Krea 2 Technical Report
Data curation, architecture, training, and prompt expansion behind Krea 2.
I worked on federated learning at Google and AutoML at Meta, then co-founded Comfy as CTO. Now I work on image models at Krea.
Writing
ML Work When You're Not at Google
A pragmatic workflow for ML teams without infinite compute, platform teams, or six-figure experiment tracking systems.
Hacking Diffusion into Qwen3 for the ARC Challenge
What happens when Qwen3 solves ARC puzzles by filling in the easy parts first? Better pixel accuracy, fewer decoding steps—and a stubborn gap in solved puzzles.
You Might Not Need RAG
We built a sophisticated RAG pipeline. Then our simpler document-reading agent outperformed it. What we learned about retrieval, ranking, and doing the dumb thing first.