AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。
AI progress is usually drawn as one upward-sloping line: more parameters, more compute, higher benchmark scores. Last week looked more like a three-dimensional coordinate system. DeepSeek added vision to its fast V4 model, giving agents a compact way to turn screenshots, charts, and documents into actions. A Google Cloud AI Research team introduced EnvHarness, a framework that makes training environments adapt to the weaknesses of the agent inside them. Etched shipped its first inference rack to Jane Street, moving its specialized hardware thesis from silicon demos into a customer data center. These developments sit at three layers - model, environment, and infrastructure - but point in the same direction. The next phase of AI will come from tightening the loop around the model: what it can perceive, what it learns from, and how cheaply its intelligence can be served. 1. DeepSeek-V4-Flash-Vision-Exp: The Agent Gets Eyes Read more