Wired: New Trick Extracts AI Inner Thoughts, Suggesting Chinese Models Distill US Rivals
Wired AI · rss · 2026-08-11
Researchers have developed a novel technique to extract hidden "reasoning traces" from leading closed-source models like Claude, GPT, and Gemini.
By analyzing these internal thought processes, the researchers found patterns in some Chinese AI models that suggest they may have been trained directly on the outputs of their leading US counterparts, a practice known as model distillation.
More from Models
- Meta Launches 30B Open Agentic Model, Zuckerberg Slams Closed AI in Manifesto — 新智元 · 2026-08-11
- OpenAI Launches GPT-5.6-Cyber for Security, Uncovering 400+ Kernel Vulnerabilities — 新智元 · 2026-08-11
- Anthropic to Add Invisible Watermarks to Claude-Generated Text and Images — The Verge AI · 2026-08-11
- MMOOC Benchmark: Multimodal LLMs Struggle to Balance In-Context and Out-of-Context Questions — VCLab-HKPU · 2026-08-11
- Muse Glimmer 30B Ships with DFlash, Boosting Inference Speed 2-4x — mervenoyann · 2026-08-11
- Grok 4.6 Drops Today: A Comparison of Supported Coding Tools — PawelHuryn · 2026-08-11