Post

Moonshot AI claims their new Kimi Linear LLM, a hybrid with O(n) linear attention, improves quality over full attention while slashing O(n²) bottlenecks. Huge if true. Chinese researchers are cooking 🔥🔥🔥. Have you sold your nvidia stocks yet? 1/2 arxiv.org/pdf/2510.26692

arxiv.orgarxiv.org

Marius Gerdes

@mglambda.bsky.social

Now they still have quadratic layers so it's still O(n²) if we're being pedantic. Thing is I'll take a small constant scaling factor in quadratic complexity over an academically pure but practically useless O(n) any day. And if you don't believe me I have an L1 cache to sell you. 2/2

3/11/25, 9:01 am0 reposts0 likes

No replies yet.