It's a good question. In this case, we had already open sourced the code in 2022 (as a companion to the 2021 Nature paper), but we'd observed that it wasn't very easy for some in the community to take advantage of that because they had to find interesting blocks to pre-train on