A “diffusion” language model. Claude, GPT and almost every other model write one word at a time, left to right. A diffusion model drafts the whole answer roughly and then sharpens every part of it at once over a few passes. Image generators already work this way.

Show Post

That’s where the speed comes from: about 1,070 words-pieces per second, against roughly 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna. Why it matters. Earlier diffusion models were fast but clearly less capable. This is the first one that matches or beats the small, fast models from the big labs.

Show Post

I regularly output things that plausibly look like a sentence.

Show Post

I was originally going to do my PhD remotely but ended up moving here to Germany because of an unscrupulous employer forcing my departure from the US. I'm so glad I'm here, though, because I have an awesome group of fellow doctoral candidates in the lab that I'm excited to be on the journey with.

Show Post

maybe France and Spain can stop giving Russians tourist visas before we get to the shooting part

Show Post