大爱 AI 日报 AI Daily 2026/11/11: Anthropic Releases Two New Models on the Same Day
每日 AI 行业动态、产品更新、前沿研究、开源项目与社媒信号
Today’s Summary
Anthropic releases Fable 5.2 and Mythos 5.2, with improvements across coding, reasoning, and vision.
Open-source models are challenging commercial benchmarks while software shifts from human-use to agent-use.
Try the two new models first, then evaluate whether agent-skills can be integrated directly into existing projects.🔥 Today’s Top 10 Highlights
1. Anthropic Releases Two New Models on the Same Day
Fable 5.2 and Mythos 5.2 launch simultaneously. Anthropic’s official account announced that both models are available today, positioned as collaborative tools for tackling the most demanding technical challenges. Both models bring improvements in coding, reasoning, and vision. Users with coding or complex reasoning needs can now switch between them directly on the Claude platform.
2. Open-Source World Model PAN 2 Surpasses NVIDIA Cosmos
Open-source diffusion model outperforms commercial benchmarks. PAN 2, released by MBZUAI, combines a 29B-parameter diffusion video generator with LLM reasoning. A technical summary compiled by BestBlogs shows a task-planning success rate of 34%. Model weights are 58 GB and available for download under the Apache 2.0 license. Developers working on autonomous driving or robotics simulation can pull the weights directly from Hugging Face, though physical consistency remains unstable in roughly two-thirds of simulated scenarios.
3. Claude Adds Whiteboard and Spreadsheet Collaboration Features
Whiteboards and Sheets now available on all paid plans. Anthropic added Whiteboards and Sheets to Claude on November 10. As shown in this post about Claude’s new whiteboard and spreadsheet features , both features are rolling out in beta across the web, desktop, and mobile apps. Claude Data, launching at the same time, is the official release of the renamed dashboard and requires connecting a data warehouse or CRM data source first. It’s normal if the entry point isn’t visible in your account yet — just describe what you need in a conversation to trigger it.
4. MiniMax Open-Sources H3-Image Generation Model
MiniMax image model weights are now publicly available. Gorden Sun posted that MiniMax has open-sourced H3-Image , an image generation model derived from the H3 video model. It supports generating images at 4K×4K resolution and accepts up to 16 reference images. The model page briefly returned a 404 at launch, with weights still uploading. Developers who need high-resolution image generation or multi-reference image compositing can watch the Hugging Face page for the upload to complete.
5. Driving Video from a Single Image with Physical Plausibility — Waymo Benchmark Tops the List
Motion Forcing ranks first on physical plausibility. Synced reported on this new framework from HKUST(GZ) , which generates driving videos containing long-tail scenarios such as lane-cutting and hard braking from a single input image. It ranks first on physical plausibility in the Waymo dataset benchmark, and the framework code is open-sourced. Autonomous driving data teams can use it to quickly augment training data for rare scenarios.
6. Tencent AI Gaming Companion Agent “Goose Dimension” Enters Beta
Tencent’s Goose Dimension enters the gaming companion space. AIBase reports that the product can perceive game screens in real time and provide instant Q&A, with support for QQ cross-device chat, desktop pet, and voice-changer interactions. The first batch of supported games includes Valorant, League of Legends, and other mainstream titles, with 4 AI characters available at launch. Users interested in trying it can visit eciyuan.qq.com to apply for beta access.
7. Grok Bot Gains Tailscale Access to Any Server
Grok Bot now supports direct Tailscale connections. A V2EX user posted about Grok Bot’s new Tailscale support for accessing any server , noting that Grok Bot can now connect to any server via Tailscale, while OpenAI Dots currently only supports Ubuntu 24.04 and 26.04 desktop editions and does not allow installation of third-party networking tools. Developers with multi-device remote access needs should factor this difference into their tool selection.
8. Xiangyang Qiaomu Open-Sources X Long-Form Auto-Translation Website Project
Open-source tool automatically scrapes X long-form posts and publishes them as a Chinese website. Xiangyang Qiaomu announced this new project , which uses Grok Bot to daily scrape and analyze high-quality long-form posts from X, automatically translates them into Chinese, and publishes them as an accessible website. The project can be adapted to monitor specific researchers or model company accounts as an information feed. Developers can find the demo website address and open-source repository link in the post’s comment section.
9. AI4M Closed-Door Session: China’s Manufacturing Data Volume Does Not Automatically Equal Advantage
Large data volume does not automatically translate into an AI materials advantage. Ji Shisan’s notes from the Future Lightcone AI for Materials closed-door session gathered perspectives from over 20 entrepreneurs, scientists, and investors. The core conclusion: more factories and longer supply chains, but records of failed samples are scattered and inconsistently formatted, with no way to link formulations to processes. The real barrier for automated laboratories lies in stable execution over long workflows — a successful single-step demo does not mean reliable operation at scale. Teams entering the AI4M space need to start by planning their data infrastructure, not their model capabilities.
10. State Media Coverage Is Not a Shield — AI Entrepreneurs Need a Clear Head
Media exposure cannot substitute for genuine product credibility. The·Muqiu posted on Jike pointing out that Evergrande, LeEco, WM Motor, and Neta Auto all appeared on CCTV or People’s Daily, yet none avoided their eventual crises. The post directly names founders in AI applications, embodied intelligence, and world models. For AI entrepreneurs seeking state media endorsement, product capability and business sustainability remain the only real moats.
◎ Industry Shifts and Personal Impact
AI Accelerates Development, But Debugging Instincts Are Quietly Atrophying
Coding skills are being displaced by AI, but the cognitive cost is underestimated. A developer shared their experience on Juejin : screenshot the error, send it to AI, wait for a solution, copy and paste. The debugger panel gets opened less and less. Things move faster, but the habit of reading source code line by line and tracing call chains has quietly faded. They mention that large-scale internationalization refactors and Vue upgrades were almost entirely AI-driven. For developers who built their skills through independent debugging, today is a good moment to ask: how much time are you still setting aside for thinking things through yourself?
⌘ Top Open-Source Projects
addyosmani/agent-skills: A Production-Grade Engineering Skills Library for AI Coding Agents
An engineering skills collection designed specifically for AI coding agents. All code for addyosmani/agent-skills is publicly available , maintained by Addy Osmani, a core contributor to Chrome DevTools, written in JavaScript, and containing production-ready engineering skills for AI coding agents. 465 new Stars were added today, bringing the total to 115,168. Developers building AI coding agents or looking for reference implementations of mature engineering skills can clone the repository and browse the available skills list directly.
◉ Social Media Picks
WordCraft Test: English Drafts Pass, Chinese Long-Form Needs a Font Fix
The open-source Word alternative has a Chinese display defect. Shikong Mao ran a full test of WordCraft — a browser-based document tool with 17,726 Stars, 4 templates, and 11 export formats. Chinese characters are preserved intact when exporting to docx. However, characters like「署」「品」「传」appear as boxes on screen, and PDF and PNG exports show the same boxes — the issue is that the web version lacks a complete Chinese font. The verdict: works fine as a scratch pad for English documents, not yet usable for Chinese long-form writing. Write first, then export to docx for proofreading.
😄 AI Fun
OpenAI Released a Handheld Console — The Comments Section Is Demanding a New Model
OpenAI’s merch store listed a Game Boy–style handheld console with a d-pad and cartridge slot for $200. The pitch: use Codex to describe a game in a few words, generate it, write it to a physical cartridge, and play. The comments section had other ideas — people are asking for GPT-6.2 . The handheld dream came true; the model update dream did not.
❓ Related Questions
Claude has added Whiteboards and Sheets — which plan do you need?
Anthropic has launched Whiteboards and Sheets for Claude, available on all paid plans. The features are currently in beta and rolling out in stages. Just describe what you need in a conversation to trigger them — no need to look for a dedicated entry point.