灰机仔 retweeted
Introducing Zero Wire: a public vulnerability intelligence feed on @CloudflareRadar, bringing the clarity of a news wire service to critical security developments. cfl.re/4iCvSbj
Many of those recent reimplementions and rewrites were done in Rust. However, to understand the implicit tradeoffs of this choice, an ablation study should be performed to see if/how faster and better the agent would do the same using the C language.
DeepSeek Harness 团队分享了他们关于 superagent 的思考:
这篇文章探讨的是一种叫 Super-agent 的智能体架构。它的核心目标是让 AI 能像人类员工一样,持续在线、同时推进多项任务,而不是每次都要人重新交代背景。Super-agent 的目标就是让 AI 从“对话工具”进化成“自主员工”。
文章给出的解法是把系统拆成几层:
双层架构:Router + Worker
系统被拆成两个角色。Router 是“前台接待”,负责接收所有消息并判断该交给谁处理;Worker 是“后台员工”,接到任务后自己闷头干。如果所有任务都挤在一个模型里处理,一来上下文会爆炸,二来一个耗时任务会堵住所有后续请求。拆开之后,Router 轻量快速,Worker 专注执行,互不干扰。这样,你在群里 at Agent 说“帮我查一下上周的销售数据”,Router 收到后把任务派给数据查询 Worker,同时你还能继续问“顺便看看这周的排期”,Router 又派给日历 Worker,两个任务并行跑。
事件驱动:不是每句话都要处理
系统不是对每条消息都调用大模型,而是先设一套规则过滤器。比如群聊里有人发“哈哈”“好的”,直接放过去;只有涉及任务、提问、at Agent 的内容才触发 AI 判断。这是因为群聊里 90% 的消息都是闲聊,如果每条都过一遍大模型,成本高、响应慢,还会让 Agent 变得“话多烦人”。用规则筛掉噪音,把模型资源留给真正需要判断的内容。
按需读取上下文
Agent 不会把整个聊天历史都塞进模型,而是只取最近几轮关键对话。更早的历史存到外部数据库里,当模型需要引用某条旧消息时,通过引用链去检索。这是因为大模型的上下文窗口再大也是有限的,塞满历史不仅浪费 token,还会让模型注意力分散,回答质量下降。比如你三天前让 Agent 记了一个需求,今天问“那个需求怎么样了”,Agent 不会把三天聊天记录全读一遍,而是直接去数据库找到那条需求记录,精准回答。
异步执行
Agent 接到耗时任务(比如生成一份几十页的报告)后,会在后台跑,前台立刻返回“正在处理,稍后通知你”。你该干嘛干嘛,不用盯着转圈。如果每次都要等模型跑完才能继续对话,用户体验极差。异步是让 Agent 真正“多任务并行”的关键。用户不需要知道任务什么时候跑完,Agent 跑完了自然会通知。
存算分离:本地轻量,云端重算
Agent 的会话状态(比如当前任务进度、上下文)存在数据库里,计算(模型推理)放在云端。本地只装一个轻量程序,用来连接代码仓库、文件系统这些工具。如果所有计算都在本地跑,普通电脑根本扛不住大模型。如果所有状态都存在云端,断网就全丢了。存算分离让两者各司其职——本地负责“连接”,云端负责“思考”。用户换设备、断网重连,Agent 的状态不会丢,因为状态在数据库里,不在本地内存里。
Super-agent 这套架构的核心思路就是:别让 AI 什么都自己扛,把听、想、做、记”拆开,各找最合适的工具来干。 这样它才能像一个人一样,同时处理多件事、记住上下文、不被打断、也不打断你。
mp.weixin.qq.com/s/zoS7GlReW…
从文件大小我们可以管窥一下 Fable5 的参数量了
🚨🇺🇸 Anthropic model weights allegedly exposed through misconfigured AWS S3 bucket
An actor using the handle "localizer" claims to have obtained a 7 GB shard of Anthropic's alleged "Claude Fable 5" model weights from a misconfigured AWS S3 bucket associated with Amazon Bedrock.
The actor claims the complete model consists of 640 shards totaling approximately 8 TB, but says only one shard was retrieved before access was restricted.
Screenshots showing a .safetensors file, model metadata, and a hexadecimal dump have been published as alleged proof. The actor is now offering the material for sale.
Claim is currently unverified. The authenticity of the files and their connection to Anthropic have not been established.
💥 No delays. No guessing. No redactions. Get the intel threat actors see, the moment they post it. darkwebinformer.com/pricing
Readers added context they thought people might want to know
The model shown here is not a Claude model, but DeepSeek V4.1 Flash. This can be verified by the fact that the file sizes match those of DS V4.1 1017. The screenshot also includes "eu-north-1," even though it does not host Fable.
x.com/torisetxd/stat…
x.com/YipingDeng5/st…
x.com/olddonkeyblog/…
我教你怎么用模型系列
ANTHROPIC IS CRACKING DOWN ON VIDEO GAME REWRITES 🚨
Claude Code picked up a new rule around Oct 27, sent from Anthropic’s servers under a flag called "tengu_heron_brook". There’s no setting to turn it off.
The full rule: “Do build original games, games in a genre, and games from the user’s own designs and assets. Don’t build a recreation of a specific existing game, even if it is renamed or called original.”
Is this the end of game rewrites, modding and everything in between?
emmm
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.














