I spend years working on training these models. Inference is always the fruit. The effort going into getting the data is the most time consuming part. I am against meta from a long time. But open sourcing the weights help move the field in general. So I have to be thankful for that.