Gemini 3.1

1. Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong (github.com via hn)

   +21 8w mmlugemmagemini

   Hey HN, Henry & Roman here from Cactus. A small, on-device model is fast and private, but sometimes wrong, but frontier models are getting expensive pretty fast.

2. A skill that saves Claude usage for thinking (judge) and hands the grunt-work coding to cheaper/free LLM models (executor) (www.reddit.com via reddit)

   8w glmdeepseekgemini+2

   Sharing a skill built with Claude Code that I've been relying on for my personal projects — hoping someone else might find it useful too. The problem: My bigger personal projects were draining my Claude limits fast — and most of that usage…

3. Animated SVG comparisons between several models (www.reddit.com via reddit)

   8w haikugpt-5sonnet+4

   I have seen some people testing models by telling them to generate images of difficult, unusual SVGs, and I thought: what if I elevate difficulty a bit and specify that it also has to be animated, and perfectly looped? I have tested Haiku…

← all threads