According to DeepSeek, the NSA (Natively Trainable Sparse Attention) mechanism is designed to improve ultra-fast long-context training and inference capabilities through dynamic hierarchical sparse strategy, coarse-grained token compression, and fine-grained token selection, potentially enhancing trading algorithms by increasing processing efficiency and reducing computational load.

According to Paolo Ardoino, the future of AI model training will not rely on the brute force of 1 million GPUs. Instead, the development of better models will significantly reduce training costs, emphasizing that access to data will remain crucial. Ardoino suggests that inference will move to local or edge computing, making the current expenditure on brute force methods seem inefficient in hindsight.