Two Minute Papers · 2026-08-07 · notable
Two Minute Papers — DeepMind's Gemma 4 training trick 'everyone should copy'
Károly Zsolnai-Fehér walks through a training technique from DeepMind's Gemma 4 technical report that he argues every lab should adopt, using the 2B–31B multimodal model family as the case study.
Two Minute Papers picks apart a DeepMind training trick from the Gemma 4 report and argues everyone should copy it.
What is it?
The Two Minute Papers episode 'DeepMind's AI Trick Everyone Should Copy' is a fresh video from Károly Zsolnai-Fehér that unpacks a training technique highlighted in Google DeepMind's Gemma 4 Technical Report (arXiv 2607.02770).
How does it work?
Two Minute Papers walks through the paper's key ablations on the 2.3B–31B Gemma 4 family — a mix of dense and Mixture-of-Experts models with a unified encoder-free 12B variant — and isolates the training step Zsolnai-Fehér flags as the 'trick everyone should copy'.
Why does it matter?
Because Gemma 4 is open-weight and its recipe is now public, any lab or open-source group training a new model can lift the technique directly — the video is a fast way for practitioners to catch the specific move before deciding whether it's worth reading the 200-plus-author report end-to-end.
Who is it for?
ML engineers, open-model trainers, and researchers tracking Gemma 4's methodology