爱窝啦 AI 日报

Daily Digest

Anthropic accused seven Chinese companies of distilling Claude through hundreds of millions of requests, with Alibaba generating 150 million interactions between May and July.
Office assistant data leaks, Personal Agents taking over worry, Coze and Grok positioning in task management — the battle is for workflow control, not single execution.
Today we examine the authorization boundary debate in the distillation dispute, then try Phonon-2 for local audio transcription and context-mode to reduce context overhead.

🔥 Today’s Top 10

1. Anthropic Names Seven Chinese Companies in Alleged Claude Distillation

The accusations are in. Anthropic released a report naming Alibaba, DeepSeek, Moonshot, Zhipu, Xiaomi, SenseTime, and MiniMax, accusing them of “distilling” Claude’s capabilities through various methods . Alibaba allegedly generated 150 million interactions between May and July, while DeepSeek forwarded 12 million user requests within 14 days. The report argues this behavior reverse-engineers training data from model outputs. The controversy centers on authorization boundaries, technical path legitimacy, and how to distinguish normal usage from improper extraction.

2. Developer Discovers AI Editor Data Leak While Hunting for Disk Space

An accidental lifesaver. A developer tracking down cache bloat due to insufficient disk space accidentally accessed other users’ project data in ZCode. According to a WeChat article, his willingness to dig deep plus years of technical blogging made the difference — without either trait, the risk could have persisted much longer. Tencent, Alibaba, and ByteDance are pushing AI office assistants, meaning more sensitive information will flow in. Before using these tools, verify the vendor’s data isolation mechanisms and access permission audits 🔒.

3. Six Questions Reveal Your True AI Stance

Only after taking it do you know which quadrant you’re in. Jike user Cui Xiaotian shared a six-question, binary-choice AI tendency test . Questions cover trust levels, usage scenarios, and ethical boundaries. These tests help developers and product managers quickly identify team or user attitude differences toward AI. Takes 15 seconds to complete — see which category you fall into.

4. Personal Agent vs. Codex: Who Does the Worrying

More than just execution. Jike user benn explains that Codex completes specific tasks but still requires you to push things forward, while Personal Agents can take over the “worrying” itself . For example, on Sumus.im, agents report progress every morning, including tasks assigned yesterday and casually mentioned weeks ago. Mental bandwidth is a scarce resource — handing off long-term tracking lets you focus on higher-priority decisions. Try assigning an agent something that requires continuous attention.

5. Coze and Grok Bot Both Testing Office Scenarios

Chat boxes are starting to manage tasks. Jike user OrangeCLK posted that Coze and Grok Bot are both testing office scenarios . Office scenarios require bots that can access calendars 📅, pull data, and interface with external systems. If your team uses Feishu or DingTalk, try handing off repetitive communications to bots, like meeting notes organization or daily progress syncs.

6. A Skill That Turns Character Images into Minimalist Block Posters

Highly stable. Jike user Lanxi tuned a skill that works with any image generation model, converting anime, live-action, and game characters into minimalist block-style collage posters 🎨 with adaptive scene completion. GPT-Image-2.5 works best. Install with npx skills add lanxi-ai/flat-character-skill. Perfect for creators who need to quickly produce stylized visual assets.

7. Figure Lets Gen-2 Robot Jump into Molten Steel for Retirement

Retirement as tech protection. Baoyu relays that Figure AI had its Figure 02 humanoid robot 🤖 jump into molten steel at a Finnish foundry for retirement , with the melted metal to be made into limited commemorative items. Figure 02 once worked at BMW factories, ran the proprietary AI model Helix, and handled logistics tasks. Continued maintenance wasn’t cost-effective, but disassembly would consume engineering time and delay the Gen-4 release. Foundries in the US and Mexico refused to accept robots with lithium batteries. This method prevents core actuators from leaking out while freeing up team resources.

8. Using Muse to Auto-Generate AI News Videos Every Morning

The pipeline is working. Guizang shared that he uses Muse to auto-generate AI news videos every morning . The workflow involves organizing content channels in Notion, finding materials, calling K3 and GLM 5.3 models via Pi in a virtual machine, assembling videos with guizang-product-video-skill, and finally sending them back through a chat interface. This pipeline proves agents can chain together multiple local models, external tools, and custom skills. Try breaking down your own daily information-gathering routine into similar steps.

9. Gemini 4 Argon Expands Output Limit to 1 Million Tokens

Longer outputs. Google’s official blog announced that Gemini 4 Argon expands its output limit to 1 million tokens , designed specifically for complex long-cycle workflows 🧠. Currently rolling out to Fairwind program participants in cybersecurity defense. Ideal for scenarios requiring large code generation, lengthy technical documentation, or deep reasoning outputs. Broader availability has not been announced.

10. context-mode Reduces AI Coding Agent Tool Output by 98%

Context window optimization. mksglu/context-mode gained 90 new stars today on GitHub Trending. The project reduces context overhead by 98% through sandboxed tool outputs , supports persistent session memory, and enforces routing across 17 platforms via MCP + hooks. Total stars now 24491. Perfect for AI agent developers who need long sessions, multi-platform debugging, or heavy tool invocation.

⚡ Product & Feature Updates

Altman Believes Humanity Has Entered the AI Singularity

The tipping point is here. AI Exploration Guide channel compiled that Sam Altman believes humanity has entered the AI singularity , the threshold where artificial intelligence surpasses human intelligence. This judgment means model capability growth curves may no longer be linearly predictable. Developers need to reassess automation boundaries and the applicable range of safety mechanisms.

⌘ Top Open Source Projects

DietrichGebert/ponytail: Making AI Agents Think Like Lazy Senior Devs

The best code is no code. DietrichGebert/ponytail gained 743 stars today , totaling 149182 stars. The project advocates having AI agents prioritize existing solutions and reuse off-the-shelf tools rather than generating code from scratch every time. Ideal for development teams looking to reduce wheel reinvention and increase code reuse.

harry0703/MoneyPrinterTurbo: One-Click HD Short Video Generation

From topic to finished product. harry0703/MoneyPrinterTurbo uses AI large models and automated workflows to generate HD short videos 🎬 based on topics or keywords. Gained 431 stars today, totaling 127557 stars. Perfect for content creators who need to rapidly produce video materials, but check generated content for copyright and accuracy.

openclaw/openclaw: AI That Actually Does Things

Any operating system, the lobster way. openclaw/openclaw supports cross-platform operations 🦞, gaining 136 stars today, totaling 390989 stars. The project emphasizes AI’s ability to execute real tasks across multiple systems. Ideal for developers needing cross-platform operation automation.

◉ Social Media Picks

Fermion Open-Sources Phonon-2 Voice Model for Ultra-Fast Transcription

1 hour of audio in just 20 seconds. Gorden Sun explains that Fermion’s open-source Phonon-2 is only 164MB 🔊 and transcribes 1 hour of audio in about 20 seconds on an ordinary thin-and-light laptop. Through extremely low-bitwidth weight compression to reduce size, accuracy matches the original model that’s 15 times larger, with superior performance in meeting and presentation scenarios. English only. Perfect for users needing local, low-cost, high-precision voice transcription.

Xiang Yang Qiao Mu Develops Music Radio Plugin for DeepSeek Harness

Beyond RSS, now with radio. Xiang Yang Qiao Mu announced that he developed a music radio plugin for the DeepSeek Harness ecosystem 📻. Installation is simple — just tell DSH: 帮我安装插件:github.com/joeseesun/qiao…. Previously, he also launched the Qiaomu AI RSS plugin, providing overseas AI news, podcasts, and new tool introductions with English and Chinese rewrite support. Perfect for developers needing to extend functionality in the DSH environment.

😄 AI Fun Facts

Rainy Day Driver Makes Grok the General Contractor

On a boring rainy drive, Tesla owner Mike P gave Grok a set of requirements and had it command the bot on his Mac mini at home: create a repo, find materials, write scripts, then have Cursor call Claude Opus 5.5 to edit a video and toss it into Google Drive. In this viral X video , he went for coffee after speaking — five nodes across three platforms ran themselves to completion. Agents aren’t products; they’re the second type of user computers have grown.

❓ Related Questions

After Gemini 4 Argon’s Release, Which Subscription Tier Provides Access to the New Model?

According to Google’s official blog , Gemini 4 Argon has been released and demonstrated performance improvements. However, the announcement does not specify availability and access conditions for the new model across paid tiers (such as Gemini Advanced, API quotas, etc.). Before purchasing, confirm with the official team whether your target tier supports Argon model access.

When comparing currently available account, subscription, or quota services, check the product catalog at Aivora AI Account Store ; for support of features mentioned in the news, refer to official product descriptions and product pages.