"Desired oververbosity 5": Claude spots OpenAI distillation fingerprints in DeepSeek reasoning traces
zeeshanp_ · x · 2026-10-29
The author notes Claude has an uncanny ability to spot odd behaviors in the reasoning traces of Chinese open models that point to distillation from closed models.
- DeepSeek's reasoning often contains text about the model itself rather than the problem.
- The strangest marker: the line "Desired oververbosity 5" appears in a significant share of traces — "desired oververbosity" is a setting in OpenAI's hidden system prompts controlling answer length on a 1–10 scale, which DeepSeek appears to have memorized during training.
- DeepSeek also famously calls itself ChatGPT, but the interesting part is that Claude can catch far more nuanced distillation indicators too.
More from Models
- GPT-6.1 Sol capacity limits stall overnight agent tasks after 45 minutes — CtrlAltDwayne · 2026-10-29
- Rumor: Anthropic reportedly has internal model beating Mythos 5.1 with better alignment — VraserX · 2026-10-29
- DIAL Benchmark Ranks Speech-to-Speech Models With 30k+ Human Ratings Across 9k Live Calls — sarahcat21 · 2026-10-29
- Benchmaxxing debate: is training on test despicable or the vertical AI superpower? — HanchungLee · 2026-10-29
- OpenRouter served 5.2T tokens in a day—back-of-envelope math questions where all the GPUs go — teortaxesTex · 2026-10-29
- Musk amplifies dev's list of AI 'wtf' moments, hailing Grok Bot as genuinely incredible — elonmusk · 2026-10-29