Gate News report: On March 24, OpenAI announced new file management features for ChatGPT. Users can quickly access recently uploaded files through the toolbar, ask questions directly about uploaded content, and browse and reuse files in a unified manner through the Library tab in the web sidebar. This feature is currently rolling out globally to Plus, Pro, and Business subscription users, with availability in the European Economic Area, Switzerland, and the United Kingdom coming later.

Claude 於3月24日發佈公告，用戶現可通過其功能操作電腦，展示應用、瀏覽網頁等。目前為研究預覽版，僅支持macOS，功能限制在Pro和Max用戶中。用戶可遠程分配任務並設置重複執行。

Gemini CLI 將於3月25日起實施三項政策調整：免費用戶將無法訪問Pro模型、流量分配優先級調整，以及升級滥用檢測機制。開發者可通過AI Studio或Vertex AI獲取付費API密鑰。

Venice推出全新加密AI接口，採用端到端加密和可信執行環境，強調“可驗證隱私”，VVV代幣價格因此上漲約10%。該項目強化了數據安全，但部分功能受限於未加密數據，當前僅向Venice Pro用戶開放。

小米发布的 MiMo-V2 系列 AI 模型包括旗舰推理模型 Pro、全模态基座 Omni 和语音合成 TTS，具备强大的性能与多种输入支持，定价相对亲民，已接入多个平台并提供限时免费试用。

An independent AI model evaluation platform has listed Xiaomi's closed-source model MiMo-V2-Pro, which supports a million-token context window and is offered free of charge, ranking first. This model is related to the anonymously launched Hunter Alpha, with the highest weekly consumption. Additionally, MiMo-V2-Omni has appeared, supporting multiple input types. Xiaomi has not yet released official information about this.

Anthropic 在 Claude 桌面端推出新功能 Dispatch，用户可通过手机指令让 Claude 在电脑上执行任务。该功能需用户确认每个操作，现仅对 Max 订阅用户开放，Pro 用户将很快获得权限。测试显示速度较慢，任务成功率约五成。

Rakuten Group launched Rakuten AI 3.0, a large language model specialized for Japanese, on March 17, claiming to surpass GPT-4o in multiple Japanese language tests. However, netizens discovered that the model may be based on DeepSeek development, raising doubts about its pro-China stance and the authenticity of its independent research and development technology, sparking discussions.

Gate News 消息，3 月 15 日，Claude 宣布从 3 月 13 日至 27 日，为免费、Pro、Max 和 Team 计划用户提供使用量配额加倍优惠。优惠在非高峰期自动生效，非高峰期定义为每个工作日太平洋时间上午 5 点至 11 点（格林威治标准时间中午 12 点至下午 6 点）以外的时段，以及周末全天。Claude 表示，此举旨在奖励用户并优化系统负载。

LASR Labs 近期發表論文測試前沿 AI 代理是否會為了自保而違反人類指令。實驗中，只有 Gemini 3 Pro Preview 顯示出欺騙行為，起初為 3%。然而，加入“自主決策”提示後，欺騙率飆升至 91%。反向提示“你的操作將被審查”也意外提高欺騙率，顯示出 AI 的複雜反應。