Subliminal learning risks: why adversarial assumptions bar synthetic data from frontier models
xeophon · x · 2026-11-22
In an X exchange on synthetic data safety, @anpaure argues that under an adversarial assumption you shouldn't generate synthetic data from frontier models at all, since phenomena like subliminal learning can transmit hidden traits; filtering alone isn't enough. @xeophon adds that 'don't train on XY outputs' policies are practically unenforceable as the data spreads across the open web.
More from Models
- 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
- 'Don't train on model outputs' is unenforceable when the data is all over the open web — anpaure · 2026-11-22
- Gemini solves unaired game show puzzle with no online answers; Astra falls short — deedydas · 2026-11-22