My experiences with Claude Fable 5:
First, discuss with Fable using Plan mode. Then, they’ll provide a detailed development plan. This is similar to how things are done with Sonnet 4.6/5.
During the development phase, Fable first operates in “auto mode”. It then works hard on its own until a result is achieved. There’s little need for human intervention along the way. Fable identifies errors and fixes them on its own, until it produces a result it’s satisfied with, at which point it “turns in its work”.
One observation: Claude Fable 5 has begun to enter the “Loop Engineering” phase. Self-evolution is something that other models currently lack.
The trigger for this recent sharp sell-off was a market misinterpretation of Meta's plan to sell off idle computing capacity. In reality, the capacity Meta is leasing out consists of older, legacy hardware suited for inference tasks, while it retains all its core, cutting-edge resources for model training; furthermore, the company continues to lease massive amounts of advanced computing power from third-party providers and has raised its full-year budget for AI hardware procurement. There is no oversupply of mid-to-high-end computing capacity.
Currently, virtually no cloud providers in the US or China have signaled plans to cut AI-related capital expenditures, suggesting the AI market rally is likely to persist; the recent pullback represents a digestion of gains accumulated during the prior run-up.
anthropic也做了龙虾/爱马仕hermes这类ai员工的功能:今天发布了 Claude Tag,让 Claude 以同事的身份常驻 Slack 频道。团队成员在频道里 @Claude 就能分配任务,Claude 会把任务拆成几个步骤,逐步完成后在 Slack 线程里交付结果。
目前以 research preview 形式提供,Claude Enterprise 和 Team 客户可以使用。
跟之前 Slack 里给 Claude 发私信或者在频道里临时呼叫不同,Claude Tag 有几个变化。
一是多人协作。
一个频道里只有一个 Claude 身份,所有人共享同一个对话上下文。你让 Claude 做了一半的事,同事可以接着往下聊,不用从头解释。
二是持续学习。
Claude 会持续关注频道里的对话,逐渐积累对团队工作的理解,不需要每次都重新交代背景。管理员可以授权它读取其他频道的信息,让它更快了解整个组织的情况。
三是主动干活。
开启 ambient 模式后,Claude 会主动推送它认为你需要知道的信息,跟进没人回复的线程,提醒被遗忘的任务。Anthropic Claude Code 产品负责人 Cat Wu 举了个例子:她把自己的 Claude Tag 连上了 Gmail,Claude 会自动读邮件,遇到重要邮件就在 Slack 里通知她。
权限控制方面,管理员可以精确指定每个频道的 Claude 能访问哪些工具和数据。法务频道的 Claude 不会把信息带到工程频道,也不会让工程师接触到法务数据。
Anthropic 自己已经是重度用户。目前产品团队 65% 的代码由内部版 Claude Tag 生成。用途也不限于写代码,团队还用它查产品数据、处理客服工单、排查疑难 bug。
Claude Tag 会替换现有的 Claude in Slack 应用,管理员有 30 天的迁移窗口。Anthropic 会给符合条件的企业和团队账户发放体验额度。底层模型用的是 Opus 4.8
AI infrastructure runs on more than compute. It runs on light.
Today in Sherman, Texas, Jensen Huang joined @CoherentCorp CEO Jim Anderson and local leaders to mark the expansion of U.S. manufacturing capacity for advanced photonics.
The technology is behind the lasers and optical links that move data across modern AI systems, strengthening the American manufacturing base behind the AI era.
Ayar Labs' partnership with @nvidia enables hyperscalers and system innovators to build optically connected AI infrastructure around NVIDIA's NVLink Fusion platform and partner ecosystem.
https://t.co/iyQSL9kkV9
#optical#photonics#semiconductor#AI#HPC#CPO