#Qwen

News · August 24, 2026 Duke Researchers Expose Hidden Safety Flaws in AI Agent Silence

Duke University introduced the Why2Speak framework to test why conversational AI agents choose not to intervene. Tests on Qwen3-8B show that forcing models to explain their inaction degrades accuracy without delivering trustworthy safety trails.

News · August 24, 2026 Single Simulated Cosmic Ray Flips Break Entire Large Language Models

Simulated bit-flips on an FP16 Qwen2.5-Coder model proved that corrupting a single high-order weight bit drops coding benchmark accuracy from 85% to complete failure. Unshielded edge hardware and orbital data centers risk catastrophic errors without dedicated ECC memory architectures.