Companies are willing to pay a lot for clean training data, and my bet is there will be a growing pile of training sets for sale on a non-exclusive basis as well.
A lot of this data - what I've seen anyway, is far cleaner than anything you'll find on the open web, with significant data on human preferences, validation, cited sources, and in the case of e.g. coding with verification that the code runs and works correctly.
Very interesting, thanks for sharing that detail. As someone who has tinkered with tokenizing/training I quickly found out this must be the case. Some people on HN don't know this. I've argued here with otherwise smart people who think there is no data preprocessing for LLMs, that they don't need it because "vectors", failing to realize the semantic depth and quality of embeddings depends on the quality of training data.