2026-05-05

SubQ

Original source

Reading Log X Post Historical Backfill LLM Research

Imported from historical reading log.

  • Main post successfully extracted via api.fxtwitter.com fallback.
  • Post by Alexander Whedon introducing SubQ as a sparse-attention LLM architecture claim: fully sub-quadratic sparse attention, 12M token context window, 52x faster than FlashAttention at 1M tokens, under 5% of Opus cost, and nearly 1,000x less compute by focusing only on relationships that matter.
  • Core framing: standard transformer attention computes many unnecessary token relationships; sparse attention focuses only on the small fraction that matters.
  • No tweet.article block present in the API response for this post, so plain tweet text was used.
  • Could not reliably access replies without hitting X login/interstitial walls; need another mirror, API route, or screenshots if reply-level analysis matters.