Post B86Iy1mGB2T67s9NxY by unfa@mastodon.social
 (DIR) More posts by unfa@mastodon.social
 (DIR) Post #B86Iy1mGB2T67s9NxY by unfa@mastodon.social
       0 likes, 0 repeats
       
       Generative "AI" is very similar to JPEG compression. The basic idea is you recognize patterns in input data, you encode the recognized ones, and then you reconstruct an approximation of the input data using the encoding and stored patterns. Only in JPEG there's a very small amount of patterns, which are little waves that are layered on top of each other to reconstruct the image in chunks. In ML there's billions of convolving patterns, so the comp. ratio is incredibly high.Fancy JPEG ate life.
       
 (DIR) Post #B86Iy26owaun9dlo0W by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @unfa one of my problems with GenAI is not that by removing details it removes information, it's that it adds back details that weren't there, that - if added by human - would've had to be deliberate, conveying extra information, but here they don't convey any information.