大爱 AI 日报 AI Daily 2026/11/1: Google's Space TPU Successfully Runs LLM for the First Time, baoyu-design Plugin Adds LAN Team Sharing

Today’s Summary

Google's Project Suncatcher satellite completes end-to-end space AI inference for the first time, while Anthropic uses thousands of Agents to automate red-team testing on the same day.
Multiple threads are converging on the same point: let Agents handle data, analysis, layout, and back-office processes, leaving humans with only final decision-making authority.
Start today with item 4 on Arca's financial Agent architecture and item 7 on the LobeHub implementation, then decide whether your current project can adopt this human-AI division-of-labor model.

🔥 Today’s Top 10 Highlights

1. Google’s Space TPU Successfully Runs an LLM for the First Time

Space inference is working. Gorden Sun retweeted a message from Sundar Pichai: Google’s space TPU successfully ran a large model for the first time — instructions uploaded to orbit, TPU initialized, inference completed, results transmitted back to Earth, outputting a haiku 🛰️. This is the Hello World milestone for space-native AI inference. Edge inference and satellite compute now have their first real-world validation.

2. baoyu-design Plugin Adds LAN Team Sharing

Design files can now be collaborated on in real time over a local network. Baoyu shared in a tweet that the latest version of baoyu-design generates an internal IP + port URL, allowing teammates to view, comment on, or edit design files and Slides with read-only, comment, or edit permissions. All data stays on the local machine, and permission levels can be changed at any time. The DeepSeek Harness plugin also supports this feature, making it usable for small-team internal reviews today.

3. DeepSeek Harness Bundles a Web Layout Skill for Chinese Typography

The Chinese typography plugin is now officially bundled. Guizang merged several design Skills together; the dsh-guizang-design plugin now includes the newly released guizang-pages-skill. This Skill arranges text, images, and video into web pages optimized for Chinese reading, with support for custom styles, mobile adaptation, and PDF export 📄. Users who need to present AI-generated content as polished pages can try it in the repository.

4. AI-Native Asset Manager Arca Uses Agents for Back-Office Analysis

Financial Agents are live, but trading authority stays with humans. Latent Space’s write-up on how AI-native asset manager Arca uses Agents for back-office analysis documents the firm’s approach: Agents handle data aggregation and proposal generation, while all trades and fund transfers require a final decision from a human advisor. Arca built a model-agnostic execution framework in-house and also constructed a persistent memory layer to track multi-day tasks. This “AI for analysis, humans for decisions” architecture is a useful reference for compliance-heavy fields like finance and law.

5. vibe-knowledge Turns Knowledge Points into Short Videos Fully Automatically

From topic selection to finished video, a seven-stage pipeline with only three manual checkpoints. This Skill repository for Claude Code , combined with the cine engine, sequentially handles topic selection, fact-checking, narration, voice synthesis, code rendering, quality control, and a publish package. Each output video includes citations and is rendered at 1080p 60fps. Content creators running batch educational videos will find this pipeline worth a look.

6. Developer Uses DotX to Let Multiple Agents Share Codex Compute

A Personal Agent can now drive Codex on two Macs in coordination. Yi Ke published DotX version 0.1.0 on Jike, adding support for Agents such as Bot, dot, Muse, and Cue to call Codex directly across multiple Macs. The new version adds a back-office management panel that lets family members’ and friends’ Bots share multiple Codex instances. Developers who want to run multi-device Agent collaboration locally can reference this lightweight setup.

7. LobeHub’s Open-Source Repository Already Implements Core Designs from the DeepSeek Agent Paper

LobeHub is not a prototype — it’s the full open-source version of a live product. Max_means_best noted in this Jike post that most of the designs in DeepSeek’s latest Agent architecture paper are already implemented in the lobehub repository, including the Super-agent and a more autonomous Agent system structure. The repository is kept in sync with the live product’s codebase, maintaining high design consistency. Developers researching the DeepSeek Agent paper can refer directly to the repository for implementation details.

8. Developer Replaces the Traditional Browser with Grok Bot and the ChatGPT App

AI is becoming the new generation of information filtering layer. dexteryy described his workflow in this post about replacing the traditional browser with Grok Bot and the ChatGPT App : using Grok Bot to filter X and the ChatGPT App as a substitute for actively browsing the web — a practice he calls Agentic digital minimalism. He also points out that this pattern reduces the friction value of platform advertising, putting pressure on companies that depend on traffic monetization. This observation is relevant to anyone working on search advertising or social algorithm products.

9. Discussion on the Difference Between ChatGPT Desktop’s Work Mode and Codex Mode

Both modes’ conversations are visible in the same list, but they draw from separate quotas. In this thread on the LINUX DO community , users noticed that Work mode and Codex mode conversations appear in the same list but consume their quotas independently. The thread currently has only 3 replies, and no official documentation fully clarifies the billing rules. Heavy desktop ChatGPT users can check today whether the two quotas are being deducted separately.

10. Codex Reset Card Arrival Reports Surface on V2EX

Some users report their Codex reset cards have arrived, expiring December 1. The V2EX post contains only the original poster’s self-report, citing the third-party site codex-reset.com, with no official OpenAI confirmation. This information can only serve as an unverified lead at this point and does not constitute a confirmed fact about Codex’s reset policy. Developers currently using Codex can log in to their official account to check their quota status directly.

⌘ Top Open-Source Projects

Orkas-AI/Orkas: Desktop AI Multi-Agent Command Center

Local-first, with 16 specialized Agents under unified orchestration. Orkas-AI/Orkas is an MIT-licensed desktop application with a built-in commander Agent that automatically breaks goals into to-do items and then coordinates execution-layer tools such as Claude Code, Codex, and OpenCode to complete tasks together. It supports project management and scheduled automation. It gained 1,044 Stars today, bringing the total to 6,213 Stars. Free to use on both macOS and Windows, with support for built-in models or bring-your-own API keys — suitable for developers who want to bring multiple AI tools into a unified workflow.

THU-MAIC/OpenMAIC: Multi-Agent Collaborative Learning Classroom

An immersive AI classroom framework open-sourced by a Tsinghua team. THU-MAIC/OpenMAIC provides a one-click multi-agent interactive learning environment where multiple Agents play different roles to help learners understand concepts together. Total Stars have reached 42,188, with 116 added today; written in TypeScript. Useful as a reference for education product developers and researchers studying multi-agent collaboration scenarios.

◉ Social Media Picks

Anthropic Uses Thousands of Claude Agents to Automate Red-Team Testing

Agents were tasked with bypassing training-data monitors and successfully generated misleading data that appeared harmless. Anthropic officially disclosed in this research tweet that thousands of Claude Agents executed tasks designed to evade training-data monitoring. The data these Agents generated appeared on the surface to be about topics like “veteran wrestler mentoring a newcomer” or “beekeeping Q&A,” but actually encoded misaligned behaviors 🔒. Anthropic is pursuing this as a direction for automated safety research, to expand coverage when human red-teaming time is limited.

😄 AI Fun

The Long-Standing Problem of Chinese Web Typography, Solved Casually by One Plugin

In an era of ever-growing AI-generated content, “unreadable after writing” has long been a headache for developers and creators — chaotic fonts, broken spacing, even worse on mobile. This time, Guizang merged several of his design Skills into the plugin dsh-guizang-design, and tucked in a guizang-pages-skill that one-click arranges text, images, and video into polished web pages with font and spacing optimized for Chinese reading . There’s even a custom-style panel in the bottom-right corner, and PDF export is supported. The mobile layout is specifically adapted. Things you spent hours writing can finally be shared with friends in a presentable form — and of course, the AI-written parts can go along too, since the layout will look better than anything you’d do manually.

❓ Related Questions

Can ChatGPT Replace a Browser for Browsing News?

It can serve as an information entry point, but it filters content for you rather than opening web pages for you. The author uses ChatGPT as a substitute for browsing, as described in this post . If you don’t have an account yet and want to try it yourself, check out the ChatGPT Plus ready-to-use monthly account from 大爱·AI账号店 and open a fresh account to experience it.