Quick test: Gemini 3.1 Pro knows the answer in its weights, Flash doesn't
giffmana · x · 2026-11-06
In a discussion over citations in a scaling law paper, giffmana ran a quick experiment: in AI Studio with web search and conversation history off, Gemini Flash doesn't know a literature detail while Gemini 3.1 Pro answers from its weights. He also argues the paper's remedy is wrong — its extrapolation "solution" basically amounts to "just use a heldout val set".
More from Models
- Meta's frontier model new 'walrus' checkpoint rumored to match crouton-alpha quality — ChrisGPT · 2026-11-06
- Ripple-65B flops in single-GPU coding test against Qwen 27B despite SOTA claims — DZhang50 · 2026-11-06
- Ripple-65B open-weight model runs at 7 tok/s locally on a 64GB Mac mini — DZhang50 · 2026-11-06
- OpenAI's next Astra, Sol and cyber-focused models could land as soon as Friday — kimmonismus · 2026-11-06
- DeepFinance benchmark: 21 real PE/IB tasks, Claude Opus 5.5 tops at 54.8% — tszzl · 2026-11-06
- ValsAI claims critical safety flaw in frontier lab models enabling high-stakes misuse — rickasaurus · 2026-11-06