Tweet by darraghcurran

April 9, 2026

Fair play if you can follow all the technical and mathematical detail here and in the blog post and paper. The plain-English version: this is a big deal. It is a meaningful architectural advance against a major LLM bottleneck. The clearest win is cheaper, more scalable inference, often faster inference too, with a secondary training benefit: reaching the same model quality with less total compute. Excited to see how others apply this in production and build on it. Kudos to @JamesONeil21 for his phenomenal work on this.

Author
darraghcurran
Date
April 9, 2026