Dust II von Counter-Strike läuft ausschließlich in einem neuronalen Netzwerk auf einer RTX 3090 – die Leistung ist mit nur 10 FPS enttäuschend

This is interesting, but mostly because it continues to demonstrate that using neural networks for generative tasks is a pretty terrible fit. An “AI” program can be good for *associative* tasks; comparing images, identifying objects, that sort of thing, and it can be good for *extrapolation* within certain contexts where a load of data is available already, like DLSS… but anything *creative* or requiring *understanding* or *thought* is simply impossible to derive mathematically. This program has ingested enough video of Counter-Strike to be able to predict how videos of Counter-Strike will look, within certain highly restrictive parameters, but you’ll never be able to feed it sufficient data for it to *know anything* about Counter-Strike.

From my position, it really looks like a lot of academics are very excited about a novel way to compress arbitrary types of data in an extremely lossy format. Like if you could use H.264 to encode multiple videos as one gigantic meta-file that you could navigate three-dimensionally, losing all the detail and meaning so you could blend seamlessly from one video to another.