DeepSeek cuts cache to 890 bytes per token, 437x smaller than V1, with Flash scoring 39 on AA Index at $0.27 per task versus V4-Pro.

According to Soumith Chintala, an efficient wrapper around FFmpeg for PyTorch has been developed, utilizing FFmpeg's fast seeking and read-ahead APIs correctly. This wrapper also optimizes memory buffer usage, avoiding unnecessary allocations and copies, which could significantly enhance video processing tasks in machine learning projects.