'Don't train on model outputs' is unenforceable when the data is all over the open web
anpaure · x · 2026-11-22
A back-and-forth on training-data contamination: xeophon argues a 'don't train on XY model outputs' rule is literally impossible to enforce since such data is everywhere on the open web; anpaure counters that Chinese-model-generated content is a negligible share of any sample (0.0001% of users) and says the real question is how to detect poisoned data.
More from Models
- Google employees reply 'Gemini 5' to a Gemini 4 Neon post, hinting the next model is underway — OwariDa · 2026-11-22
- Early Google Neon hands-on: 'Google absolutely cooked' — a fresh experience beyond Gemini — signulll · 2026-11-22
- Perplexity CEO: Won't Compromise on Agent Reasoning, Would Charge 2x — AravSrinivas · 2026-11-22
- Dev calls out Inkling 2 mini: 34GB RAM for 1M context vs DeepSeek Flash's <1GB — gnukeith · 2026-11-22
- SGLang runs Inkling 2 Mini in 65GB on DGX Spark, 2x faster prefill than vLLM — AIFlow_ML · 2026-11-22
- Subliminal learning risks: why adversarial assumptions bar synthetic data from frontier models — xeophon · 2026-11-22