“A founding bet of Reflection AI was that pre-training for large language models was converging on a known paradigm, allowing the company to leverage open-weight models instead of investing in its own pre-training.”