The author did a great job with this video. I converted it into a Chinese version and revised it based on the thesis, and shared it with everyone.
The focus of this video is how DeepSeek V4.1 Flash reduces the KV cache required to save historical information. It mainly does four things: Reduce duplicate storage: 40 layers share four global caches by group, three for the encoder and one for the decoder. Compressed history entries: the encoder combines two tokens…
Installed Rive CLI, "cute little duck" designed using deepseek-v4-flash.
Jack Ma has singing talent, but is illiterate about AI. From his conversation with Musk
Jack Ma has singing talent, but is illiterate about AI. From his conversation with Musk, it can be seen that Alibaba does not have a technical gene, but it has a trading gene. As soon as Alibaba's stock price rebounded a little, it was slapped by Anthropic: Claude accused Alibaba, Dark Side of the Moon, Xiaomi, Zhipu, DeepSeek, SenseTime, and MiniMax of illegally distilling their AI models.
20260911 These 100 lessons are not stored for you to copy your homework.
Only by understanding the underlying logic can we promote the healthy development of the storage industry. I don’t comment on stocks or recommend stocks, I only share the logic of the storage industry; after mastering the logic of the storage industry, you can analyze and make decisions yourself. All contents of this account do not constitute any investment advice.
You have reached the end of the archive
All of deepseek