Gemini 4 Carbon keeps saying 'we' instead of 'I' on PostTrainBench
maksym_andr · x · 2026-11-13
maksymandr notices Gemini 4 Carbon (and Argon) consistently refer to themselves as 'we' rather than 'I' on PostTrainBench. The traces show models reporting elaborate post-training runs (SFT/DPO on Qwen3-4B, decontamination checks, sampling-config alignment), so the pronoun shift may hint at post-training artifacts. xeophon adds: like biology, if you don't want models to gain a capability, don't train on and for it.
More from Models
- InferenceBench: agents beat vLLM defaults but lose to simple hyperparameter search — maksym_andr · 2026-11-13
- Inkling joins Eval³ leaderboard that scores models on auditing evals for reward hacks — simonguozirui · 2026-11-13
- MiniMax H3 outputs silently changed: same workflow now yields flickering artifacts — Cequejedisestvrai · 2026-11-13
- Claude Starts Refusing to Help with Pokémon ROM Hacks and Fan-Game Development — Lucas_Zxc2833 · 2026-11-13
- Orchestrator rolls out to Pro subs, pairing Hydrogen and Argon to save quota — clmt · 2026-11-13
- Liquid AI's LFM hits 'college graduate' intelligence at 120 tps on a MacBook Pro — JosephJacks_ · 2026-11-13