“When existing evaluation benchmarks don't cover desired capabilities like creating slide decks, OpenAI's team creates new, internal evals to measure performance on those specific tasks.”