Raw LLM Responses

0:35 A dataset creates nothing. A model trained on the dataset can create Images. The dataset is roughly 240,000 GB - the model is roughly 2 GB. A direct copy would be, according to physics, impossible - there is nothing in the model that is for the purpose of making copies, in fact it is quite the opposite as it deliberately adds noise as perquisite so the chance to find similarities is according to recent studies well below 0.033% 2:50 The actual version of glaze that works is extremely visible on the image. Also we don't even know if the same is going to happen in base model training. What they have done is train single or a few pictures and looked for a difference. But those are not the same circumstances when base model training is happening, the process takes weeks, if not months and billions of pictures. You can't simulate that, because every time you would want to test if Glaze worked you had to train a new base model, meaning buying time in a big GPU cluster and letting it run for a month for over half a million dollar. They are promising something that can NOT be tested by anyone - unfalsifiable research is just a hypothesis - they believe it does something, they don't actually know. 4:30 Stability trains in Germany under EU Law, where machine learning laws already exist. See Article 2, 3 and 4 (commercial use) in Directive 790/2019.