今天的核心信号是 **AI 产业正在从"能力竞赛"转向"信任与安全竞赛"**。Bun 用 Claude Code 重写百万行代码后引发的安全性质疑、微软软件包二次被凭证窃取器污染、以及 Meta 被迫删除面部识别代码，都在提醒行业：AI 的规模化落地必须建立在可信任的基础之上。与此同时，OpenAI 和 Anthropic 的 IPO 竞赛与 Apple WWDC 的 AI 全面升级表明，资本和产品两端的竞争都在加速——但谁能在安全与速度之间找到平衡，谁才能赢得下一个阶段。
AllNewsPapersProjects★ Top picks (4+)
📰 Industry News
OpenAI Files Confidentially for IPO, Following Anthropic
OpenAI has confidentially submitted an S-1 filing with the SEC, just one week after its main rival Anthropic filed to go public, escalating the IPO race between the two AI giants.
Bun Team Rewrites 1M Lines of Code with Claude Code in 9 Days, Sparking Security Concerns
The Bun project completed its largest-ever AI refactoring with 6,755 commits, but the developer community is debating whether a 99.8% test pass rate truly represents code safety.
Apple WWDC 2026 Unveils New Siri AI and Apple Intelligence Upgrades
Apple introduced an AI-powered new Siri at WWDC with more natural conversation, personalized context understanding, and announced free cloud AI API access for small developers.
Google NotebookLM Upgraded to Gemini 3.5 with Cloud Computer and Source Retrieval
NotebookLM gets a full upgrade to the Gemini 3.5 model with more accurate responses, plus new cloud computing capabilities and source-finding features.
Microsoft AI Chief Says Superintelligence Is Near, But Won't Take Your Job
Mustafa Suleyman stated in an interview that superintelligence is imminent but human jobs won't be fully replaced, emphasizing a future of human-AI collaboration.
Ant Group Launches Overseas AI Payment Solution for Global Agent Operations
The solution helps users and merchants assess agent trustworthiness, paving the way for commercialized AI agent payments.
Amap Releases ABot-Earth0.5: 3D-Native High-Consistency Scene Generation Model
The model moves beyond 2D distillation to native 3D for high-consistency scene generation, now open for beta testing.
UK Government Invests $1 Billion in AI Supercomputer to Reduce US Tech Dependence
The UK believes state-backed computing infrastructure will boost domestic chip startups and reduce reliance on American technology.
Meta Removes Face-Recognition System from Its Smart Glasses App
Following a WIRED report, Meta deleted facial recognition code from the latest Meta AI app without explaining why or whether it will return.
Microsoft Packages Laced with Credential Stealer for Second Time in Weeks
73 malicious packages run self-replicating stealers as soon as opened by an AI agent, exposing AI supply chain security risks.
Amazon Launches AI-Generated Custom Merch Feature
Amazon expands print-on-demand services, letting users generate AI designs via Alexa text prompts for T-shirts, water bottles, and other products.
Former ByteDance Seed Team Member Gu Quanquan Responds to Departure Controversy
Gu denied claims about not being a team lead, as multiple members from the department recently left to start ventures.
📄 Papers
BloomBench: First Cognitively-Inspired Bilingual Multimodal Benchmark
This benchmark rigorously diagnoses real reasoning abilities of VLMs from a human cognitive perspective, filling gaps in fragmented existing evaluations.
Proposes a non-parametric learning algorithm that compares past reasoning traces to generate improvement signals, boosting reasoning without requiring large training samples.
Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation
Studies post-hoc compression of long chain-of-thought traces from reasoning models, reducing training tokens to 8.6-21.0% of original while maintaining performance.
Robots Need More than VLA and World Models
This position paper argues the bottleneck in robot intelligence is not just policy learning but the lack of mechanisms to convert abundant unstructured behavioral data into grounded robot supervision.
Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models
Proposes IPT as intermediate perceptual representations to help VLMs infer spatial information not directly observable.
UnpredictaBench: Benchmark for Evaluating Distributional Randomness in LLMs
Tests LLMs' ability to capture true underlying distributions, finding many models collapse to single answers and fail to simulate real system uncertainty.
Augmenting Attention with Exponentially Decaying Memory Improves Query-Aware KV Sparsity
LayerRoute: Input-Conditioned Adaptive Layer Skipping via LoRA Fine-Tuning
Learns to selectively skip Transformer blocks per input, optimizing for heterogeneity between structured tool calls and open-ended reasoning in agentic tasks.
🔧 Open Source
mvanhorn/last30days-skill
AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web, then synthesizes a grounded summary.
colbymchenry/codegraph
Pre-indexed code knowledge graph supporting Claude Code, Codex, Gemini, Cursor, and other major AI coding tools, reducing token consumption and tool calls.
Compresses tool outputs, logs, files, and RAG chunks before reaching the LLM, reducing tokens by 60-95% with same answers. Available as library, proxy, and MCP server.
XingYu-Zhong/DeepSeek-GUI
AI agent workspace for DeepSeek models with built-in Code and Claw modes, embeddable into applications.
Panniantong/Agent-Reach
Gives AI agents "eyes" to browse the entire internet, supporting Twitter, Reddit, YouTube, GitHub, Bilibili, Xiaohongshu, and more — one CLI, zero API fees.
alchaincyf/huashu-design
HTML-native design skill for Claude Code supporting high-fidelity prototypes, slides, and animations with 20 design philosophies and a 5-dimension review system.
openai/plugins
OpenAI's official plugin repository, providing third-party extension capabilities for ChatGPT and other products.
RyanCodrai/turbovec
Vector index built on TurboQuant, written in Rust with Python bindings, pursuing high-performance vector retrieval.
Leonxlnx/taste-skill
A "taste skill" frontend tool designed to stop AI from generating boring, generic "slop," giving AI good aesthetic judgment.
No items match this filter.
💡 Today's Take
The core signal today is that the AI industry is shifting from a "capability race" to a "trust and safety race." Bun's million-line Claude Code rewrite sparking security concerns, Microsoft packages being infected with credential stealers for the second time, and Meta being forced to remove facial recognition code all remind the industry that scaled AI deployment must be built on trust. Meanwhile, the OpenAI and Anthropic IPO race, alongside Apple WWDC's full AI upgrade, shows acceleration on both capital and product fronts — but whoever finds the balance between speed and safety will win the next phase.