# @Krubbled — 2026-04-24

♥0 ↻0 · https://x.com/Krubbled/status/2047737473271488892

Deepseek v4 makes you really appreciate how difficult the 1M+ token context engineering problem is. They had to reinvent attention like 4 different times (MLA -> DSA -> CSA + HCA) just to make inference affordable.

tags: author:krubbled, kind:tweet, on:deepseek-v4-pro, year:2026
cited on: deepseek-v4-pro