今天最值得关注的信号是 **AI 芯片人才争夺战升级**——OpenAI 芯片元老在量产前夜转投 Anthropic，同时英伟达 RTX Spark 正式将 AI 硬件推向 Windows PC，硬件与人才的双重竞争正在重塑 AI 基础设施格局。在安全方面，OpenAI 的 Lockdown Mode 和 SABER 基准的发布表明，行业正在从「模型能力竞赛」转向「Agent 安全与可控性」的深水区。此外，开源社区涌现出一批以「AI 技能包」为核心的项目（gsap-skills、last30days-skill、taste-skill），预示着 AI Agent 生态正在从通用能力向「专业工具化」演进，开发者可以通过标准化技能模块为 Agent 注入特定领域能力，这可能是下一波 Agent 应用爆发的关键基础设施。
AllNewsPapersProjects★ Top picks (4+)
📰 Industry News
Breaking] OpenAI Chip Lead Defects to Anthropic Just Before Mass Production
The key engineer who built OpenAI's first chip from scratch jumps to Anthropic, shaking the industry.
Breaking] Nvidia Launches RTX Spark: AI Hardware Officially Comes to Windows PCs
At Computex 2026, Nvidia unveils the Windows version of its Blackwell GB10 superchip, with Microsoft launching Surface Laptop Ultra and Dev Box simultaneously.
OpenAI Unveils "Lockdown Mode" to Defend Against Prompt Injection Attacks
New feature aims to protect sensitive data and reduce information leakage risks from prompt injections.
OpenAI Still Building a "Super App," Internal Says "Chat is Dead"
A senior OpenAI employee reveals the company is developing a super app beyond traditional chat interfaces.
Trump Administration May Take Equity Stake in OpenAI
President Trump says he's discussing deals where "the American people can benefit from the success of AI."
White House AI Advisor Sriram Krishnan Departs to Found New Institution
Krishnan reportedly plans to continue shaping Trump's AI policy through a new organization.
Meta AI App Starts Generating Clickbait Articles with AI
Meta AI app adds a "For You" section where topics, images, and text are all AI-generated with questionable quality.
Notion Restores Access to Anthropic After Service Disruption
Notion's head of product was "astonished" at "the amount of people RT-ing this."
Open-Source Chinese Video Framework Generates 5-Minute AI Long Videos Without Failure
Achieves high consistency, low latency, and real-time super-resolution, reaching global first tier.
PixelBloom Completes Series C Funding, Goes All-In on AI Office Agents
StarLink Capital portfolio company secures new funding focused on AI office solutions.
📄 Papers
SABER: Benchmarking Operational Safety of LLM Coding Agents in Stateful Project Workspaces
New benchmark evaluates coding agent safety after action sequences in realistic project environments, going beyond simple refusal of unsafe prompts.
Code2LoRA: Hypernetwork-Generated Adapters for Code Language Models under Software Evolution
Generates repository-specific LoRA adapters via hypernetwork, injecting repo knowledge with zero inference token overhead, handling codebase evolution.
BRepCLIP: Contrastive Multimodal Pretraining on BRep Primitives for CAD Understanding
First framework to align boundary representation geometry with language and image embeddings, filling CAD representation learning gap.
AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents
Inserts inference step between scene perception and tool use to generate IntentFrames, controlling probe budget and tool selection.
ForeSci: Evaluating LLM Agents for Forward-Looking AI Research Judgment
Evaluates whether agents can make forward-looking research decisions from historical evidence, with 500 tasks across four rapidly evolving AI domains.
Benchmark Everything Everywhere All at Once
Proposes Benchmark Agent for fully autonomous benchmark construction and updating, addressing sustainability and performance saturation issues.
AffordanceVLA: A Vision-Language-Action Model Empowering Action Generation through Affordance-Aware Understanding
Introduces structured affordance forecasting as task-oriented intermediate representation, bridging VLM semantic spaces and embodied control policies.
Learning Geometric Representations from Videos for Spatial Intelligent MLLMs
Learns geometric representations using only 2D video sequences, unlocking spatial intelligence in multimodal LLMs.
The Shape of Addition: Geometric Structures of Arithmetic in LLMs
Discovers geometric structure in addition operations, proposes Noisy Quantization Model explaining LLM arithmetic errors as "Geometric Slippage."
LLM Anonymization Against Agentic Re-Identification
Studies threat model of Agentic LLMs re-identifying via web search, exploring balance between anonymization and utility retention.
AI-powered job application framework built on Claude Code, automatically evaluating jobs, tailoring CVs, writing cover letters, and preparing for interviews.
Gives AI "taste," preventing generation of boring, generic "slop."
graphify](https://github.com/safishamsi/graphify)
Converts any folder of code, SQL schemas, scripts, docs, papers, images, or videos into a queryable knowledge graph.
No items match this filter.
💡 Today's Take
The most notable signal today is the **escalation of the AI chip talent war**—OpenAI's chip lead defects to Anthropic just before mass production, while Nvidia's RTX Spark officially brings AI hardware to Windows PCs, with hardware and talent competition reshaping the AI infrastructure landscape. On the security front, the release of OpenAI's Lockdown Mode and the SABER benchmark signals the industry shifting from "model capability competition" to the deep waters of "Agent safety and controllability." Additionally, the open-source community is seeing a surge of projects centered on "AI skill packs" (gsap-skills, last30days-skill, taste-skill), indicating the AI Agent ecosystem is evolving from general capabilities toward "professional tooling"—developers can inject domain-specific abilities into Agents through standardized skill modules, which could be the key infrastructure for the next wave of Agent application breakthroughs.