Consider real world physics – and logic.
@dw_innovation Marcus' observations are spot-on with respect to the lack of spatiotemporal continuity in all these new models, this was the first thing I noticed while watching the demos last week too: the represented worlds seem as consistent as a dream.
You can test this same issue easily by playing/simulating Crowther's Adventure or similar with ChatGPT: move around different rooms and things will change. This is of course not surprising if you know how these architectures work.
@dw_innovation I do wonder though how many people do notice the Escherian spaces, continuity errors, odd motions, etc. Maybe our society won't care about these details, similar to how we accept the limitations of modern tech and resulting in solutions like lossy compression: not many care about artifacts, reduced colour space and so on.
That's an interesting thought. In the field of verification / disinfo countering, there's also the (well documented) phenomenon that people don't really care about the fact that something has been thoroughly analyzed and debunked.
"Well, maybe it's fake, but it could be true."
"It's a good story!"
"Yeah, well, maybe this is not her, but I still think she's a terrible person."
Add comment