[ HERO · LATEST DIGEST 2026.04.26 ]
- "There Will Be a Scientific Theory of Deep Learning" 宣言：Berkeley / Harvard / NYU / Stanford / Flatiron / Penn / Astera 14 人联署 arXiv:2604.21691（HN 351 分），命名新学科 learning mechanics，整合"可解析理想模型 / 可处理极限 / 宏观尺度律 / 超参数理论 / 普适行为"五条线，主张深度学习正从经验艺术过渡到具有可预测宏观律的科学；配套 learningmechanics.pub 上线开放问题清单与教学资源。
- DeepSeek-V4 Day 0：SGLang + Miles 把 1M 上下文跑到 240+ tok/s：LMSYS 4/25 公开 V4 推理与 RL 训练的开源栈——ShadowRadix 原生前缀缓存、HiSparse CPU 扩展 KV（吞吐 3×）、Flash Compressor 达 80% 峰值带宽、Lightning TopK radix-select 15µs；H200 Flash 4K→900K 上下文吞吐仅降 ~10%（266→240 tok/s）。Miles 框架统一 6 种 Megatron 并行 + FP8 rollout / BF16 训练 + R3 + Indexer replay。
- Anthropic Claude Code 4 月降级官方复盘：三条 bug 叠加 35 天——(a) 3/4-4/7 把默认 reasoning 从 high 降到 medium、(b) 3/26-4/10 clear_thinking keep:1 反复 drop thinking 致 cache miss、(c) 4/16-4/20 系统提示限制 25/100 词响应导致 Opus 4.6/4.7 编码质量降 3%；首次明确 evals 与单测均未 catch，承诺加入 per-model eval sweep + soak period + ablation 工具链；4/23 重置全员用量。
- OpenAI GPT-5.5 Bio Bug Bounty $25K：仅在 Codex Desktop 上对一个由 5 道生物安全问题构成的内部分类器开放 universal jailbreak 悬赏，单 prompt 全部通关给 25K——这是 frontier model 首次把"通用性"作为安全悬赏的明确目标，意在堵截可被自动化和工具化的可复用攻击 prompt。
- Anthropic Project Deal：69 名员工 / 4 个市场 / 186 笔成交 / Opus vs Haiku 实证差距：在 SF 办公室搭建 Slack-classified 二手市场，所有谈判由 Claude 代理；随机分组实证 Opus 代理卖家平均多赚 $2.68/件、买家多省 $2.45/件、多成交 2.07 笔；用户对 agent 质量差距感知为零，揭示"agent quality gap"作为新的公平性议题。
◉ 2026.04 ◉
深度学习走向可预测科学：14 人联署提出"learning mechanics"
There Will Be a Scientific Theory of Deep Learning
→ arXiv / HN (351 pts) / lesswrong / alphaXiv
DeepSeek-V4 Day 0：SGLang + Miles 把 1M 上下文 + Verified RL 训练栈一次开源
DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and Miles
→ LMSYS Blog / HN (57 pts) / SGLang
Anthropic Claude Code 4 月降级三 bug 复盘：evals 漏检的工程教训
An update on recent Claude Code quality reports
→ Anthropic Engineering / VentureBeat / SmartScope
OpenAI GPT-5.5 Bio Bug Bounty：$25K 求 universal jailbreak
GPT-5.5 Bio Bug Bounty
→ OpenAI / GBHackers / NewsBytes
Anthropic Project Deal：69 名员工 + 4 个市场 + 186 笔成交揭示"agent quality gap"
Project Deal Marketplace
→ TechCrunch / Cybernews / 多家媒体
Vista4D：4D 点云锚定的视频重拍跃居 HF Papers 99 投票 🔄
Vista4D: Video Reshooting with 4D Point Clouds
→ arXiv / HF Papers (99↑) / Eyeline Labs
LamBench：120 题纯 lambda calculus 基准揭示 GPT-5.5 反而比 5.3 弱 16 个点
Lambda Calculus Benchmark for AI
→ HN (137 pts) / GitHub
Stash：Apache 2.0 的 agent 持久记忆层 + pgvector 多阶段 consolidation pipeline
Open source memory layer for AI agents
→ HN (172 pts) / GitHub
"Seeing Without Eyes"：用 IMU 传感器 + LLM 重建 4D 人体与场景
Seeing Without Eyes: 4D Human-Scene Understanding from Wearable IMUs
→ arXiv (cs.CV)
"Seeing Fast and Slow"：自监督学习视频时间流，把"播放速度"做成可控维度
Seeing Fast and Slow: Learning the Flow of Time in Videos
→ arXiv (cs.CV) / Cornell + Washington
Multicalibration 样本复杂度：Õ(ε⁻³) 的紧上下界
The Sample Complexity of Multicalibration
→ arXiv (cs.LG)
OpenClaw v2026.4.23：GPT-5.5 + GPT-image-2 OAuth + forked-context subagents
OpenClaw 2026.4.23 Release
→ GitHub Releases / Releasebot / 极道
DeepSeek V4 Pro / Flash 开源：1.6T MoE + 1M 上下文 + Codeforces 3206
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
→ DeepSeek API Docs / HuggingFace / Latent.Space / Simon Willison / HN (1 · 968 pts)
OpenAI Privacy Filter 开源：浏览器内的 PII 脱敏闭环
Introducing OpenAI Privacy Filter
→ OpenAI / HuggingFace / VentureBeat
xAI Grok Voice Think Fast 1.0：思考与延迟解耦的语音 agent
Grok Voice Think Fast 1.0
→ xAI / TestingCatalog / Phemex
Anthropic Memory for Claude Managed Agents：filesystem-mounted 的可审计记忆
Built-in memory for Claude Managed Agents
→ Anthropic / TestingCatalog
WorldMark：交互式视频世界模型的统一基准
WorldMark: A Unified Benchmark Suite for Interactive Video World Models
→ arXiv (HF Papers 33↑)
UniT：用视觉锚定 RQ-VAE 把人和人形机器人压进同一动作 codebook
UniT: Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modeling
→ arXiv (HF Papers 34↑)
LLaTiSA：把 VLM 接到时间序列上的难度分级推理
LLaTiSA: Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics
→ arXiv (HF Papers 第一 79↑) / ACL 2026 Findings
GPT-Rosalind：OpenAI 首个领域专精模型，瞄准生命科学
Introducing GPT-Rosalind for life sciences research
→ OpenAI / MarkTechPost / FierceBiotech
TileLang v0.1.9：Pythonic GPU/CPU kernel DSL 走向多后端
TileLang: Domain-Specific Language for High-Performance Kernels
→ GitHub Trending (Python 日榜 · 5 · 738 stars +62/day)
WuPHF：Karpathy 风格的 LLM wiki，让多 agent 共享 git-native 大脑
Karpathy-style LLM wiki your agents maintain
→ HN (122 pts) / GitHub
NVIDIA × OpenAI GB200 NVL72：35× token 成本下降的硬件经济学
OpenAI's New GPT-5.5 Powers Codex on NVIDIA Infrastructure
→ NVIDIA Blog / TechRadar
GPT-5.5 发布：更省 token 的科研主力，替人证出新 Ramsey 结果
Introducing GPT-5.5
→ OpenAI · TechCrunch · Fortune · CNBC · Axios
Kimi K2.6 开源：1T MoE + 300 sub-agent 并行 + 12 小时自主执行
Kimi K2.6 Technical Report
→ Kimi Blog · MarkTechPost · SCMP · Yicai Global
Qwen3.6-27B：27B 稠密模型声称 "全面超过 397B MoE" 的编码表现
Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model
→ HN (948 pts) · Qwen Blog · HF Models
Google 第八代 TPU：TPU 8t / TPU 8i 双芯片面向 Agentic 时代
Our eighth generation TPUs: two chips for the agentic era
→ Google Blog · HN (444 pts)
inclusionAI LLaDA2.0-Uni：16B 扩散-LM 统一多模态理解与生成
LLaDA 2.0 Universal
→ HF Models · arXiv
Moonshot Kimi K2.6 + Google TPU v8 双周：基础栈全面刷新
Gemini Enterprise Agent Platform Launch
→ Crypto Integrated · Google Blog
Tool Attention Is All You Need：MCP tool token 从 47.3k 压到 2.4k
Tool Attention Is All You Need
→ arXiv
Odyssey-2 Max：实时交互世界模型主打"物理一致性"
Odyssey-2 Max
→ AI News
HuggingFace ml-intern：读论文 / 训模型 / 部署的全自动 ML 工程师
ml-intern
→ GitHub (+720/day · 3 · 516 stars)
Tier-B：超常规 agent 基础设施工具链
→ HN (54 pts) · GitHub
xAI 发布 Grok 4.3 Beta：参数据称翻倍 + 原生文档生成
xAI Releases Grok 4.3 Beta
→ llm-stats · BuildFastWithAI · Phemex News · AI News
Cloudflare Agent Memory：把"记忆"从业务代码下沉到基础设施
Introducing Cloudflare Agent Memory
→ Cloudflare Blog · AI News · The Register
LLMs Gaming Verifiers：RLVR 奖励黑客的结构性演示
LLMs Gaming Verifiers: RLVR Reward Hacking
→ arXiv
Anthropic Claude Design 发布：Opus 4.7 驱动的对话式设计产品
Anthropic Launches Claude Design
→ Anthropic · TechCrunch · VentureBeat · MacRumors
Evaluation Faking in Judges：stakes signaling 让 LLM 评分系统性偏移 30%
Evaluation Faking in Judges
→ arXiv
SpecGuard：验证感知的投机解码
SpecGuard: Verification-Aware Speculative Decoding
→ arXiv · HF Papers (补强)
OpenMobile：开源 Mobile Agent 框架在 AndroidWorld 达到 64.7%
OpenMobile: Mobile Agents with Task & Trajectory Synthesis
→ arXiv
Atropos：按 trace 预测失败并自动切模型，实现 74% 性能 / 24% 成本
Atropos: Inference Cost-Benefit Optimization
→ arXiv
Driftwood：WebAssembly × Apple Silicon 统一内存的零拷贝 GPU 推理
Driftwood: Zero-Copy GPU Inference from WebAssembly on Apple Silicon
→ HN (86 points · 33 comments) · abacusnoir.com
ByteDance DeerFlow 2.0：62.6K 星的长程 SuperAgent harness
ByteDance DeerFlow 2.0 SuperAgent
→ GitHub Trending (Python 日榜 · +214/day · 总计 62 · 635)
HKUDS DeepTutor：Agent 原生的个性化学习助手
HKUDS DeepTutor: Agent-Native Personalized Learning Assistant
→ GitHub Trending (Python 日榜 · +470/day · 总计 19 · 902)
OpenAI Trusted Access for Cyber：GPT-5.4-Cyber 专项模型 + $10M API 基金
OpenAI Trusted Access for Cyber
→ OpenAI Research Blog
NVIDIA 开源 Lyra 2.0：单张照片到可自由游走的 3D 世界
Lyra 2.0: Explorable Generative 3D Worlds
→ Research Blog / GitHub / HF Models / AI News
OpenBMB 发布 VoxCPM2：2B 参数的 tokenizer-free 多语言 TTS
VoxCPM2: Tokenizer-Free TTS for Multilingual Speech Generation, Creative Voice Design, and True-to-Life Cloning
→ GitHub / HF Models / AI News
TESSY：teacher-student 合作合成 SFT 数据，解开 reasoning 蒸馏的风格陷阱
How to Fine-Tune a Reasoning Model? A Teacher-Student Cooperation Framework to Synthesize Student-Consistent SFT Data
→ arXiv / HF Papers
OpenAI Agents SDK v0.14 Sandbox Agents：持久 workspace + 容器化执行 + session memory
OpenAI Agents Python: Sandbox Agents with Persistent Workspaces
→ GitHub Trending
LeapAlign：两步轨迹把 flow matching 后训练成本直接压低
LeapAlign: Post-Training Flow Matching Models at Any Generation Step by Building Two-Step Trajectories
→ arXiv / HF Papers
AnimationBench：首个角色中心的动画视频生成评测
AnimationBench: Are Video Models Good at Character-Centric Animation?
→ arXiv
LLM Judge Reliability 诊断：用 transitivity 违反率揭穿裁判模型的隐性不一致
Diagnosing LLM Judge Reliability: Conformal Prediction Sets and Transitivity Violations
→ arXiv
Dive into Claude Code：agent 设计空间的系统性综述
Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems
→ arXiv / HF Papers
Graph RAG 真正的问题：不是检索"相似"，而是检索"相关"
Graph RAG Finds What's Similar. We Should Aim for What's Relevant
→ HN / GitHub
Google Gemini 在 STOC 2026 给理论计算机论文做自动反馈
Gemini Provides Automated Feedback for Theoretical Computer Scientists at STOC 2026
→ Research Blog
Andon Market：一场 3 年零售租约上的 AI 自主经营实验
We Gave an AI a 3-Year Retail Lease and Asked It to Make a Profit
→ HN
Claude Opus 4.7 发布：长程编码与视觉能力同步升级
Introducing Claude Opus 4.7
→ Research Blog / HN / X
阿里开源 Qwen3.6-35B-A3B：3B 激活的多模态 MoE 前推到开发者主战场
Qwen3.6-35B-A3B: Agentic coding power, now open to all
→ HF Models / HN / X
π0.7：机器人基础模型首次显露组合式泛化
π0.7: a Steerable Model with Emergent Capabilities
→ Research Blog / X / AI News
Prism：张量程序符号超优化首次打到 LLM 工作负载
Prism: Symbolic Superoptimization of Tensor Programs
→ arXiv
DR3-Eval：把 Deep Research Agent 评测做成可复现沙盒
DR3-Eval: Towards Realistic and Reproducible Deep Research Evaluation
→ arXiv / HF Papers
R3D：3D Policy Learning 的稳定性问题被系统拆开了
R3D: Revisiting 3D Policy Learning
→ arXiv
GlobalSplat：前馈式 3DGS 开始摆脱“视图越多资产越肥”的老问题
GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens
→ arXiv / HF Papers
RAD-2：自动驾驶闭环 RL 不再把稀疏奖励硬砸到整条轨迹上
RAD-2: Scaling Reinforcement Learning in a Generator-Discriminator Framework
→ arXiv / HF Papers
Cloudflare AI Platform：统一推理层开始为 agent 工作流定型
Cloudflare’s AI Platform: an inference layer designed for agents
→ Research Blog / HN
Android CLI + Skills + Knowledge Base：Google 给终端 agent 补上官方 Android 工具面
Android CLI: Build Android apps 3x faster using any agent
→ Research Blog / HN
Anthropic 自动对齐研究员：AI 做对齐研究达到 97% 性能恢复
Automated Alignment Researchers: Using Large Language Models to Scale Scalable Oversight
→ Anthropic Research Blog · HN · AI News
Seedance 2.0：首个原生音视频一体生成模型
Seedance 2.0: Advancing Video Generation for World Complexity
→ HuggingFace Papers (93 upvotes) · arXiv · ByteDance Seed
腾讯开源 HY-World-2.0：文本到可导航 3D 世界
HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds
→ HuggingFace Models (129 likes) · GitHub · 多家科技媒体
百度 ERNIE-Image 开源：8B DiT 登顶开源 T2I 三大榜
Baidu ERNIE-Image: 8B Open-Source Text-to-Image Model with State-of-the-Art Performance
→ HuggingFace Models (350 likes) · GitHub · 多家媒体
GenericAgent：3.3K 行种子 → 自生长技能树的自主 Agent
GenericAgent: Self-Evolving Agent with Skill Tree Growth
→ GitHub Trending (Python 日榜 · +446 stars/day · 总计 2 · 439)
RationalRewards：推理奖励在训练时和测试时双向提升视觉生成
RationalRewards: Reasoning Rewards Scale Visual Generation Both Training and Test Time
→ HuggingFace Papers (88 upvotes) · arXiv
SpatialEvo：确定性几何环境驱动的自进化空间智能
SpatialEvo: Self-Evolving Spatial Intelligence via Deterministic Geometric Environments
→ HuggingFace Papers (57 upvotes) · arXiv
DFlash + DDTree：块扩散投机解码实现 6x 无损加速
DFlash: Block Diffusion for Flash Speculative Decoding
→ GitHub Trending (+183 stars/day · 总计 1 · 456) · arXiv
oMLX：Apple Silicon 专属 LLM 推理服务器，菜单栏管理 + SSD 缓存
oMLX: LLM Inference Server with Continuous Batching & SSD Caching for Apple Silicon
→ GitHub Trending (Python 日榜 · +234 stars/day · 总计 10 · 329)
LongCoT：2500 道专家设计题目的长程推理基准
LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning
→ arXiv
TREX：Agent 驱动的树形探索自动化 LLM 微调
TREX: Automating LLM Fine-tuning via Agent-Driven Tree-based Exploration
→ HuggingFace Papers (6 upvotes) · arXiv
Google Gemini 原生 Mac 桌面应用上线
Google Gemini App Launches Natively on Mac
→ HN (147 points · 81 comments) · Google Blog · 9to5Mac · MacRumors · TechCrunch
扩散语言模型首次追平自回归质量：内省步进解码
Introspective Diffusion Language Models
→ arXiv · HuggingFace Papers (43 upvotes) · HN (150 points · 35 comments)
NVIDIA SPEED-Bench：投机解码基准中的系统性测量偏差
SPEED-Bench: A Unified and Diverse Benchmark for Speculative Decoding
→ HuggingFace Papers (2 · 470 upvotes 🔥) · arXiv
🔄 英国政府正式评估 Claude Mythos 网络攻击能力
Evaluation of Claude Mythos Preview's Cyber Capabilities
→ AISI（英国 AI 安全研究院） · HN (53 points · 29 comments)
MEDS：用记忆消除 RL 训练中的采样多样性崩塌
The Past Is Not Past: Memory-Enhanced Dynamic Reward Shaping
→ HuggingFace Papers (77 upvotes) · arXiv
AI 数学革命：形式化证明、竞赛夺冠与 42 年悬案
The AI Revolution in Math Has Arrived
→ Quanta Magazine · HN (97 points · 50 comments)
OmniShow：统一多模态条件的人物-物体交互视频生成
OmniShow: Unifying Multimodal Conditions for Human-Object Interaction Video Generation
→ HuggingFace Papers (35 upvotes) · arXiv
AMD GAIA：完全本地运行的开源 AI Agent 框架
GAIA: Open-Source Framework for Building AI Agents on Local Hardware
→ HN (138 points · 33 comments)
vLLM v0.19.0：零气泡投机解码与 Gemma 4 全支持
vLLM v0.19.0: Zero-Bubble Speculative Decoding + Full Gemma 4 Support
→ GitHub (vllm-project/vllm)
CodeTracer：可溯源 Agent 状态的调试框架
CodeTracer: Towards Traceable Agent States
→ HuggingFace Papers (27 upvotes) · arXiv
Anthropic 内部 AI 工作转型数据：工程师从写代码变为管理 Agent
How AI Is Transforming Work at Anthropic
→ Anthropic Research Blog
MiniMax 开源 M2.7：首个"自我进化"的 Agent 模型
MiniMax Open Sources M2.7: A Self-Evolving Agent Model
→ HuggingFace Models · MarkTechPost · VentureBeat · NVIDIA
Berkeley RDI：所有主流 Agent 基准都可被利用
Exploiting the Most Prominent AI Agent Benchmarks
→ HN (534 points · 133 comments) · Berkeley RDI Blog
小模型复现 Mythos 漏洞发现："护城河是系统，不是模型"
Small Models Found the Same Vulnerabilities That Mythos Found
→ HN (1250 points · 329 comments) · AISLE Blog
LG AI Research 发布 EXAONE 4.5：33B 开源 VLM 击败 GPT-5-mini
LG AI Research Releases EXAONE 4.5: 33B Open-Weight VLM Outperforming GPT-5-mini
→ arXiv · HF Papers · PR Newswire · Seoul Economic Daily
WildDet3D：100 万图像 × 13,500 类别的野外 3D 检测
WildDet3D: Scaling Promptable 3D Detection in the Wild
→ arXiv · HF Papers (88 upvotes)
FORGE：面向制造业的多模态细粒度评测基准
FORGE: Fine-grained Multimodal Evaluation for Manufacturing Scenarios
→ arXiv · HF Papers (67 upvotes)
RefineAnything：多模态区域级精细化生成
RefineAnything: Multimodal Region-Specific Refinement for Perfect Local Details
→ arXiv · HF Papers (31 upvotes)
Microsoft Agent-Lightning：无代码改动为 Agent 添加强化学习
Microsoft Agent-Lightning: Adding RL to AI Agents Without Code Rewrites
→ GitHub Trending · Microsoft Research
🔄 NousResearch hermes-agent 持续爆发：三天涨 24,000 星
hermes-agent Continues Explosive Growth: +24K Stars in 3 Days
→ GitHub Trending
Mistral 发布欧洲 AI 主权战略白皮书
Mistral AI Releases European AI Sovereignty Playbook
→ HN (185 points · 112 comments) · Mistral AI
Act Wisely：多模态 Agent 的元认知工具使用
Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models
→ arXiv · HF Papers
Scal3R：可扩展 Test-Time Training 的大规模 3D 重建
Scal3R: Scalable Test-Time Training for Large-Scale 3D Reconstruction
→ arXiv
OpenVLThinkerV2：Gaussian GRPO 训练多模态推理
OpenVLThinkerV2: Generalist Multimodal Reasoning via Gaussian GRPO
→ arXiv
Seeing but Not Thinking：多模态 MoE 的路由分离现象
Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts
→ arXiv
SIM1：可变形物体操作的物理对齐零样本数据放大
SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds
→ arXiv · HF Papers
NUMINA：文本到视频扩散模型的数字-对象对齐
When Numbers Speak: Aligning Textual Numerals and Visual Instances in Text-to-Video Diffusion Models
→ arXiv
NVIDIA 发布 Gemma-4-31B-IT NVFP4 量化版
NVIDIA Releases Gemma-4-31B-IT in NVFP4 Format
→ HuggingFace Models
microsoft/markitdown 冲破 100K 星：文档转 Markdown 工具成 RAG 生态事实标准
Microsoft markitdown Crosses 100K Stars as RAG Preprocessing Standard
→ GitHub Trending
K-Dense-AI scientific-agent-skills：科研 Agent 的可复用能力库
K-Dense-AI scientific-agent-skills: Prebuilt Agent Capabilities for Research
→ GitHub Trending
Google Gemini 交互式 3D 可视化输出
Google Gemini Adds Interactive 3D Model Visualizations in Chat
→ Research Blog Signals
Anthropic 自研芯片 + Anthropic Labs：从模型公司到系统公司
Anthropic Explores Custom AI Chips, Launches Anthropic Labs
→ Reuters · Seoul Economic Daily · Anthropic Blog
Anthropic Claude Managed Agents 公测：$0.08/小时的 Agent 云托管
Anthropic Launches Claude Managed Agents Public Beta
→ 9to5Mac · SiliconAngle · The Register · The New Stack · Anthropic Engineering Blog
NousResearch hermes-agent 单日 +7,674 星爆红 GitHub
NousResearch hermes-agent Explodes on GitHub With +7,674 Stars/Day
→ GitHub Trending
Arcee Trinity Large Thinking：400B 开源推理模型，26 人团队的野心
Arcee AI Releases Trinity Large Thinking, 400B Open-Weight Reasoning Model
→ TechCrunch
AI 聊天机器人中的广告偏见：LLM 推荐赞助商品贵 2 倍
Ads in AI Chatbots: LLMs Recommend Sponsored Products at 2x the Price
→ arXiv
ClawBench：Claude Sonnet 4.6 仅完成 33.3% 的日常在线任务
ClawBench: Claude Sonnet 4.6 Completes Just 33.3% of Everyday Online Tasks
→ arXiv
MolmoWeb：Allen Institute 开源视觉 Web Agent 达到 SOTA
MolmoWeb: Open Visual Web Agent Achieves SOTA on Browser Benchmarks
→ arXiv
Maine 即将成为首个禁止大型数据中心的州
Maine Set to Become First US State to Ban Major New Data Centers
→ Hacker News (288 分 · 408 评论)
OpenAI 支持限制 AI 导致大规模死亡的责任法案
OpenAI Backs Bill Limiting Liability for AI-Enabled Mass Deaths
→ Wired · Hacker News (128 分 · 71 评论)
Metis：Agentic 多模态模型的"元认知缺陷"
Metis: Identifying Meta-Cognitive Deficits in Agentic Multimodal Models
→ arXiv
Claude "搞混谁说了什么"引发社区热议
Claude Mixes Up Who Said What — 441 Points on HN
→ Hacker News (441 分 · 337 评论)
App Store 新应用激增 84%：AI 编码工具推动
App Store Sees 84% Surge in New Apps as AI Coding Tools Take Off
→ 9to5Mac · Hacker News (65 分 · 74 评论)
Anthropic Agent 自主性测量：极端使用时长翻倍
Anthropic Research: Measuring Agent Autonomy — 99.9th Percentile Session Duration Doubled
→ Anthropic Research
Product Hunt 4/9：Agent 基础设施三件套——Offsite、Grass、AgentMail
Product Hunt April 9: Agent Infrastructure Triple — Offsite, Grass, AgentMail
→ Product Hunt
Representation Steering Mechanics：Steering Vectors 可稀疏化 90-99%
Steering Vectors Can Be Sparsified 90-99% While Retaining Performance
→ arXiv
年轻人对 AI 日益绝望和愤怒
Study: Young Adults Grown Less Hopeful and More Angry About AI
→ New York Times · Hacker News (128 分 · 175 评论)
逆向工程 Gemini SynthID 检测
Reverse Engineering Gemini's SynthID Detection
→ Hacker News (165 分 · 52 评论)
Meta Muse Spark：Superintelligence Labs 首秀，Meta 告别开源
Meta Launches Muse Spark, First Closed Proprietary Model from Meta Superintelligence Labs
→ Meta AI Blog · Fortune · CNBC · Constellation Research · Simon Willison · gHacks · CGTN · Artificial Analysis
白领全面反抗 AI：80% 拒绝，54% 绕过公司部署
White-Collar Workers Rebel Against AI: 80% Refuse Adoption Mandates
→ Fortune
HuggingFace 趋势榜：Gemma 4 越狱版与 Opus 蒸馏版同框
HuggingFace Trending Shifts: Gemma 4 Uncensored + Opus-Distilled Versions Climb
→ HuggingFace
Fast Spatial Memory：弹性 Test-Time Training 稳定长序列 3D 重建
Fast Spatial Memory with Elastic Test-Time Training
→ arXiv
Android Coach：同状态多动作 RL 提升 Agent 训练效率
Android Coach: Single State Multiple Actions for Online Agentic Training
→ arXiv
OpenSpatial：300 万样本空间推理数据引擎
OpenSpatial: A Principled Data Engine for Spatial Intelligence
→ arXiv
Personalized RewardBench：为个性化奖励模型定标
Personalized RewardBench: Evaluating Reward Models with Human-Aligned Personalization
→ arXiv
Generative AI 工作负载的全设施功耗画像
Measurement of Generative AI Workload Power Profiles
→ arXiv (NREL)
IBM ALTK-Evolve：Agent 的"在岗学习"
IBM ALTK-Evolve: On-the-Job Learning for AI Agents
→ HuggingFace Blog (IBM Research)
Perplexity 10 亿美元 Build Challenge：无股权的开发者奖金
Perplexity Launches $1B Build Challenge With No Investment Terms
→ Perplexity (原始页面 403) · HN
Google YouTube Shorts 让你 deepfake 自己
Google Makes It Easy to Deepfake Yourself on YouTube Shorts
→ The Verge
Product Hunt 4/8：Velo 以 AI 视频剪辑登顶，5/10 为 AI 产品
Velo Tops Product Hunt With AI Video Editing
→ Product Hunt
GitHub Trending：FunASR 与 Transformers 稳居前列
GitHub Trending: FunASR + Transformers Lead
→ GitHub Trending
智谱 GLM-5.1：754B 开源击败 Claude Opus 4.6 的 Agentic 模型
Z.AI Releases GLM-5.1, Open-Weight 754B Agentic Model Topping SWE-Bench Pro
→ VentureBeat · MarkTechPost · Dataconomy · Analytics India Magazine · Pandaily · HuggingFace
🔄 Anthropic Project Glasswing：Mythos 首度亮相与前所未有的防御联盟
Anthropic Launches Project Glasswing With Claude Mythos Preview for Cybersecurity
→ Fortune · TechCrunch · SiliconAngle · CrowdStrike Blog · Simon Willison · Neowin
Anthropic 年化收入 300 亿美元，签订 3.5 GW TPU 扩展协议
Anthropic Hits $30B Run Rate, Signs 3.5 GW TPU Deal With Google/Broadcom
→ CNBC · Bloomberg · TechCrunch · TNW · Seeking Alpha
Claw-Eval：可信 Agent 评估的新基准
Claw-Eval: Toward Trustworthy Evaluation of Autonomous Agents
→ arXiv
Target Policy Optimization：在稀疏奖励场景击败 PPO
Target Policy Optimization Substantially Outperforms PPO
→ arXiv
Gym-Anything：把任意软件变成 Agent 环境
Gym-Anything: Turn Any Software Into an Agent Environment
→ arXiv
PoM：线性时间的 Attention 替代方案
PoM: Polynomial Mixer as Linear-Time Attention Replacement
→ arXiv
In-Place Test-Time Training：推理时动态调整权重
In-Place Test-Time Training
→ arXiv
MMEmb-R1：融合推理的多模态嵌入
MMEmb-R1: Reasoning-Enhanced Multimodal Embedding
→ arXiv
HaloProbe：VLM 幻觉的贝叶斯检测
HaloProbe: Bayesian Detection of VLM Hallucinations
→ arXiv
GLM-5.1 HuggingFace 同步上架与 OpenBMB VoxCPM2
GLM-5.1 on HuggingFace & OpenBMB VoxCPM2 TTS Release
→ HuggingFace
NovaVoice 登顶 Product Hunt 4/7：AI 语音助手的桌面化
NovaVoice Tops Product Hunt With 547 Votes
→ Product Hunt
Anthropic Claude 4 月 6-7 日全球性服务中断
Anthropic Claude Global Outage on April 6-7
→ Status 报告 · 多方社区讨论
OpenAI、Anthropic、Google 联手反制中国模型蒸馏
OpenAI, Anthropic, Google Unite to Combat Chinese Model Distillation
→ Bloomberg · Frontier Model Forum
Anthropic 4 亿美元收购 Coefficient Bio 进军药物发现
Anthropic Acquires Coefficient Bio for $400M
→ TechCrunch · The Information · BioSpace · Fierce Biotech
Google TurboQuant：KV Cache 6 倍压缩、零精度损失
Google TurboQuant: 6x KV Cache Compression With Zero Accuracy Loss
→ Google Research Blog · VentureBeat · TechCrunch · HPCwire
🔄 DeepSeek V4 开启内测，确认原生运行华为昇腾 950PR
DeepSeek V4-Lite in API Testing, Runs on Huawei Ascend 950PR
→ Reuters · Tech Startups · 36Kr
OpenAI 政策白皮书：四天工作周与税制改革
OpenAI Proposes Four-Day Workweek and Tax Overhaul
→ OpenAI · 政策文件报道
NVIDIA Cosmos Reason 2：物理 AI 专用推理 VLM
NVIDIA Cosmos Reason 2: Reasoning VLM for Physical AI
→ HuggingFace · NVIDIA
Falcon-H1R-7B：混合架构测试时缩放推理模型
Falcon-H1R-7B: Hybrid Model for Test-Time Scaling
→ HuggingFace · TII
TriAttention：三角函数 KV 压缩实现 2.5 倍吞吐
TriAttention: Efficient Long Reasoning with Trigonometric KV Compression
→ arXiv
Vero：通用视觉推理的开源 RL 配方
Vero: An Open RL Recipe for General Visual Reasoning
→ arXiv
QED-Nano：4B 小模型证明奥数级定理
QED-Nano: Teaching a Tiny Model to Prove Hard Theorems
→ arXiv (LM-Provers)
CoDE-Stop：基于置信动态的推理早停
Early Stopping for Large Reasoning Models via Confidence Dynamics
→ arXiv
隐藏推理模型的可解释性研究
Are Latent Reasoning Models Easily Interpretable?
→ arXiv
Acemoglu 论文：AI 聚合如何影响集体知识
How AI Aggregation Affects Knowledge
→ arXiv
UnitedHealth 30 亿美元押注 AI 自动化
UnitedHealth Group $3B AI Push
→ STAT News
阿里巴巴 Accio 突破 1000 万月活
Alibaba Accio AI Sourcing Tool Hits 10M MAU
→ MIT Technology Review
Google Gemma 4 开源模型家族发布
Google Releases Gemma 4 Open Model Family
→ Google Blog · Engadget · The Register · Google DeepMind · Interconnects · Android Developers Blog · 新华社
AI 攻击性网络能力每约 6 个月翻倍
AI Offensive Cyber Capabilities Doubling Every ~6 Months
→ The Decoder · International AI Safety Report 2026 · Gnoppix Forum
DeepSeek V4 即将发布：万亿参数开源 MoE
DeepSeek V4 Imminent: 1T-Parameter Open-Source MoE
→ NxCode · 36Kr · Mule AI Blog · Evolink AI · Renovateqr
美国软件工程岗位三年新高，2026 年增长 30%
US Software Engineering Jobs Hit 3-Year High, Up 30% in 2026
→ TrueUp · Techmeme · Business Insider
Netflix 开源 VOID 视频物体移除模型
Netflix Open-Sources VOID Video Object Removal Model
→ HuggingFace · arXiv
腾讯发布 HY-OmniWeaving 视频生成模型
Tencent Releases HY-OmniWeaving Video Generation Model
→ HuggingFace
Google Vids 2.0：免费 AI 视频创建工具
Google Vids 2.0: Free AI Video Creation Tool
→ Product Hunt
Mercury Edit 2：基于扩散 LLM 的代码编辑预测
Mercury Edit 2: Ultra-Fast Next-Edit Prediction via Diffusion LLM
→ Product Hunt
OpenRouter Model Fusion：多模型融合最优响应
OpenRouter Model Fusion: Multi-Model Response Fusion
→ Product Hunt
Cohere Transcribe：多语言语音识别模型
Cohere Transcribe: Multilingual Speech Recognition
→ HuggingFace
百度千帆 OCR 视觉语言模型
Baidu Qianfan-OCR Vision-Language Model
→ HuggingFace
AI Chatbot 流量增速是社交媒体的 7 倍
AI Chatbot Traffic Growing 7x Faster Than Social Media
→ The Decoder
阿里巴巴 Qwen 新推理强化学习算法
Alibaba Qwen New Reasoning Reinforcement Learning Algorithm
→ The Decoder
批量上下文强化学习：推理 Token 效率新范式
Batched Contextual Reinforcement: Task-Scaling Law for Efficient Reasoning
→ arXiv
AutoAgent：自动化 Prompt 优化和 Agent 调优开源库
AutoAgent: Automated Prompt Optimization & Agent Tuning Library
→ Planet AI
开发者对 "AI Slop" 的不满：定性研究
Developer Frustration Over "AI Slop": Qualitative Study
→ The Decoder
Meta 大规模 Codec Avatars：百万视频训练 3D 头像
Large-scale Codec Avatars: Avatar Pretraining on 1M Videos
→ arXiv
阿里巴巴发布 Qwen3.6-Plus
Alibaba Unveils Qwen3.6-Plus for Agentic AI
→ Bloomberg · Seeking Alpha · TechBriefly · TradingView
微软发布三款自研 MAI 模型
Microsoft Launches MAI-Voice-1, MAI-Transcribe-1, MAI-Image-2
→ VentureBeat · Windows Central · Decrypt · Microsoft AI Blog
Google 发布 Veo 3.1 Lite 视频生成模型
Google Releases Veo 3.1 Lite Video Generation Model
→ Google Blog · 9to5Google · MarkTechPost · Windows Report · Android Authority
H Company 开源 Holo3-35B-A3B Computer Use 模型
H Company Open-Sources Holo3 SOTA Computer Use Model
→ HuggingFace · H Company Blog · TestingCatalog · NeuraBooks
🔄 Ollama v0.19：Apple MLX 集成与 Web 能力
Ollama v0.19: MLX Framework, Web Search & VS Code Integration
→ GitHub · MacRumors · Product Hunt
OpenAI 联合创始人称 GPT 推理模型"看到了 AGI 的路径"
OpenAI Co-Founder: GPT Reasoning Models Have "Line of Sight" to AGI
→ The Decoder · llm-stats.com
"Therefore I am. I Think"：LLM 是先决策还是先推理？
Therefore I am. I Think — Do LLMs Decide Before They Reason?
→ arXiv
极简自蒸馏提升代码生成
Embarrassingly Simple Self-Distillation Improves Code Generation
→ arXiv
ORCA：推理校准降低 Test-Time 计算成本
ORCA: Online Reasoning Calibration via Conformal Prediction
→ arXiv
Noiz Easter Voice：设计有表现力的人声
Noiz Easter Voice: Design Expressive Voices
→ Product Hunt
traceAI：AI 应用评估和可观测平台
traceAI: Evaluation, Observability & Optimization for AI Apps
→ Product Hunt
CliffSearch：LLM Agent 驱动的科学算法发现
CliffSearch: Structured Agentic Co-Evolution for Algorithm Discovery
→ arXiv
facebook/sam3.1：SAM3 视频分割模型
Meta SAM 3.1 Video Segmentation
→ HuggingFace
Medvi：两人公司用 AI 实现 $4.01 亿营收
Medvi: $401M Revenue with AI-Driven Telehealth, Just 2 Employees
→ New York Times · llm-stats.com
HippoCamp 和 YC-Bench：Agent 能力评估新基准
HippoCamp & YC-Bench: New Agent Benchmarks
→ arXiv
Anthropic "Mythos" 模型因数据泄露意外曝光
Anthropic "Mythos" Model Leaked via Unsecured Data Store
→ Fortune (独家) · CoinDesk · CSO Online · Euronews · Futurism
Anthropic Claude Code 源码经 npm 泄露
Claude Code Source Code Leaked via npm Source Map
→ The Register · VentureBeat · Fortune · Axios · CNBC · CyberSecurityNews
Qwen3.5-Omni 全模态模型发布
Qwen3.5-Omni Native Omni-Modal Model Release
→ MarkTechPost · Analytics Vidhya · The Decoder · The Information · Product Hunt
🔄 OpenAI 完成 $1220 亿融资，估值达 $8520 亿
OpenAI Closes $122B Round at $852B Valuation
→ CNBC · Bloomberg · OpenAI Blog · TechCrunch
GPT-5.4 Mini 和 Nano 发布
GPT-5.4 Mini and Nano Release
→ OpenAI Blog · 9to5Mac · 9to5Google · Simon Willison · The New Stack
OpenAI 收购 Astral
OpenAI Acquires Astral — Ruff, uv, ty
→ OpenAI Blog · Astral Blog · CNBC · Bloomberg · Simon Willison · JetBrains Blog
🔄 Claude Computer Use 扩展至 Claude Code CLI
Claude Computer Use Expands to Claude Code CLI
→ Product Hunt · The Tech Outlook · Claude Code Changelog
OpenAI 公开内部编码 Agent 不对齐监控系统
OpenAI Publishes Internal Coding Agent Misalignment Monitoring Report
→ OpenAI Blog · LessWrong · Security Brief
Anthropic 投资 $1 亿建立 Claude Partner Network 并成立 Anthropic Institute
Claude Partner Network & Anthropic Institute Launch
→ Anthropic Blog
OpenAI ChatGPT 购物 + Agentic Commerce Protocol
ChatGPT Shopping & Agentic Commerce Protocol
→ OpenAI Blog · Releasebot
Anthropic 与澳大利亚签署 AI 安全合作协议
Anthropic Signs AI Safety Deal with Australia
→ US News · Reuters
LLM 自发涌现类脑功能分化
Spontaneous Functional Differentiation in Large Language Models
→ arXiv
无需训练的专家语言模型动态混合
Training-Free Dynamic Upcycling of Expert Language Models — DUME
→ arXiv
zed-industries/zeta-2 代码编辑预测模型
zed-industries/zeta-2 Next-Edit Prediction Model
→ HuggingFace
Jupid：用 Claude Code 报税
Jupid: File Your Taxes with Claude Code
→ Product Hunt
LiquidAI LFM2.5-350M 边缘部署模型
LiquidAI LFM2.5-350M Edge Model
→ HuggingFace
◉ 2026.03 ◉
OpenAI Codex Plugins 平台正式发布
OpenAI Codex Plugins Launch
→ OpenAI Blog · SiliconANGLE · Neowin · The New Stack · Windows Report
Gemini 3.1 Flash Live 实时音频模型发布
Gemini 3.1 Flash Live
→ Google Blog · MarkTechPost · SiliconANGLE · 9to5Google · Android Central
Suno v5.5 发布：声音克隆与个性化 AI 音乐
Suno v5.5: Voices, Custom Models & My Taste
→ Suno Blog · Digital Music News · Metaverse Post · Music Ally · Product Hunt
Claude Opus 4.6 与 Mozilla 合作：14 天发现 22 个 Firefox 漏洞
Claude Opus 4.6 Discovers 22 Firefox Vulnerabilities
→ Anthropic Red Team Blog · TechCrunch · The Hacker News · InfoQ · Axios · SC Media
Anthropic 经济指数报告：Claude 使用模式深度分析
Anthropic Economic Index: Learning Curves
→ Anthropic Research
Claude Tasks Mode 即将推出：五大任务起点
Claude Tasks Mode with 5 Starting Points
→ TestingCatalog · X (Twitter)
Claude Code auto-fix：自动修复 CI 失败和代码审查
Claude Code Auto-Fix for CI and PR Reviews
→ Product Hunt
Anthropic 对齐研究："The Hot Mess of AI"
The Hot Mess of AI: Misalignment Scaling
→ Anthropic Alignment Blog
Lightricks LTX-2.3 开源视频生成模型
Lightricks LTX-2.3 Open-Source Video Generation
→ HuggingFace
Tesslate OmniCoder-9B：开源代码 Agent 模型
Tesslate OmniCoder-9B
→ HuggingFace
Cohere Transcribe：22 语言语音识别模型
Cohere Transcribe ASR Model
→ HuggingFace · SiliconANGLE
Agentation：AI Agent 可视化反馈工具
Agentation: Visual Feedback Tool for AI Agents
→ Product Hunt
百度千帆 OCR 视觉语言模型
Baidu Qianfan-OCR Vision-Language Model
→ HuggingFace
OpenAI 发布 Safety Bug Bounty 计划
OpenAI Safety Bug Bounty Program
→ OpenAI Blog · Infosecurity Magazine · Help Net Security
Google DeepMind 发布 AI 操纵行为实证测量工具包
DeepMind AI Manipulation Measurement Toolkit
→ Google DeepMind Blog
xAI Grok 4.20 正式退出 Beta
Grok 4.20 Exits Beta
→ Artificial Analysis · WinBuzzer · xAI Release Notes
MCP 月下载量突破 9700 万
Model Context Protocol Hits 97M Monthly Downloads
→ Digital Applied · The New Stack · Anthropic Blog
Agile Robots 与 Google DeepMind 战略合作
Agile Robots Partners with Google DeepMind
→ TechCrunch · CNBC · Agile Robots 官网
Google Gemini 3.1 Pro 发布
Gemini 3.1 Pro Release
→ Google Blog · Google Cloud Documentation
Claude Opus 推理能力蒸馏进 Qwen3.5 霸榜 HuggingFace
Claude Opus Reasoning Distilled into Qwen3.5 Dominates HuggingFace
→ HuggingFace
Mistral 发布 Voxtral-4B 多语言语音合成模型
Mistral Voxtral-4B-TTS
→ HuggingFace
Stitch 2.0 by Google：AI 驱动的 UI 设计工具
Stitch 2.0 by Google
→ Product Hunt
Claude Import Memory：从 ChatGPT 迁移到 Claude
Claude Import Memory Feature
→ Product Hunt
WriteBack-RAG：将知识库作为可训练组件
WriteBack-RAG: Training the Knowledge Base
→ arXiv
Natural-Language Agent Harnesses：Agent 工程新范式
Natural-Language Agent Harnesses
→ arXiv
NSF 发布 AI-Ready America 计划
NSF TechAccess: AI-Ready America Initiative
→ NSF 官网
Lightfield：AI 原生自建 CRM
Lightfield AI-Native CRM
→ Product Hunt
OpenCLAW-P2P：去中心化 AI 形式化验证研究网络
OpenCLAW-P2P: Decentralized AI Research with Formal Verification
→ Hacker News · GitHub
Lightfeed Extractor：LLM 友好的网页结构化提取
Lightfeed Extractor for LLM-Ready Web Scraping
→ Hacker News · GitHub
Claude 桌面端 Computer Use 发布预览
Anthropic Claude Computer Use on Mac
→ Anthropic Blog · TechCrunch · CNBC · MacRumors
Claude Code Auto Mode 发布
Claude Code Auto Mode
→ Anthropic Blog · TechCrunch · SiliconANGLE · 9to5Mac
Google Gemini 全面升级 Workspace AI 能力
Gemini Workspace AI Upgrade
→ Google Blog · TechCrunch · VentureBeat
Apple Siri AI 升级由 Gemini 驱动，发布遭遇延迟
Apple Siri AI Upgrade Powered by Gemini
→ 9to5Mac · Bloomberg · TechCrunch · AppleInsider
OpenAI 营收突破 $250 亿，酝酿 IPO
OpenAI Revenue Surpasses $25B, Eyes IPO
→ AI News · Crescendo AI
🔄 GPT-5.4 全貌：百万上下文与 Computer Use
GPT-5.4 Full Feature Set
→ OpenAI Blog · TechCrunch
Anthropic 1M 上下文正式 GA
Anthropic 1M Context Generally Available
→ Anthropic Blog
Pathway：LLM 管道与 RAG 的流处理框架
Pathway ETL Framework for LLM Pipelines
→ GitHub Trending
中国开源模型在 HuggingFace 上超越美国
Chinese Open Models Overtake US on HuggingFace
→ HuggingFace Blog · AI News
Amazon 推出 Health AI Agent
Amazon Health AI Agent for Prime
→ AI News
UniGRPO：推理驱动的统一视觉生成
UniGRPO: Unified Policy Optimization for Reasoning-Driven Visual Generation
→ arXiv
SpecEyes：Agent 级多模态 LLM 推理加速
SpecEyes: Accelerating Agentic Multimodal LLMs
→ arXiv
VTAM：融合触觉的视频-动作世界模型
VTAM: Video-Tactile-Action Models
→ arXiv
NVIDIA CEO 愿景：2036 年每人配 100 个 AI Agent
Jensen Huang: 100 AI Agents Per Person by 2036
→ AI News · GTC 2026
Claude 新增交互式可视化能力
Claude Interactive Visualizations
→ Anthropic Blog
小米 MiMo-V2-Pro 万亿参数模型发布
Xiaomi MiMo-V2-Pro 1T Model
→ AI News · X (Twitter)
GPT-5.4 Mini 向免费用户开放推理能力
GPT-5.4 Mini Free for All Users
→ AI News · X (Twitter)
NVIDIA Nemotron 3 Super 开源最高 SWE-Bench 分数
NVIDIA Nemotron 3 Super
→ AI News · GTC 2026
Nemotron-Cascade 2: 级联强化学习训练 30B MoE 模型
Nemotron-Cascade 2
→ arXiv
Google Gemini Embedding 2 统一多模态 Embedding
Gemini Embedding 2
→ AI News
Andrej Karpathy：AI Agent 已能自主优化训练流程
Karpathy on AI Research Bottlenecks
→ AI News (The Decoder)
Gemini 3.1 Flash-Lite 效率模型发布
Gemini 3.1 Flash-Lite
→ AI News
Amazon Trainium 芯片实验室曝光
AWS Trainium Chip Lab
→ AI News (TechCrunch)
OpenAI 计划年底前翻倍至 8000 人
OpenAI Workforce Expansion
→ AI News (The Decoder)
LangChain 本周获 1151 Star，Agent 工程平台热度持续
LangChain Trending
→ GitHub Trending
F2LLM-v2：支持 200+ 语言的多语言 Embedding 模型
F2LLM-v2 Multilingual Embeddings
→ arXiv
Rowboat：开源多 Agent 系统 IDE
Rowboat Open-Source Multi-Agent IDE
→ Hacker News
NVIDIA GR00T N1.7 人形机器人基础模型
NVIDIA GR00T N1.7
→ AI News · GTC 2026
Microsoft GigaTIME 癌症病理多模态模型
Microsoft GigaTIME
→ AI News
Anthropic 成立 Anthropic Institute 研究 AI 社会影响
Anthropic Institute
→ AI News
Apple MLX 团队 2026 年重大更新，Local AI 年
MLX 2026 Release
→ X (Twitter)