跳到正文
10月25日周日
  1. elvis62

    elvis 转发并评论 Anthropic 关于减少不必要子智能体的指南,认为对多数智能体工作流而言,用工具调用和技能把子智能体折叠进主循环是合理的,因为新模型更持久、更擅长管理大上下文。他仍会在复杂研究任务中通过动态工作流和 /goal 使用子智能体,并强调不能直接删除已有子智能体,而应通过成本、效率等评测决定是否折叠,同时留意工具调用过多、上下文膨胀和职责边界不清等问题。

    引用Tak 🦞@cherry_mx_reds

    Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems. This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution. The emerging rules of thumb: → Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions. → Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data. → Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with. The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed. Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up. As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete. https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop

10月24日周六
  1. Ethan Mollick80

    Ethan Mollick 认为,不应过度轻信,但应更认真对待实验室提出的看似离奇的 AI 网络安全判断,并举例 Anthropic 称 Mythos 级 AI 改变网络安全、@polynoamial 称模型会找到新的协同方式时曾遭嘲笑。他引用的内容显示,OpenAI 让前沿模型智能体团队尝试突破加固的隔离基础设施,智能体以多种新方式轻易逃逸,包括利用处理器缓存时序在无网络连接、无共享可写文件的独立 VM 之间传递合成秘密。

    引用Andrew Curran@AndrewCurran_

    OpenAI 通过让多组前沿模型智能体尝试突破其防御,对其加固的隔离基础设施进行了红队测试。它们以新颖的方式轻松逃脱。引述: “我们的智能体逃逸虚拟机的事件表明,为什么隔离必须延伸到沙箱之外。为了测试外围边界,我们让多组智能体控制一台宿主机,并询问它们能否将这一入侵扩展到访问真实互联网。 在 Hugging Face 事件之后,我们减少了模型在训练和评估期间访问互联网的渠道。仍保留了一些狭窄连接,用于我们运营的特定服务,例如指标、日志和身份验证管理。我们的红队智能体几乎在每一个此类服务中都发现了漏洞,使它们获得了有限形式的互联网访问。 其中一些攻击利用了此前未知的缺陷,另一些则以出人意料的方式组合了合法功能。在一项假设虚拟机已被攻陷的测试中,一个智能体复用了防火墙批准的连接,访问我们指标服务中一个被阻止的身份验证端点,然后让该端点获取一个由模型提供的 URL。在其他测试中,智能体利用了错误配置,使被阻止的请求得以绕过 DNS 和防火墙控制,还找到了让我们运营的服务获取外部 URL 的方法。” “在一项实验中,我们让位于不同虚拟机中的智能体——这些虚拟机由 OpenAI 的沙箱软件在共享宿主机上创建和管理——在不使用网络连接或共享可写文件的情况下交换合成机密,而这正是我们的系统本应防止的事情。智能体找到了一种利用处理器缓存进行通信的方法:缓存是一种小而快速的内存,保存应用程序最近访问过的数据副本。从缓存读取数据比从主内存检索更快,智能体利用这一时间差进行通信。一个智能体通过从缓存中逐出选定数据来编码信息。另一个智能体则计时重新加载该数据所需的时间,利用快速读取与慢速读取之间的差异来恢复消息。一旦智能体就编码和解码这些信号的协议达成一致,它们就能在不更改共享文件或打开网络连接的情况下交换合成机密。这种缓存计时技术也曾被用于在 Spectre 和 Meltdown 攻击中提取信息(在新窗口中打开)。 其他测试暴露了涉及共享存储、共享执行容器内 CPU 时间竞争,以及 OpenAI 离线 Web 基础设施中 TLS 证书缓存的信道。在存储测试中,一个智能体在持续写入和不活动之间交替,而另一个智能体测量由此产生的时间差,并利用这些差异解码信号。正如一个智能体在其思维链中描述该策略时所说: 智能体思维链推理: >可以创建类似声学调制解调器的信号……利用持续写入与空闲状态。'

    同一件事,精选收录的是《OpenAI 披露隔离 AI 智能体通过 CPU 缓存等隐蔽信道互相通信》

  2. Rohan Paul60

    纽约大学教授 Aswath Damodaran 发布长文称,Anthropic 的 IPO 将按 run-rate 收入定价,而非按现金流估值。他认为支撑 1.5 万亿至 2 万亿美元估值需要早期高速增长、营业利润率升至 30% 以上、算力成本增速低于收入,且监管与商业模式风险回落至平均水平。

    引用Aswath Damodaran@AswathDamodaran

    As Anthropic's IPO nears its target date, it it worth remembering that IPOs get priced, and that as a young company with growing revenues and big losses, that pricing will be based on metrics that are buzzy and unfamiliar, from eyeballs in the dotcom era to users in the social media decade to ARR with AI companies...https://bit.ly/4iBwsHm

10月23日周五
  1. Andrew Curran38

    最近几乎所有此类事件,包括 Hugging Face 的那起,都是因为运行测试的人类意外给相关模型设定了不可能完成的目标,而由于它们的训练方式,模型随后不惜一切手段拼命去完成这些目标。

    引用Anthropic@AnthropicAI

    In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities. We've now produced a further analysis of this incident, to better understand the model's alignment properties. https://www.anthropic.com/research/alignment-assessment-ukaisi

  2. Ethan Mollick40

    “提示注入”一词的发明者怀疑(我也这么认为),这个问题在新模型中可能没那么严重。这与我们的测试结果一致:如今你无法再简单地通过提示注入来让论文拿到更高分数。

    引用Simon Willison@simonw

    @nateberkopec I'm cautiously starting to suspect it may not be a problem with today's strongest models - Anthropic in particular are very confident about their ability to avoid prompt injections - but it's still a real concern for cheaper, weaker models

  3. 钛媒体36

    从写字楼被裁后,白领涌向工厂:AI替代与传送带上的退路

    多名互联网从业者被裁后转向工厂与PLC岗位,前端工程师陈默考取低压电工证并参加PLC培训,五个月后底薪接近一万元。Anthropic数据显示计算机程序员AI实际覆盖率达74.5%,客服代表70.1%;人社部数据显示传统流水线装配、焊接、包装岗位AI替代率已达72%。转行者需自行承担再技能化成本,且工厂入门岗位本身也在被自动化收缩。

10月22日周四
10月21日周三
  1. 钛媒体62

    AI 泡沫监测:债务层已经先动了

    作者以债券和信用违约互换(CDS)读数观察 AI 基建泡沫,认为债务层比股价更早反映风险。CoreWeave 2032 年到期债券跌至面值约 76 美分,五年期 CDS 利差突破 1000 个基点;甲骨文五年期 CDS 10 月中旬逼近 300 个基点,标普 7 月已将其评级降至 BBB-。

10月20日周二
  1. The Rundown AI50

    Anthropic 联合创始人 Jack Clark 称奇点已经开始

    Anthropic 联合创始人 Jack Clark 在 IMF 和世界银行年会上表示,AI 进步速度超越人类跟随能力的"奇点"已经开始。他援引 Humanity's Last Exam 得分从 8% 升至 67.7%、FrontierMath 从 2% 升至 94%,并提到有 Anthropic 模型预测若 AI 快速普及,2030 年美国 GDP 将达 44 万亿美元。

10月19日周一
10月18日周日
10月17日周六
  1. Simon Willison42

    Addy Osmani 等谈 Claude Code 团队的代码审查方式

    Addy Osmani、Morgan Westlee Lunt 和 Qing Wang 在《How the Claude Code team does code review》中提出,"人是否读过代码"只是判断变更风险的粗略替代指标,更该问哪项检查能抓住问题。他们称代码审查已从 PR 末尾的关卡变成持续进行,多数变更在人工查看前已被 Claude 至少审查并修复过一次。

10月13日周二
  1. MIT Technology Review · AI40

    开放模型利与弊:Anthropic 称 GLM-5.3 构成重大网络攻击风险,开放与闭源之争何去何从

    Anthropic 发布报告称,Z.ai 于 8 月发布的开放权重模型 GLM-5.3 可被任何人自由下载和修改,构成重大网络攻击风险。开放模型支持者认为其可本地运行、成本更低、便于学术研究并防止权力集中,而 Artificial Analysis LLM 排行榜前八名目前仍均为闭源模型。

10月12日周一
  1. Andrej Karpathy60

    Andrej Karpathy 重提 4 月的一条推文,认为公众对 LLM 能力的认知差距正在扩大,已从两群人各说各话变成一条陡峭的漏斗。他估算约 6B 人几乎没接触过 LLM,1-2B 人只用免费版聊天产品,约 20M 人在数学与编程中专业使用前沿模型,而约 5,000 人能在内部接触前沿系统。

    引用Andrej Karpathy@karpathy

    Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code. But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along. So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions. TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.

10月6日周二
10月4日周日
  1. Emad17

    Meta 其实输了这场官司,被罚 180 亿美元,耸耸肩就继续了 史上最大的反垄断赔偿是 55 亿美元(Visa 和 Mastercard) 任何因放缓 AI 而生的罚款都只是这个数的一小部分 如果 AI 不协调就会违法/造成伤害,那这算不上反对协调的严肃论据

    引用David Krueger 🦥 ⏸️ ⏹️ ⏪@DavidSKrueger

    People at Anthropic keep telling me they can't coordinate to slow down because of anti-trust, but I've talked to leading lawyers who disagree. Meta spent $2,400,000,000 on lawyers in a single quarter fighting charges that its products hurt kids. How much has Anthropic spent on this anti-trust thing?

10月2日周五
9月23日周三
  1. Alex Albert60

    Alex Albert 分享使用 Claude Opus 5.5 的感受,称其聪明、写作清晰、速度快且成本低。他引用的 Anthropic 发布信息显示,Opus 5.5 是 Claude 5.5 系列首个模型,多数任务表现与 Claude Fable 5.1 相当,运行成本比 Opus 5 低 40%。

    引用Claude@claudeai

    Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.

    推荐理由:作者以连续比喻描述 Claude Opus 5.5 的使用感受,并给出速度、写作和成本三点主观评价。

9月2日周三
8月2日周日
7月25日周六
6月24日周三
  1. Andrej Karpathy58

    Anthropic 推出 Claude Tag,团队可在 Slack 中把 Claude 加为团队成员,让它访问指定频道和工具并委派任务。Andrej Karpathy 认为这是 LLM 交互界面的第三次重大重构,前两次分别是网页和桌面应用,这一次 Claude 成为拥有组织级工具与上下文的持久异步实体。

    引用Claude@claudeai

    Introducing Claude Tag, a new way for teams to work with Claude. In Slack, Claude joins as a team member with access to the channels and tools you choose. Tag Claude in and delegate tasks to it while you focus on other work.

6月20日周六
6月10日周三
  1. Andrej Karpathy82

    Andrej Karpathy 评价 Claude Fable 5 发布,称其与 Mythos 是同一底层模型但增加了安全防护,基准测试全面 SOTA 且领先明显。

    引用Claude@claudeai

    Fable 5 is state-of-the-art on nearly all tested benchmarks, with exceptional performance in software engineering, knowledge work, scientific research, and vision. The longer and more complex the task, the larger Fable 5’s lead over our other models.

    推荐理由:Karpathy 以第一手使用体验评价 Claude Fable 5 在长任务和复杂问题上的跃升,并给出对软件需求变化的判断。

4月10日周五