DeepSeek is training a 2T-parameter model and plans to build an 8T-parameter one | Hacker News Reader