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
Iran threatens to hit Gulf states if US launches new strikes
No one can afford to make Myst games anymore
Literary Sins
Let’s take apart your phone
A physicist rigged his pet hamster’s wheel to upload to Strava
Nikita Bier steps down as X’s head of product
The title cards in Blade Runner are amazing
Travis Kalanick’s robotics startup Atoms taps former Uber finance chief as CFO
Nvidia’s Vera Whitepaper Has a Thread Loose
Meta launches Muse Code, an AI agent for large code bases
I'll be stepping back from leading product for X
Prime Agent: A self-improving RLM agent
Trump’s DOJ gains oversight of OpenAI’s green-card employee sponsorships
OpenAI says my prepaid credits were consumed, refuses to show any record
When online commenters detect my art as AI
Moove raises $250M to become the backbone of the robotaxi industry
How Lightspeed found its newest hire … via Instagram DM
Lightspeed is building its edge on followers, not just funds
Cross-compilation with Wails (2025)
Klaviyo acquires Elias Torres’ Agency in full-circle reunion for tech founders
I'm switching my phone from Android to Linux
Meta Ran Ads That Contained AI-Generated Child Sexual Abuse Imagery
The Job Board Is Lying to You
Jeff Dean and other top AI researchers are leaving Google to launch their own startup
How to Survive in a Louisiana Swamp
Ban the Throbber
Show HN: My receipt printer prints an original artwork every morning
Muse Code and Muse Spark 1.2
You're not uncool. Making friends as an adult is just hard (2021)
Sula: A Gemini protocol server written in Scryer Prolog
Zed DeltaDB
London cops handed victim's new address and number to her stalker, watchdog says
What happens if you put work into the second dimension?
Born Against, or why hobby programming communities are against LLM usage
Show HN: Hacker News with reduced priority for AI driven content
Microsoft's AI Sales Mostly Come from OpenAI, Disclosures Show
Beating GPT-5.6 Sol on retrieval with 100x cheaper open models
Show HN: ClickBench Playground – a playground for 110 database systems
Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence (2025)
COLDCARD's Random Numbers Weren't
Antigravity 2.0
Reddit aims to make ‘karma’ less important for first-time posters with shift to AI moderation tools
Lucid Motors just delayed its affordable EV. Now what?
Atlassian Rovo Exfiltrates Data, Bypassing Controls
Phishers are hijacking legitimate cloud infrastructure
AI is a bubble, just like dot-com
Meet the eight startups pitching at Startup Battlefield Australia
Senators demand crackdown on wildfire "prediction market" bets
I'm Scared a Stranger Will Call My Novel AI, So I Built GitHub for Words
Your model already knows the answer: how benchmark answers leak into LLMs
PSA: Apple’s Private Relay can leak your real IP address
Celld: Self-hosted, distributed Durable Objects
Launch HN: HyperProbe (YC S26) – Agents that do read-only debugging in prod
I’m leaving OpenAI to build telepathy
Hollywood and Shanghai Cinema in the 1930s [pdf]
Stop sending me your errors
Detach from your work to make it work
Something is changing in the unit economics of software
Discovery Loop
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