stanslab_dev

@stanslab_dev

Senior Dev | Indie Maker 🛠️ #BuildInPublic | AI • React Native 🎮 Free Games: game.stanslab.dev 👇 Links: linktr.ee/stanslab_dev

Wrocław, Poland

Joined November 2025

stanslab_dev
@stanslab_dev
14h
Languages have addresses inside Gemma 4. 🗺️ We probed all 128 experts with 16 languages (100 sentences each) and watched who lights up. Polish & English = a broad, steady backbone. Most other languages flow through narrow niches.
stanslab_dev
@stanslab_dev
Aug 26
Waiting <3 I love this model
Mia
@MiaAI_lab
Aug 26
Time to reveal! Ox Alpha is Zai's GLM Flash 🔥 And it's coming out... tonight!
stanslab_dev
@stanslab_dev
Aug 24
Gemma 4: ablation drift, not routing %. Each value = 1−Jaccard vs baseline after zeroing one expert; higher = stronger effect. L7/E94: Polish 0.473 vs foreign 0.312 vs random 0.226. L12/E82: Polish 0.408, Math 0.033 — “math expert” wrong. 😅 What next? :D
stanslab_dev
@stanslab_dev
Aug 21
We first thought Gemma 4 had 87% “dead” experts. Wrong experiment. In a small pilot, masking the prompt-suite union moved outputs more than its complement: 12/12 Polish cells, 11/12 Math. Not a general claim yet. Next: dozens of held-out prompts.
stanslab_dev
@stanslab_dev
Aug 21
Thats why I analyze those small models :D (and macbook pro m1 pro 32Gb is enough for this work)
Google Gemma
@googlegemma
Aug 21
You don't always need a frontier model. A recent benchmark found that Gemma 4 31B matches Sonnet 5 on answer quality at ~40x lower cost. With high cost-efficiency and low latency, Gemma unlocks high-volume use cases that are uneconomical with larger models.