行业追踪 · 当日动态流 60 条报道
当日 24h 要点速览
-
Anthropic · 发布 Claude 防工业级模型蒸馏安全报告，并披露全新面向研发者的智能体沙箱环境协议。
-
NVIDIA · 与微软联合推出面向本地 AI 推理的次世代硬件方案，单机最高支持 128GB 统一内存及 1 PFLOPS 算力。
-
Mistral AI · 公布 Ministral 3 模型与微型卫星星座联合部署进展，完成在轨边缘智能推理测试。
要闻精选
Anthropic's newest ad is creeping people out
Popular sugar substitutes linked to faster brain aging
Harness Engineering
Codex Resets
ECHR confirms ethical veganism is a protected philosophical belief
Juggling for Blind People
Even Microsoft couldn't make Windows 11 work well on 8GB of RAM
手搓了一个 agent 架构能减少 长 coding 任务的 llm 往返和 token 开销 80%左右。有 benchmark！
Gap in Mochizuki's proof of ABC confirmed by Lean
NYC may require landlords and realtors to disclose the use of AI in listings
AI Mania Is Eviscerating Global Decision-Making
Real-Time LuaTeX: Recompiling Large Documents in 1ms [pdf]
curl can be used to send emails with SMTP
Judge a book by its first pages
Classic Amiga titles, free to download
Hardcore IndieWeb: Run your own website 100% independently for only $0.01/day
Pico W firmware creates driverless USB WiFi bridge (Layer-2)
UnifiedIR for Julia
Co-evolution of self-replication and function in a digital primordial soup
Ada: An AI business intelligence software from CSV and Excel(yes LLMs but more)
Cornell's Interactive Wall of Birds
Supplement that binds to microplastics may remove them from our body
SpaceX and the myth of independent Wall St research
Intel Starts Shipping High-NA EUV Silicon
Typing Speed Test, but for Developers
Waymo says San Francisco service has resumed after one-hour pause
What's the deal with all the random weekly quota resets for agents lately?
Kimi: Threat or menace?
[Rewrite Bun in MC++] 7 天 20+万行代码 C++26 模块化, 重写 Bun 行为艺术复刻, 已能真实运行顶级开源 Web 框架项目
REO Trucks I4 4WD Pickup Truck Starts at $21,500
The unreasonable difficulty of time series forecasting
Show HN: Get alerts for good seats at 70mm IMAX showings of The Odyssey
现在机场价格这么离谱了吗
In Germany if you say a restaurant is just ok they send the gestapo after you
The Kimi K3 Moment
Goodbye, and Thanks for All the Bikesheds
Overtraining as the path to human-like AI
OpenEvidence is bribing spouses of CMI and Genentech leadership
Heresy
All the EVs that were discontinued or killed off in the US this year
Young adults are poor despite every metric which suggests otherwise
Setting up your spare Mac for Claude Code to control, a step-by-step guide
Fake food delivery site for the dopamine
Frozen 2 should be Rated R
No link between acetaminophen use during pregnancy and adverse birth outcomes
Our Approach to Bioresilience: Isomorphic Labs and Google DeepMind
Federal employees can download TikTok on their work phones again
能否把逻辑跟 AI 说清楚，让它先不要写实际代码(pyside6 界面)。而是先用最简单的 html 把页面/按钮布局和伪数据把图表/折线图等递进的逻辑关系表达出来？ 经过审核，觉得可以再用代码实施呢？
Gleam Is Now on Tangled
The Fermi Paradox, Percolation, and Inbreeding
If You Build It, They Will Come
Elixir-lang.org has a new design
What else do people draw on gradient.horse?
Traders are increasingly betting against SpaceX just weeks after IPO
Show HN: Q3Edit – Edit and play Quake 3 maps in the browser
British runner Josh Kerr breaks world record for mile which stood for 27 years
A 600-mile road trip (and data) proves EV charging doesn’t suck anymore
LG ThinQ Terms of Use
EU ban on destruction of unsold clothes and shoes enters into application
社区音量 · 领军人物与专家观点 X / 社区精选
Demis Hassabis @demishassabis · 2 小时前
祝贺科研团队！将大模型多智能体诊断系统应用于真实前瞻性临床试验，取得令人振奋的准确度突破。
Andrej Karpathy @karpathy · 4 小时前
当端侧小模型的推理能力超越两年前的千亿大模型时，软件的系统架构将彻底重塑，智能体正在成为原生操作系统层。
Guillermo Rauch @rauchg · 6 小时前
毫秒级启动、与模型解耦的极简 CLI 工具是未来。内置 MCP、Skills 与子智能体后，终端将重焕新生。
Ethan Mollick @emollick · 8 小时前
如今已有十余款可在现代移动设备上全本地运行的紧凑模型，整体表现逼近甚至超越昔日的 GPT-4 基准。
产品发布 · Show HN & Product Hunt 当日发布
OpenCode 2.0
▲ 312全开源 AI 编程智能体，原生支持终端、桌面与网页端协同工作流。
MemorySync MCP
▲ 185为自主智能体提供跨会话持久记忆与沙箱化知识库检索的开放服务。
Articos Agent
▲ 278在 Claude 与 ChatGPT 中模拟真实买家反应，上线前自动验证营销与产品策略。
ExcaliClaw
▲ 146支持 AI 智能体通过协议实时绘制、调整与解释技术架构手绘图的工作区。
开源热榜 · GitHub Trending Trending
专为智能体打造的轻量技能协议工具包，支持极简 npm / npx 直接调用。
功能完备的本地模型交互与多智能体统一管理控制台界面。
微软官方 1-bit LLM 推理与训练框架，针对能效与端侧极致优化。
低显存环境下运行超大参数模型的分层推理引擎，大幅降低本地部署门槛。
模型生态 · OpenRouter 调用排名 & 新模型 24h 统计
Top 模型调用份额 (Tokens)
01
DeepSeek V4.1 Flash DeepSeek · 6.02T
02
Step 5 Preview StepFun · 3.25T
03
GLM 5.3 Flash Zhipu AI · 2.03T
04
Claude Opus 5.5 Anthropic · 1.42T
05
GPT-6 Luna OpenAI · 1.18T
06
Nemotron 3 Ultra NVIDIA · 1.05T
近期模型上线与定价
Muse Wonder Meta · 1.0M ctx · 今日
$4.00 / $20.00 Nemotron 3.5 Bolt NVIDIA · 256k ctx · 昨日
免费调用 MiniMax Abab 7 MiniMax · 512k ctx · 前日
$0.80 / $2.40 融资并购 · 创投交易 当日记录
$90M E 轮
医疗 AI 语音及临床工作流公司完成 9000 万美元融资，加速推进生成式临床文档自动化与多院区落地。
Healthcare AISeries E
$49M 政府采购
气象气球与深度学习天气预测初创公司获美国空军近 5000 万美元订单，提供海量高空大气训练数据。
Climate AIDefense
并购重组 破产保护 / 资产出售
生成式对话初创公司启动资产出售程序，部分核心业务与团队将整合进主流影视与互动娱乐巨头。
M&AConsumer AI