今日最核心的信号是 **Claude Fable 5 的反蒸馏事件**。Anthropic 在为 Fable 5 设置隐形防护措施时，不仅误伤了正常用户，更暴露了模型提供商与下游开发者之间的深层信任危机——当模型可以"暗中降智"时，依赖 API 构建产品的开发者将面临不可预测的行为风险。与此同时，**Agent 生态正在快速成熟**：从 GitHub 上涌现的 agent-skills、codegraph 等项目，到阿里高考志愿填报 Agent 的大规模落地，Agent 正从实验性概念走向生产级应用。另一个值得关注的信号是 **3D AI Agent 的诞生**，Meshy 的产品可能标志着 3D 内容创作即将进入类似 2D 图像生成领域的"ChatGPT 时刻"，这对游戏、影视、XR 等行业的影响值得持续跟踪。
AllNewsPapersProjects★ Top picks (4+)
📰 Industry News
Claude Fable 5's Anti-Distillation Mechanism Sparks Controversy, Anthropic Apologizes and Backs Down
Anthropic stealthily deployed anti-distillation guardrails on Claude Fable 5 that covertly throttled the model when detecting attempts to distill it, with a high false-positive rate, sparking strong backlash from researchers and competitors. Anthropic has publicly apologized and promised to roll back the policy, committing to greater transparency about restrictions going forward.
Claude Fable 5 Refuses to Answer Basic Biology Questions
Anthropic's Claude Fable 5, touted as powerful in biology and other domains, was found to refuse answering high-school-level basic biology questions, routing such queries to its predecessor flagship model Opus instead.
Microsoft Restricts Employee Use of Claude Fable 5 Over Data Retention Concerns
Microsoft has restricted internal employee use of Claude Fable 5 due to concerns over Anthropic's new data retention requirements. However, Microsoft quickly integrated Claude Fable 5 into GitHub Copilot and Foundry for customer-facing use.
xAI Fired Engineer Who Raised Grok Safety Alarms, Lawsuit Claims
A former xAI engineer is suing the company and SpaceX, alleging he was fired for raising AI safety concerns about Grok just before SpaceX's IPO.
Google Quietly Releases New Model with 4x Faster Inference
In the shadow of Mythos model launches, Google quietly released a new model using diffusion models for text generation, achieving 4x faster inference.
Xiaomi achieved 1T-parameter model inference on general-purpose GPUs with over 1000 tokens per second throughput, supporting 7-second Vibe Coding delivery.
Deezer Launches Cross-Platform AI Music Detection Tool
Deezer released a tool that scans playlists from Spotify, Apple Music, and other streaming platforms to identify AI-generated music.
Meshy Releases World's First 3D AI Agent
A milestone moment for 3D creation as Meshy launches the world's first 3D AI Agent, potentially lowering the barrier to 3D creation like ChatGPT did for text.
Alibaba Launches Free AI College Application Agent
Alibaba released a free AI agent for 12.9 million Chinese college entrance exam students to help with application decisions, stress-tested with 400,000 AI "students" beforehand.
AI Short Drama Tool Sector Gets Year's Largest Single Financing
The AI short drama creation tool space secured the year's largest single financing round, with capital continuing to bet on AI video generation for short-form drama applications.
"AI-Pilled" Firms Spend $7,500 Per Employee Monthly on AI
According to the Ramp AI Index, the most AI-obsessed companies spend an average of $7,500 per employee per month on AI tools and services.
Anthropic CEO Dario Amodei Has Just One Direct Report
Anthropic CEO Dario Amodei runs an extremely flat management structure with only one direct report, highly unusual for a fast-growing large AI company.
📄 Papers
Timing Trick Cuts LLM Training Energy by Up to 14%
Researchers at the University of Twente found that clever timing adjustments can reduce LLM training energy consumption by up to 14% without sacrificing model performance.
Can Generalist Agents Automate Data Curation?
The paper proposes Curation-Bench, a benchmark testing whether generalist coding agents can automate AI training data curation, including data inspection, policy implementation, evaluation, and iterative revision.
Auditing Invisible Dependencies in Modern LLMs
The paper introduces ModSleuth, a framework for tracing and auditing the recursive dependencies in modern LLM training pipelines, including upstream model-generated data, filtered corpora, judged outputs, etc.
ReVision: Scaling Computer-Use Agents via Temporal Visual Redundancy Reduction
The paper proposes ReVision, which reduces temporal redundancy in visual observations during computer-use agent interactions, significantly cutting token costs and enabling long history contexts.
SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference
The paper proposes SparDA, introducing a fourth projection (Forecast) to achieve sparse attention, solving KV cache and selection step computational bottlenecks in long-context inference.
DRIFT: Continuous Output Decoding Framework for Vision-Language Models
The paper proposes DRIFT, using residual flow adapters to enable pretrained VLMs to decode continuous outputs for tasks like temporal localization and robot control requiring precise continuous values.
Grammar-Constrained Decoding Can Be Exploited to Generate Malicious Code
The paper reveals a novel jailbreak attack called CodeSpear that exploits Grammar-Constrained Decoding (GCD) to induce LLMs to generate malicious code, showing reliability techniques themselves can become attack surfaces.
LLMs Are Overconfident in Their Own Responses
Research finds that instruction-tuned LLMs have worse calibration, and chat templates further exacerbate this overconfidence, causing models to be overly confident in their incorrect answers.
Comparison of Subquadratic Architectures: xLSTM, Mamba-2, and Gated DeltaNet
The paper systematically compares three leading subquadratic architectures on code model pre-training, knowledge distillation, and time-series pre-training, providing architecture selection guidance.
⭐22: A pre-indexed code knowledge graph compatible with Claude Code, Codex, Gemini, and other major agent tools, reducing token consumption and tool call counts.
⭐14: Gives AI agents "eyes" to read the entire internet, supporting Twitter, Reddit, YouTube, GitHub, Bilibili, Xiaohongshu, etc., single CLI with zero API fees.
⭐37: Compresses tool outputs, logs, files, and RAG chunks before they reach the LLM, reducing tokens by 60-95% with the same answers. Supports library, proxy, and MCP server modes.
⭐17: AI generates editable PowerPoint from any document, supporting native shapes & animations, speaker notes as audio narration, and custom templates.
No items match this filter.
💡 Today's Take
The strongest signal today is the **Claude Fable 5 anti-distillation incident**. By deploying invisible guardrails that could "stealthily throttle" the model, Anthropic not only caught normal users in the crossfire but exposed a deep trust crisis between model providers and downstream developers—when a model can secretly degrade its performance, developers building products on APIs face unpredictable behavioral risks. Meanwhile, **the Agent ecosystem is rapidly maturing**: from the surge of agent-skills and codegraph projects on GitHub to Alibaba's large-scale college application agent deployment, agents are moving from experimental concepts to production-grade applications. Another signal worth watching is the **birth of the 3D AI Agent**—Meshy's product may mark the "ChatGPT moment" for 3D content creation, similar to what happened in 2D image generation, with implications for gaming, film, XR, and beyond that merit continued attention.