facebook.com
16 reactions · 6 comments | It is becoming clear that local battles over data centers may soon be less relevant if the pace of model efficiency continues at its current pace. Attached are three video clips that were made right here in my house with an off the shelf 32GB Intel B70 GPU card you can buy on Amazon alongside the open weight MiniMax H3 video generation model which comes from, where else, China. These clips were generated with H3's "reference" template. I gave it a still image of me on my set and asked the model to replace me with Captain Picard and Jerry Seinfeld using a script I provided. What's even more remarkable is Picard's phaser rifle - it rendered that based off just a 2D still image I uploaded as a second reference file. The likenesses of each (along with their voices) were already contained within the model's training data. While this is an excellent visual representation of local AI model quality, the local language models (similar to ChatGPT, Claude, etc) are moving at an equally expontential path. A local model I have running on this very same hardware is able to do most general consumer AI tasks along with some remarkably good coding capabilities without ever needing to utilize a datacenter for the task. | Lon.TV
It is becoming clear that local battles over data centers may soon be less relevant if the pace of model efficiency continues at its current pace. Attached are three video clips that were made right...
