This is the long version of what I previously post. Deepseek tried out some tricks during llm training. And the reason they can do it is also because previous open source llm model and papers. OpenAI found those tricks too but they didn’t open source it. nostr:note19caz355x4kyjvy5hn96n76rfp84lecsvrgn69zuuvcrgl7jjy93q3eq0ry