The pre-training it’s had though (the GPT stuff) is where the weightings are. And they’re proprietary so MS isn’t going to tell you what they are. Yes, the data is yours. But the model that operates it is theirs and nobody has any idea what those weightings are.
There is the coincidental problem that the historic AFP data set for the past two decades leans heavily on non-white people doing terror things, and poor people being seriously over-represented in the data set. It’s not going to be as useful for - for example - white collar crime.
God they are so close to getting it aren’t they, like yes guys keep thinking, you collect too much data…. yes I understand and that’s a problem… and what could we possibly do about that?
Comments