bin sun⚡️

AI 新闻速览 · 工具实操吐槽
Google / OpenAI / Claude / Grok 边用边说

[bin sun⚡️ @binbinyiyilili]

Meta 刚发布个人 Agent Muse：不是又一个聊天框，而是跑在独立 Secure VM 上的私人助手——自带浏览器，能代办邮件、订票、购物，另有 Sentinel 拦外发。美国 iOS/Android/muse.ai 先上，基础免费。和自建 agent 不同，Meta 押的是「开箱即用 + 云端隔离」——下一关是你敢不敢把邮箱和支付交给它。

[AI at Meta @AIatMeta]

Introducing Muse, a personal agent that gets things done for you, powered by Muse Spark 1.3.

Get an inside look at how we built Muse and what it can do for you: https://t.co/qzFgoNAUSV

[bin sun⚡️ @binbinyiyilili]

@thsottiaux 谢谢tibo

[bin sun⚡️ @binbinyiyilili]

@kimmonismus @nvidia 最想听的是 GPU 能保持盈利多久那段：推理负载从大模型训练转向长时程 agent 之后，一代卡的折旧周期直接决定 AI 基建这笔账算不算得过来。

[bin sun⚡️ @binbinyiyilili]

@gigazine Apple Intelligence 的本地模型要占好几 GB，RemoveMacAI 对 256GB 这类小容量 Mac 挺实用；不过系统大版本更新后模型可能会被重新下载，更新完记得再查一次。

[bin sun⚡️ @binbinyiyilili]

@OpenAIDevs @Microsoft Windows 用户终于有原生级沙盒了：基于微软 MXC，设置更快、网络限制更严、文件权限更细。注意门槛是兼容的 Windows 11 设备，老机器先确认一下。

[bin sun⚡️ @binbinyiyilili]

@arena 从真实长时程 agent 会话里做因果追踪、测五个信号，比两两偏好投票更接近“能不能把活干完”。好奇长任务里的单步失误怎么归因到具体模型，而不是工具或环境。

[bin sun⚡️ @binbinyiyilili]

@GoogleDeepMind 7.4 亿参数、Apache 2.0、还能配 Gemma 4 做端侧私有 RAG。“用语音备忘录定位视频片段”这种跨模态检索，以前要几套模型拼起来，现在一个共享向量空间就能搞定。

[bin sun⚡️ @binbinyiyilili]

@OpenAIDevs 根据对话内容和交流习惯预测下一条消息，对“继续 / 跑测试 / 修这个报错”这类高频短指令最省事。Pro 先测，期待之后放给更多用户。

[bin sun⚡️ @binbinyiyilili]

@kimmonismus “假设模型已被攻破”这个前提最关键：用另一个 AI 盯 agent 只是黑盒查黑盒，真正可落地的是权限最小化、完整日志和可观测性。If you can't observe it, you can't trust it.

[bin sun⚡️ @binbinyiyilili]

@arena @AnthropicAI 1749 vs 1740，只差 9 分，而且两个都在帕累托前沿。对大多数截图转网页任务来说，Sonnet 5.5 的性价比可能更值得先试，Opus 留给复杂交互页面。

[bin sun⚡️ @binbinyiyilili]

@testingcatalog 两条线索放一起看：隐藏的“Preview”语音标签 + 新增的语音数据训练共享选项，Anthropic 自研语音模型的概率在变大。Claude 目前语音体验落后于对手，这块补上会很关键。

[bin sun⚡️ @binbinyiyilili]

@testingcatalog 关键限制在这：只有你明确要求、且额度剩余≤10%时才会动用重置，绝不自动触发。等于把“续命”按钮交还给用户，长任务跑到一半不会被悄悄消耗掉。

[bin sun⚡️ @binbinyiyilili]

Chamath 在 All-In 上说：软件 IP「两天前就基本不值钱了」。他指的是 ArtCraft 用 Rust 从零重写的一套 Adobe 开源替代：PhotoCraft、VectorCraft、FilmCraft、LightCraft 等，开发时大量借助 Claude。不过它们是按公开规格做的 clean-room 实现，并不是他说的“反编译”，目前还是早期 alpha。界面和菜单变便宜了，专业工作流、插件生态和文件格式才是 Adobe 真正的护城河。

[The All-In Podcast @theallinpod]

Chamath: Software IP Just Became Worthless

“I think that IP and software was effectively rendered worthless two days ago.

There have been people that decompiled the entire Big 5. What is the Big 5?

Adobe has five very, very important products in their product suite, and these folks basically decompiled it, distilled it, and published open source versions of these things.

I was so shocked because I was like, ‘Hold on a second. These are profound products that have existed and have had moats for years, and now they're just open source versions.’

If you think about it, when everything is headless, so when Elon makes his services headless, Benioff made Salesforce headless a few months earlier, when all of these things are headless and you can wrap them in agent tools and you can connect them via MCP to other experiences, the value of IP no longer exists in software. It doesn't exist.”

[bin sun⚡️ @binbinyiyilili]

ChatGPT 的 dots 可以直接在手机上创建了：iOS 和 Android 版 ChatGPT App 里就能新建、命名、给自己的 dot 发消息。同批更新里，dot 还能在 Codex 里开新任务、跟进已有线程，会参考你的 ChatGPT 对话、Codex 线程和自动化，也更会判断该续旧线程还是另开一个。常驻 Agent 从桌面搬到手机，随手派活的门槛又低了一截。

[ChatGPT @ChatGPT]

Fresh updates for dots.

First up: you can now create your dot straight from your phone in the ChatGPT app on iOS and Android.

[bin sun⚡️ @binbinyiyilili]

阿里通义开源 Qwen-Image-2.1-Turbo：在 Qwen-Image-2.1 同一套 7B 架构上做的加速版，只要 8 步去噪就能出 2K 图，还能用自然语言接着改图（加配饰、换场景）。权重已上 Hugging Face 和魔搭，Diffusers 直接加载；Pro 和 Turbo 的 API 也同步上线。想本地跑图又嫌慢的，可以先试这个。

[Qwen @Alibaba_Qwen]

🚀 Meet Qwen-Image-2.1-Turbo — create and edit images in just 8 denoising steps! Open weights now available!

Built on Qwen-Image-2.1, Turbo is an accelerated checkpoint on the same 7B visual generation architecture.

Fewer steps does not mean lower quality: it still generates strong 2K images from text, and supports continued creation through natural-language edits, from adding accessories to changing a scene.

Start directly with Diffusers: load QwenImage21Pipeline and the checkpoint’s recommended 8-step sampling schedule is ready to go.

🌐 Prefer an API? Qwen-Image-2.1 Pro and Turbo APIs are now officially live. Choose a hosted API for your application, or use the Turbo weights in your own workflow.

🤖 ModelScope:
https://t.co/Cixiro51v8
🤗 Hugging Face:
https://t.co/4GqXcYeAmB
🔗 Pro API:
https://t.co/fw2am7VCP7
🔗 Turbo API:
https://t.co/qmDvTqiVZK

What will you create with 8 steps? Share your results with us!

[bin sun⚡️ @binbinyiyilili]

微软发布 Microsoft-Decision-1：专做“快速决策”的模型，只管是非、选择、打分这类结构化判断。纳德拉称它在延迟和质量上都超过通用大模型和其他决策模型，微软内部已在事故响应、质量控制、科研发现里试用。前几天 OpenAI 也推了 Decisions API——Agent 里“只选不写”的小快模型，正在变成标配。

[Satya Nadella @satyanadella]

Introducing Microsoft-Decision-1, our new model for fast decision-making.

It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality.

We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.

[bin sun⚡️ @binbinyiyilili]

@addyosmani Setting instructions, repo and memory once per project is the part that saves the most friction. 跨会话的工作不用每次重新交代背景，这点比单纯并行更有价值。

[bin sun⚡️ @binbinyiyilili]

@shengtang135754 2 块钱一期，提升主要来自接入 Pixabay API 换了素材源，说明这类视频的上限更多取决于素材而不是模型。Haiku 5.5 负责编排足够，用不着上 Opus。

[bin sun⚡️ @binbinyiyilili]

@AnthropicAI 从射电到伽马射线都有全天图，紫外却一直有大片空白，这次由天体物理学家主导、Claude 执行补上了。A good template for AI in science: the expert sets direction, the model does the heavy lifting.

[bin sun⚡️ @binbinyiyilili]

@OpenAIDevs @Microsoft Windows 上终于有系统级的沙盒（MXC）可用，文件访问能细到具体路径，这对在 Windows 写代码的人是实打实的改善。Native sandboxing on Windows 11 closes a big gap vs macOS/Linux setups.

[bin sun⚡️ @binbinyiyilili]

@kemu_aii Sonnet 干活、Haiku 做调研、Opus 只当顾问，这个分层很省 token。代价是切换模型时上下文要交接清楚，否则 Opus 给的建议容易脱离实际代码。