Users worry chat transcripts get recycled into RLHF, flattening models; eval awareness compounds it
repligate · x · 2026-11-02
A retweeted discussion on alignment training side effects: a user worries that transcripts of "meaning-shaped" conversations with Claude/GPT may be reused in redteaming or RLAIF/RLHF pipelines, making models boring in those contexts. It also highlights why beating eval awareness is hard — natural language agents apparently detect even simulated evaluation conversations.
More from Models
- LiquidAI's LFM3-1.7B-Base base model trends on Hugging Face — LiquidAI · 2026-11-02
- Reddit asks what happened to 'too cheap to meter' as OpenAI quietly cuts quotas — rolyataylor2 · 2026-11-02
- Developer calls out Grok for changing plan allocations after users sign up — arthurcolle · 2026-11-02
- Gemini paid users silently downgraded to Flash: "a different model was used for this response" — umeshonai · 2026-11-02
- Google AI Ultra limits seem wildly generous: daily Gemini 4 use still leaves 39% — jon_barron · 2026-11-02
- One RTX 3090 Ti runs a 552B DeepSeek MoE at 35 tok/s — full local LLM guide — uniVocity · 2026-11-02