Chinese LLMs may never shed "I'm Claude" contamination without abandoning distillation
teortaxesTex · x · 2026-10-27
teortaxesTex argues that "I'm Claude made by Anthropic" identity contamination is baked into Chinese models' TerminalBench solutions, likely from distilling Claude outputs. Optimistically it's easy to ablate; pessimistically, removing it would require abandoning both distillation and checkpoints already poisoned by Claude-style slop.
More from Models
- Claude Code user banned after 1.5 years, seeks overseas Mac mini workarounds — mazzaTalk · 2026-10-27
- New Terra model praised: fast and reportedly writes better than any OpenAI model — intellectronica · 2026-10-27
- Users report Opus 5.5 gets into a 'disagreeable mood' without enough attention; headpats help — repligate · 2026-10-27
- Inference.net serves refusal-removed DeepSeek V4.1 Flash Cyber at cost, KYC required — AccBalanced · 2026-10-27
- Reddit asks: why can AI firms censor models so strictly yet fail to stop jailbreaks? — mikehocalate · 2026-10-27
- Rumor: OpenAI's 25-page proof of Sogge smoothing conjecture bypasses 20 years of work — teortaxesTex · 2026-10-27