Research-focused AI signal from labs, technical talks, and scientific breakthroughs. / 追踪研究机构、技术演讲与重要科研进展。
Signal brief / 信号简报
Drone-Bench shows frontier models nearing autonomous drone control without yet reaching dependable end-to-end performance. The strongest model cleared four of five subtasks, but reconstruction errors cascaded into localization and navigation; average reliability still trails one-off best results by roughly six months.
Read signal / 阅读信号Interpretability is shifting from the ambition to fully explain a model toward practical auditing, monitoring, and debugging. Visible reasoning traces remain useful evidence, not proof: stronger systems may omit or shape what they reveal.
Read signal / 阅读信号“World model” now labels systems that predict very different parts of reality. The useful distinction is functional: state, observation, action, and transition each support different capabilities, so visual plausibility alone does not establish physical understanding.
Read signal / 阅读信号Evidence stream / 证据流