According to CoinMeta, the Databricks team stated that they put GPT-6, Astra, and Opus into an automatic optimization loop, allowing the models to continuously rewrite, test, and improve GPU and kernel. As a result, they managed to rank first in all four categories of NVIDIA SOL to ExecBench. This benchmark included 235 GPU and kernel tasks, covering basic operators, complex fusion operators, low-precision calculations, and real large model inference. The model token spent approximately $70,000 throughout the process. The system is built upon KDA and Humanize, enabling AI to write its own kernel and conduct correctness and performance tests, making continuous modifications based on the results. The KDA team will also apply a similar method to Kimi Delta Attention in the near future, achieving 2.96 times the speed of the official FlashKDA on NVIDIA B300. These two achievements are part of different applications within the same agent core optimization framework.