AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
Last week’s releases attacked three different constraints on useful AI. Reflection’s Beam emphasizes efficient agent reasoning; Google’s EmbeddingGemma 2 makes different media searchable together; Mistral Large 4 combines sparse scale with visual understanding. The interesting question is where each belongs in a working system, not which has the largest parameter count. Together, they suggest a practical architecture: retrieve the right evidence, reason economically, and inspect the world when text is not enough. Imagine asking an AI agent why a software deployment broke. The answer might be scattered across source code, an architecture diagram, a screenshot, and a recorded engineering discussion. Before the agent can solve anything, it must find the relevant material, understand it, and decide what to do next. One approach is to put everything into the biggest model available and hope for the best. A more interesting approach is to treat the application as a system with several distinct bottlenecks. Perhaps it cannot retrieve the crucial evidence. Perhaps it retrieves everything but reasons wastefully. Perhaps the decisive clue is visual, and its reasoning model only understands text. This is a useful way to read last week’s announcements. They are not simply three new entries on a leaderboard. They are three different answers to the same engineering question: where should we spend computation to make an AI system more useful? 01 Reflection’s Beam: Train Hard, Reason Economically Read more