2026年3月19日deepseek.com
DeepSeek R2 技术报告泄露，引发外界对训练路线与成本结构猜测
Reports circulated on March 19 that a technical document tied to DeepSeek R2 had leaked, prompting renewed discussion about architecture choices, training efficiency, and deployment strategy. Even without formal confirmation, the episode drew attention to how closely the market is watching cost-performance tradeoffs outside the major U.S. labs.
Will 点评
我对泄露新闻一向保持克制，因为二手材料最容易被拿来神化或抹黑。但 DeepSeek 每次引发讨论，都说明市场在重新评估中国团队的工程效率和成本结构，这个变量已经没人能忽视。
#DeepSeek#Leak#Research
2026年3月18日cas.go.jp
日本政府发布 AI 战略白皮书，重点转向产业落地与治理并行
Japan's government published a new AI strategy white paper on March 18, outlining updated priorities around industrial adoption, talent development, national competitiveness, and governance. The document is notable because it frames AI as a policy execution issue tied to productivity and public trust, not only a research agenda.
Will 点评
我很在意日本这类白皮书，因为它往往决定企业后续预算、补贴和合规节奏。日本市场一旦开始把 AI 从概念讨论切到制度化推进，真正受益的会是那些能把产品做得稳、做得可解释的团队。
#Japan#Policy#Governance
2026年3月17日fda.gov
AI 医疗诊断产品通过 FDA 510(k)，监管落地进入新阶段
An AI-assisted medical diagnostic system reportedly received FDA 510(k) clearance on March 17, signaling that regulated clinical AI is moving further into real deployment pathways. The event matters because approval milestones reshape the conversation from laboratory promise toward reimbursement, workflow integration, and legal accountability.
Will 点评
我一直觉得医疗 AI 的真正门槛不是 demo 准不准，而是能不能穿过监管、责任和临床流程三道门。510(k) 这种节点一旦出现，行业讨论就会从想象力回到采购、集成和问责。
#Healthcare#FDA#Diagnostics
2026年3月15日ai.meta.com
Meta 发布 Llama 4，继续用开源生态争夺模型分发权
Meta unveiled Llama 4 on March 15 as the next major generation of its open model family, emphasizing scale, multimodality, and broader deployment flexibility. The release reinforced Meta's strategy of using openness and ecosystem reach to shape how AI models are adopted downstream.
Will 点评
我一直觉得 Meta 的核心竞争点不是单次发布有多炸，而是它不断把分发权往开源生态里拉。对开发者来说，可改、可托管、可嵌入自家流程，很多时候比榜单第一更有现实吸引力。
#Meta#Llama#Open Source
2026年3月14日blog.google
Google 更新 Gemini 2.0 Pro，多模态理解继续向生产级靠拢
Google announced a March 14 update to Gemini 2.0 Pro focused on broader multimodal input handling and improved reasoning across mixed media tasks. The change matters because it points to a product direction where multimodal capability is expected to operate reliably inside everyday workflows, not only as a showcase feature.
Will 点评
我对多模态一直比较克制，因为很多演示都强在展示，弱在连续工作。但如果 Gemini 2.0 Pro 真把图文音视频放进一个稳定推理框架里，它在搜索、办公和代理链路上的联动空间会非常大。
#Google#Gemini#Multimodal
2026年3月12日cursor.com
Cursor 1.0 正式版发布，AI 编程工具进入产品化新阶段
Cursor announced version 1.0 on March 12, positioning the release as a stable foundation for AI-assisted software development rather than an experimental editor layer. The milestone suggests the category is shifting from feature races toward reliability, team governance, and repeatable engineering workflows.
Will 点评
我一直把 Cursor 当成 AI 编程产品化能力的观察窗口。做到 1.0 不只是功能变多，而是说明它开始对稳定性、团队协作和默认工作流负责，这才是真正能吃到企业预算的节点。
#Cursor#Coding#Developer Tools
2026年3月11日modelcontextprotocol.io
MCP 被 15 家主流工具采用，模型上下文接口开始走向标准化
On March 11, the Model Context Protocol gained another wave of support as roughly 15 mainstream tools signaled adoption or compatibility. The shift is important because standards for context exchange can reduce the friction of connecting models with editors, data systems, and agent runtimes.
Will 点评
我一直认为 2026 年真正重要的不只是更强模型，而是上下游终于开始讲同一种接口语言。OpenClaw 这种多工具编排场景最怕各家各写一套，MCP 一旦成形，集成成本会立刻下降一个量级。
#MCP#Standards#Ecosystem
2026年3月10日anthropic.com
Claude Sonnet 4 发布，主打更稳的长上下文与企业代理能力
Anthropic introduced Claude Sonnet 4 on March 10 as the next general-purpose model in its Claude lineup, highlighting more reliable long-context reasoning and better enterprise workflow execution. The release was framed less as a flashy leap and more as a refinement for teams that depend on consistent performance inside agentic and document-heavy tasks.
Will 点评
我对 Sonnet 线一直很关注，因为它最接近真正可落地的工作模型。OpenClaw 这类多步骤协作场景里，稳定性、工具调用和长上下文一致性，比跑分上的戏剧性提升更值钱。
#Anthropic#Claude#Agents
2026年3月10日blog.google
Google 扩展 Gemini 在 Docs、Sheets、Slides 与 Drive 中的办公能力
Google announced on March 10 that Gemini in Workspace would gain broader drafting, spreadsheet-building, slide-generation, and Drive answer features for Google AI Ultra and Pro subscribers. The practical significance is that Gemini is being embedded deeper into default work surfaces, making AI assistance feel less like an add-on and more like built-in office infrastructure.
Will 点评
我一直觉得 AI 办公的胜负手不是谁更会写一段漂亮文案，而是谁能把上下文、权限和协作流接得最顺。入口还在邮件、文档和云盘里，Google 的主场优势就很明显。
#Google#Gemini#Productivity
2026年3月9日openai.com
OpenAI 宣布收购 Promptfoo，强化 LLM 评测与红队能力
OpenAI said on March 9 that it plans to acquire Promptfoo and integrate its security testing and evaluation technology into OpenAI Frontier. The announcement matters because it treats evaluation, red teaming, and traceability as part of the product surface for enterprise agents rather than optional tooling for advanced teams.
Will 点评
我很认同把评测能力前置。模型越来越像基础设施之后，真正拉开差距的是你能不能持续发现问题、量化风险、快速回归，而不是只会发版本海报。
#OpenAI#Evaluation#Safety
2026年3月8日openai.com
OpenAI Operator 全量开放，网页代理正式进入大众使用阶段
OpenAI expanded Operator to general availability on March 8, making its browser-based task agent accessible beyond earlier limited cohorts. The move signaled confidence that web automation assistants are becoming a mainstream product surface rather than a research preview for power users.
Will 点评
我对 Operator 的判断一直很明确：它的意义不在演示能点几次网页，而在于是否能把不稳定的人类浏览器操作收敛成可复用流程。OpenClaw 如果要继续放大自动化价值，这条路迟早都得打通。
#OpenAI#Operator#Agents
2026年3月6日github.blog
GitHub Copilot 新增 Agent 模式，开始争夺完整开发流程入口
GitHub introduced a new Agent mode for Copilot on March 6, extending the product from inline completion toward more autonomous coding assistance across broader tasks. The update suggests that developer AI tools are converging on a model where planning, editing, and validation live inside the same interface.
Will 点评
我不把 Agent 模式看成一个小功能，而是平台控制权的延伸。谁能从补全走到任务拆解、改文件、跑验证，谁就不再只是助手，而是在重写开发者默认工作台的边界。
#GitHub#Copilot#Agents
2026年3月5日anthropic.com
Anthropic 回应美国国防部相关限制措施
Anthropic published a March 5 statement after receiving a March 4 letter from the U.S. Department of Defense, saying the action had a narrow scope tied to direct DoD contract use and that the company would challenge it in court. The update is notable because it shows how quickly frontier AI policy is shifting from abstract safety language into concrete procurement and national-security constraints.
Will 点评
我更关注的不是标题里的冲突感，而是前沿模型公司开始被迫把边界写进现实世界的治理框架。到了这个阶段，能力再强，没有制度化约束也走不远。
#Anthropic#Policy#Defense