Commentary on X
@_xjdr i buy it. >1 distillation white paper this the big one, and more a paradigm shift/new understanding of how/why it works. 30b dense that is 5.5/4.8 ish. that + i buy my fancy tokens from noumena and it's a done deal. my bet is 6-12 months out for that whole picture.
4:23 PM · Jul 14, 2026@_xjdr I’ve been pretty blown away by how useful it is. I think prior to this, the gpt oss models were their best small models. I use Luna all the time now for munging data!
6:56 PM · Jul 14, 2026@_xjdr glm 5.2 to me is very gpt 5.5 shaped, this and deepseek pro v4 are so good
4:08 PM · Jul 14, 2026View all
i think we are a 3T param base, 1 more iteration / improvement on open weights model SOTA (k3 could hit both) and 1 distillation white paper away from things looking very very different . with proper UX, current open weights models are good enough for 90% of queries from 90% of people (probably more) also most people reading this are outside of that 90% (while some of you are very specifically included )
4:06 PM · Jul 14, 2026I've become skeptical about any specific phase transition. Frontier labs continue to propagate the frontier and cover these last miles in a way that makes their pricing defensible, and stronger models + more compute + more money&staff still compound. This could go on for a while
i think we are a 3T param base, 1 more iteration / improvement on open weights model SOTA (k3 could hit both) and 1 distillation white paper away from things looking very very different . with proper UX, current open weights models are good enough for 90% of queries from 90% of people (probably more) also most people reading this are outside of that 90% (while some of you are very specifically included )