V4.1-Flash 的 KV cache 相比上一代仅需 1/4 的 HBM 和 1/8 的 SSD 存储。由于缓存命中费用通常占智能体成本的较大部分,压缩缓存可显著降低相关成本。 · V4.1-Flash’s KV cache requires just 1/4 of the HBM and 1/8 of the SSD storage used by the previous generation. Because cache-hit charges often account for a large share of agent costs, the compression can significantly reduce those costs.
Navier-Stokes 工作中没有任何人类或智能体查看用户数据。ChatGPT 和 Codex 会使用用户反馈及去标识化数据进行整体改进,Mark Chen 称其他大语言模型公司也会这样做。 · No human or agent viewed user data as part of the Navier-Stokes effort. ChatGPT and Codex do use user feedback and de-identified data for broader improvement, and Mark Chen said every LLM company does the same.