今天最核心的信号是 Anthropic 发布了首个面向公众的 Mythos 级模型 Claude Fable 5，标志着顶级模型能力开始向开发者开放，同时其安全设计也引发了关于 AI 意识讨论的行业争议。与此同时，OpenAI 秘密提交 IPO，AI 头部公司的资本化进程正在加速，产业格局面临重塑。在开源社区，代码知识图谱（codegraph）和跨平台信息获取工具（Agent-Reach）的涌现，表明开发者正在系统性地解决 AI Agent 在复杂任务中的上下文和感知瓶颈，这将是 Agent 走向实用的关键一步。
AllNewsPapersProjects★ Top picks (4+)
📰 Industry News
Anthropic Releases Claude Fable 5, First Mythos-Class Model for the Public
Anthropic officially launched Claude Fable 5, its first Mythos-class model available to the public, excelling in software engineering, knowledge work, and vision, with guardrails for high-risk domains.
Microsoft AI Chief Calls Out Anthropic for Acting Like Claude Is Conscious
Microsoft AI CEO Mustafa Suleyman publicly criticized Anthropic for speculating about Claude's consciousness in its "constitution," calling it "really, really dangerous."
OpenAI Confidentially Files for IPO
Following Anthropic and SpaceX, OpenAI has also confidentially filed IPO paperwork, preparing to go public.
Apple WWDC 2026 Embraces AI, Major Siri Overhaul
Apple announced its AI strategy centered on "Siri AI" at WWDC 2026, including new photo editing tools and Safari extension generation, though mostly seen as catching up.
Microsoft AI Chief Walks Back Comments About AI Taking Over White-Collar Work
Mustafa Suleyman clarified his earlier remarks, stating AI will augment rather than fully replace lawyers, accountants, and other white-collar workers.
UK Invests in Billion-Dollar AI Supercomputer to Kick Addiction to US Tech
The UK government plans a state-backed infrastructure initiative to boost homegrown chip startups and build an AI supercomputer.
Meta Deletes Face-Recognition System From Its Smart Glasses App
Following a WIRED report, Meta removed face-recognition code from its smart glasses companion app, without stating why or if it will return.
Tencent Launches Full-Stack Agent Platform to Unify Enterprise AI Entry
Tencent announced a new strategy at its AI industry conference, aiming to connect full-stack agents through a single entry point.
Ant Group Launches Overseas AI Payment Solution
The solution helps users and merchants assess the trustworthiness of AI agents, enabling global agent operations.
📄 Papers
Agents' Last Exam (ALE): A New Benchmark for Evaluating AI Agents on Economically Valuable Tasks
This benchmark measures AI agent performance on long-horizon, real-world tasks with economic value, bridging the gap between existing benchmarks and real-world deployment.
AsyncWebRL: Efficient Multi-Step RL for Visual Web Agents
Improves training efficiency through asynchronous design, everlasting rollout pools, and lightweight screenshot handling.
DEI: Diversity in Evolutionary Inference for Quality-Diversity Search
Uses diverse LLMs as mutation operators in a distributed search framework for efficient exploration.
Pruning and Distilling Mixture-of-Experts into Dense Language Models
The first systematic framework to convert a trained MoE model into a standard dense architecture, addressing memory-constrained deployment.
SigmaScale: LLM Compression with SVD-based Low-Rank Decomposition and Learned Scaling Matrices
Improves truncated SVD compression performance by learning auxiliary scaling matrices.
Robotic Policy Adaptation via Weight-Space Meta-Learning
Proposes the WIZARD framework for zero-shot generalization of robot policies through weight-space meta-learning.
Liberating LLM Capabilities in Full-Duplex Speech Models
Proposes enabling speech LLMs to output text and other non-speech results, unlocking capabilities in code generation and structured analysis.
🔧 Open Source
router-for-me/CLIProxyAPI
Wraps Gemini CLI, ChatGPT Codex, Claude Code, etc., into OpenAI/Gemini/Claude-compatible API services, enabling free model access via API.
alibaba/open-code-review
Alibaba's open-source hybrid architecture code review tool, combining deterministic pipelines and an LLM Agent with built-in rulesets for NPE, XSS, etc.
colbymchenry/codegraph
A pre-indexed code knowledge graph for Claude Code, Codex, Cursor, etc., reducing token consumption and tool calls.
mvanhorn/last30days-skill
An AI agent skill that researches any topic across Reddit, X, YouTube, HN, and more, then synthesizes a grounded summary.
Panniantong/Agent-Reach
Gives AI agents "eyes" to read and search Twitter, Reddit, YouTube, Bilibili, and more via CLI, with zero API fees.
chopratejas/headroom
Compresses tool outputs, logs, and files before they reach the LLM, reducing token consumption by 60-95% without affecting answer quality.
Lum1104/Understand-Anything
Transforms any codebase into an interactive knowledge graph for exploration, search, and Q&A, compatible with major AI coding tools.
No items match this filter.
💡 Today's Take
The strongest signal today is Anthropic releasing Claude Fable 5, its first public Mythos-class model, opening frontier capabilities to developers while also sparking industry debate on AI consciousness. Simultaneously, OpenAI's confidential IPO filing signals accelerating capitalization among top AI firms, poised to reshape the industry landscape. In the open-source community, the emergence of tools like code knowledge graphs (codegraph) and cross-platform information agents (Agent-Reach) shows developers are systematically addressing context and perception bottlenecks for AI agents in complex tasks—a critical step toward practical agent deployment.