Different RL judges train different design taste into models, taste hard to instill
round · x · 2026-11-03
Jasper Lu's experiment trained Qwen3.8-27B with Sol vs. Sonnet as RL judges on the same webdev environments, producing models with distinct design behaviors — though both judges preferred Sonnet's lineage. Commentators argue taste is too contextual to train into model instincts, raising questions about blending models to evoke tasteful output.
More from Models
- Mistral Large 4 tech report: 52B-active/1T MoE, fits on four GB300 GPUs at 606GB — aronchick · 2026-11-03
- Leaked Mistral Large 4 abstract: 1T-param MoE with 52B active, fits on 4 GB300 GPUs — qtnx_ · 2026-11-03
- Tester: LongCat's website model is 'dumb as rocks' in Russian, far below parity claims — teortaxesTex · 2026-11-03
- OpenAI's new LifeVal benchmark reportedly tests everyday tasks; GPT-6 Astra hits 100% on several — mark_k · 2026-11-03
- 36 runs, 12 open models, one medical exam question — and all disagree with the answer key — MaziyarPanahi · 2026-11-03
- User: Claude has lost its holistic coding ability, forcing a switch to other models — BrainSurfing · 2026-11-03