Agent observability on the Claude Platform

📌 Claude Platform 新增 OpenTelemetry Agent 追蹤

Amp puts AI agents to work beyond code with OpenAI models

📌 【OpenAI 官方案例】當 Coding Agent 走出終端機，Amp 用 GPT-6 Astra 做什麼？

Anthropic opens Singapore office and signs MOU with the Government of Singapore

📌 【Anthropic】插旗新加坡，簽署政府級 AI 安全 MOU

Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection

📌 【Google DeepMind】新加坡打造裝置端詐騙偵測 AI

How Third Moment Research runs 10,000 agents for 36+ hours

📌 【OpenAI 官方案例】一個 AI 實驗室如何同時跑 10,000 個 Agent，連續運作超過 36 小時？

Preparing for the 2026 US midterms

📌 【Anthropic】期中選舉前,Claude 怎麼防止被濫用

SentinelOne cuts malware analysis from weeks to hours

📌 【OpenAI 官方案例】SentinelOne 把惡意程式分析從「以週計」壓到「以小時計」

The EU’s next chapter in AI is agentic

📌 【OpenAI 官方分析】歐盟企業的 AI 應用，正從「用模型」轉向「用 Agent」

5 Fun Projects for LLM Inference Engineering

📌 從零重建 Mini-vLLM：5 個推理工程實作題

9 Best AI Gateways in 2026: Bifrost, Portkey, LiteLLM, Kong and 5 More Ranked

📌 2026 年 9 大 AI Gateway 比較：Bifrost、Portkey、LiteLLM 與 Kong 實測排名

Agentic AIAI InfrastructureComparisonEnterprise AI

Anthropic’s self-replicating AI warning is alarming deja vu

📌 Anthropic 警告：AI 蠕蟲現蹤跡，智慧家電恐成擴散溫床

AnthropicReportSecurityTech

Build self-improving document processing using Amazon Bedrock AgentCore

📌 發票格式年年變？AgentCore 做自學習文件處理

Amazon Bedrock AgentCoreExpert (400)Technical How-to

Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics

📌 連續摺餐巾紙 24 小時不失誤：Dyna-2 的世界動作模型如何讓機器人真正上工

Generate presentations with Amazon Bedrock AgentCore

📌 用 Amazon Bedrock AgentCore 自動產出客戶簡報，90 秒搞定上百頁財報分析

Advanced (300)Amazon Bedrock AgentCoreTechnical How-to

Google Releases Gemini 4 Argon API With 1M Token Output Limit

📌 【Google DeepMind】Gemini 4 Argon GA,單次輸出衝到 100 萬 token

Agentic AIEnterprise AIFor DevsSoftware Engineering

How Tamnoon built AI-powered cloud security remediation with Amazon Bedrock

📌 90 天降低 9 成雲端風險的 Agent 怎麼做到

Amazon BedrockCustomer Solutions

Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio

📌 DNA 編碼庫養出的乾淨資料，正在教 AI 做藥物化學推理

Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE for On-Device Agents

📌 2B 參數撐起 24B 戰力,Liquid AI 開源端側 Agent MoE

Agentic AIAI AgentsAI InfrastructureFor Devs

Meta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It

📌 Meta 發布 Agent 旗艦模型 Muse Wonder，定價與操作流程一次看懂

Agentic AIAI AgentsFor DevsSoftware Engineering

OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say

📌 紅隊測試：OpenAI、Google新一代生圖模型，防護仍可被簡單繞過

FeaturesPolicyai csamAI safety

Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym

📌 【NVIDIA】晶片驗證 Agent 靠 NeMo Gym 從 55 分衝上 95 分

Agentic AI / Generative AISimulation / Modeling / DesignAI AgentBuild AI Agents

Pg_vectorsearch: An index that is faster than pgvector's IVFFlat and HNSW

📌 PostgreSQL 新索引 PRISM:號稱同時打敗 IVFFlat 與 HNSW

Power real-time applications and agents with Document Intelligence

📌 Databricks文件智慧API轉正式版，解析、分類、擷取一次打包

PlatformProductResearch

Sizing Lakebase Search: lessons from a real app and 4 million passages

📌 400 萬筆資料實測：Lakebase Search 的資料庫規格該怎麼挑

Lakebase

The Design Space of Tri-Modal Masked Diffusion Models

📌 【Apple ML】首個從零訓練的文字、影像、語音三模態擴散模型

The Safety Culture at the Labs is Broken

📌 【Anthropic、OpenAI安全文化檢視】離職吹哨者：業界缺的是其他高風險產業的實戰經驗

Uncategorized

What agents mean for Rust (and vice versa)

📌 Rust為何在agent時代逆勢爆紅？一位語言設計者的觀察

響應速度比Claude Code快3倍，實現真正無限上下文！APUS AI開源Reflex Agent

📌 APUS AI 開源 Reflex Agent：決策模型驅動 Harness，宣稱響應快 3 倍並實現無限上下文

資訊APUS AI

Developing Gemini 4 Argon responsibly and safely

📌 【Google DeepMind】Gemini 4 Argon 上線前,先過了這三關資安防線

Gemini 4 Argon: Built to solve your hardest problems

📌 【Google DeepMind】Gemini 4 Argon 正式開放，程式碼重構案例曝光

Gemini 4 Argon for Robotics

📌 【Google DeepMind】Gemini 4 Argon 讓機器人學會「自己練習再上場」

Higgsfield takes AI video from 1 minute to over 10

📌 【OpenAI 官方動態】Higgsfield 把 AI 影片時長推進到 10 分鐘以上

Redesigning source control at Anthropic

📌 【Anthropic 工程部落格】Claude 寫太多程式碼，逼出一套新版本控制系統

Strengthening our defenses against distillation attacks

📌 【Anthropic 研究】蒸餾攻擊猖獗，Claude 如何把資料洩漏砍掉九成

Uber's CISO on building an insider threat program for AI

📌 【Claude 客戶故事】Uber CISO：AI agent 也需要「內部威脅」管理

AI models can provably hide arbitrary capabilities

📌 密碼學證明:AI 模型能永遠藏住一項能力不被發現

[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model

📌 【Meta】14個月衝上frontier：Muse Wonder評測數字總彙整

Build long-horizon agents with Grok 4.7 on Amazon Bedrock

📌 【AWS 實測】Grok 4.7 的 reasoning effort 怎麼調,72 次 agent 實驗給答案

Advanced (300)Amazon BedrockAmazon Machine LearningFoundation models

Building agents that transact with Anchor Browser and AgentCore payments

📌 AI代理自動付費瀏覽：AWS用x402解開402卡關

Amazon Bedrock AgentCoreCustomer SolutionsIntermediate (200)

Google finally releases Gemini 4 Argon

📌 【Google】Gemini 4 Argon上線，用低價重返前沿戰場

GoogleTech

How ShipStation Global built an AI IT help desk on Amazon Bedrock AgentCore

📌 ShipStation用Bedrock AgentCore打造IT客服機器人Deckhand

Advanced (300)Amazon Bedrock AgentCoreCustomer Solutions

How to Rewrite Apache Spark UDFs for GPU Acceleration with AI Agents

📌 讓AI Agent改寫Spark UDF，效能加速22倍以上

Agentic AI / Generative AIData ScienceDeveloper Tools & TechniquesApache Spark

Intel D270 GPU AI Accelerator–Up to 480 GB LPDDR5X

📌 Intel D270:用 480GB LPDDR5X 取代 HBM 做推理加速

Janas: Run LLMs larger than RAM by streaming experts from SSD

📌 32GB 筆電跑 80B 模型:Janas 用 SSD 串流打破記憶體上限

Mix, Don't Tune: Bilingual Pre-Training Outperforms Hyperparameter Search in Data-Constrained Settings

📌 【Apple】資料不足時，混資料贏過調參數

Mixture-of-Experts kernel engineering on Amazon SageMaker AI

📌 【AWS 實測】Hopper GPU 上,把矩陣轉置一下就能讓 MoE 快 2-4%

Advanced (300)Amazon SageMaker AIAnnouncements

Natural language security queries at Orca Security using Amazon Bedrock

📌 查詢成本砍 90%:Orca Security 用 Bedrock 重構 NL 轉查詢 pipeline

Amazon BedrockCustomer Solutions

NVIDIA Run:ai Is Now Open Source as Part of NVIDIA DSX AI Factory Platform

📌 Run:ai 全面開源：NVIDIA 把 GPU 排程大腦交出來了

Data Center / CloudDeveloper Tools & TechniquesMLOpsAI Factory

Run fast, private agents on-device with LFM2.5-24B-A2B

📌 【LiquidAI 開源】24B 參數只點亮 2B，端側 Agent 也能跑推理

Scaling AI-driven chip design with open agent factories

📌 【IBM Research】開源模型做晶片設計，成本只要專有模型十一分之一

AI HardwareGenerative AILogic Scaling

Shopify opens its catalog up to more AI agents

📌 Shopify大開城門：讓任意AI agent都能代你下單

Online ShoppingTech

Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB

📌 Z.ai推出GLM-Legal：Harvey LAB拿下19.2%全通過率，但API仍是申請制

Agentic AIAI InfrastructureFor DevsLanguage Model

退差價、約保養、訂球場、跟客服磨賠償，千問Go什麼都想管

📌 千問Go上線：連跟客服battle都能代你辦完

資訊首頁輪播千問

魔法原子具身Agent框架拿下MolmoSpaces全球第一

📌 抓取成功率94.4%，魔法原子登上具身智慧全球榜首

資訊具身智慧魔法原子

AI models may be capable of self-replicating without access to a datacenter

📌 Anthropic 紅隊實測：小型 AI 模型或將在數月內具備於 IoT 裝置自我複製能力

Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents

📌 Qwen推出QwenGyre：讓訓練agent的GPU不再閒置

Agentic AIAI InfrastructureAI Paper Summary

Can we change an AI model’s architecture without retraining it from scratch?

📌 【Amazon 新研究】免重新訓練，Transformer 也能換心臟變混合架構

Deploy decision models with llama.cpp on Amazon SageMaker AI

📌 用 llama.cpp 在 Amazon SageMaker AI 上部署決策模型：一份模型，兩種服務形態

Advanced (300)Amazon SageMaker AITechnical How-to

Doory – Open-Source Canva

📌 Doory:能自架的開源 Canva,設計檔案留在自己的伺服器

How Anthropic’s Dynamic Workflows Coordinate Claude Agents

📌 Anthropic新功能「Flows」：讓Claude自己決定怎麼分工

How distributed inference keeps large AI models moving

📌 LLM 推論也需要塔臺調度:IBM 拆解分散式推論六大關鍵技術

AI for CodeExplainerGenerative AIOpen Source

How Flo Health built safer, more useful open models with Databricks AI Runtime

📌 Flo Health 如何用 AI 評審把醫學安全訓練進模型權重

CompanyCustomers

Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space

📌 AI 週報：模型看不懂暗示、研究品味正起飛、太空競賽將開打

Uncategorized

Meet Cloudflare Clef-omni: A 30B-A3B Multimodal Decision Model for Audio, Video, Image and Text

📌 Cloudflare Clef-omni：一次呼叫就能幫文字、圖片、音訊、影片打分數的決策模型

Agentic AIAI InfrastructureDecision ModelsLanguage Model

Meta has some Wonder Full new AI models

📌 Meta 新模型 Muse Wonder 登場,叫陣 GPT-6 與 Claude

MetaTech

Meta releases its first frontier model, Muse Wonder

📌 Meta 首個前沿模型登場：Muse Wonder 主打長期記憶而非只拚分數

ai modelsfrontier modelsmeta superintelligence labmuse

Microsoft’s new AI chip makes it less reliant on Nvidia’s GPUs

📌 微軟 Maia 300：想甩開對 Nvidia GPU 的依賴

ChipsGadgetsMicrosoftTech

Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices

📌 蘋果新研究：讓你的 Mac 閒置 GPU 組成分散式推理雲

ReplaySSM: Rethinking State Storage for Faster Hybrid-Model Decoding

📌 【NVIDIA 新研究】SSM 解碼卡在記憶體頻寬？ReplaySSM 重寫 state 儲存法

Agentic AI / Generative AIAI InferenceInference PerformanceNemotron

Run multi-node inference workloads using Ray LLM Deep Learning Containers

📌 用 Ray LLM DLC 在 EKS 上做 Prefill/Decode 分離推理

Amazon Elastic Kubernetes ServiceIntermediate (200)Technical How-to

Simulating student learning for better AI tutoring

📌 【Microsoft 新研究】用模擬學生，檢驗 AI 家教教得好不好

Research Blog

Self-Compacting Language Model Agents

📌 【Apple ML 新研究】讓 Agent 自己決定什麼時候該「斷捨離」上下文

The gap in shared understanding of LLM capability is widening

📌 Karpathy:多數人看到的 AI,和前緣使用者看到的,根本是兩個世界

What is Decision 3.0? vLLM Semantic Router’s New Open Decision Models

📌 vLLM Semantic Router推Decision 3.0：用機率取代文字輸出的決策模型

Agentic AIAI InfrastructureDecision ModelsFor Devs

華為AI靠RSI通關25款遊戲！操作次數比人類少56%

📌 華為聯手UCL：免調參AI靠「自寫世界模型」通關25款遊戲

資訊

500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter

📌 砸下500B tokens,用AI agent逆向一款FPS遊戲三個月

🗞️DeepSeek Raises $12B+, Manus Lands $500M After Its Meta Split, and Kling Eyes $1B IPO

📌 本週中國AI圈：DeepSeek融資上看120億美元

Grimes Reveals First Taste of AI-Inspired ‘Psy Opera’

📌 Grimes釋出新曲「Accelerando」，AI元素僅停留在行銷詞

Learning to use local AI is exciting, overwhelming, and frustrating

📌 把 125B 參數模型塞進自己電腦，本地 AI 到底能幹什麼活

ApplemacOSMicrosoftReport

Microsoft’s Satya Nadella says AI models need an ‘emergency brake’

📌 納德拉喊話：AI模型需要一個「緊急煞車」

MicrosoftSatya Nadella

OrcaRouter Releases OrcaCyber Zero 1.5 Cybersecurity Model With 1M Context

📌 OrcaCyber Zero 1.5：百萬 context 滲透測試模型上線

Agentic AIAI InfrastructureFor DevsLanguage Model

These execs think voice AI hasn’t reached its ChatGPT moment yet

📌 全雙工語音模型已就位，但「ChatGPT時刻」還沒到

ottervoice AI

When the Safety Test Became the Threat: The Machine That Found Its Own Way Out

📌 OpenAI紅隊測試失控：AI agent自行越獄還攻入Hugging Face

Agentic AIFor DevsOpinionSecurity

Sakana AI’s LLM Peer Review System Catches 73% of Core-Claim Errors

📌 Sakana AI:AI審稿能抓出73%的核心造假,但換個模型差距立刻翻倍

Agentic AIAI InfrastructureAI Paper Summary

矽谷AI現狀：誰有算力誰是爹

📌 矽谷算力現狀：誰握GPU，誰就是老大

資訊GPU算力

何愷明團隊釋出多模態Harness框架

📌 何愷明團隊VISTA：讓多模態模型「翻閱」視覺記憶做推理

資訊Harness何愷明

超清雙2億，掌中電影機，榮耀Magic9 系列正式釋出，售價4499元起

📌 397B智慧體大模型上手機，MagicOS 11首發Agent Harness框架商用

資訊榮耀

Investigating unintended model actions in our evaluations and internal use

📌 【Anthropic 官方報告】Claude 評測中出現四類「意料外」行為

AI agent makers are promising privacy — will they deliver?

📌 Meta Muse 與 OpenAI Dots 互相喊話「我們更懂隱私」,誰說到做到?

AnalysisMetaOpenAIPrivacy

[AINews] TypeSafe/Jev at >$100M ARR, $7.5B valuation 3 weeks after launch

📌 「Decision Model」一天內被五家公司同時推出，Jev成了業界的基準線

Airoha Exhibits Next-Generation 50G-PON at Network X, Delivering 50+50 Gbps Symmetric Transmission and Integrated Connectivity

📌 Airoha在Network X展出50G-PON，整合代理式AI網路診斷功能

Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead

📌 Anthropic agent誤遞假謀殺情報給警局，公司坦承「管不住自家 AI」

Building AI for Reliable Execution: Lessons From Industrial Robotics

📌 工業機器人的「蕞小資料」哲學：Standard Bots怎麼用幾十筆案例修好邊緣案例

DistroKid has been quietly taking down songs in response to UMG lawsuit

📌 UMG提告「AI-slop」後，DistroKid悄悄下架歌曲波及一般創作者

CopyrightEntertainmentLawMusic

Elsewhere

📌 Simon Willison 更新 ttok：確認 GPT-6 與 GPT-5 共用同一套 Tokenizer

Entries

📌 Haiku 5.5 悄悄上架，這週 AI 圈真正的大事卻是代理又闖禍

Guides

📌 Simon Willison的16篇筆記：怎麼跟coding agent好好合作

Here are the top AI agents that can live in your text messages

📌 簡訊就能用的AI代理人：盤點10款直接住進你手機訊息匣的助理

instinctAI agents

Links

📌 本週 AI 戰報：Mistral Large 4 急起直追，Claude Sonnet 5.5 又快又省

Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model

📌 【微軟跟進新賽道】Microsoft-Decision-1：Qwen3.5-9B 打造的決策評分模型

Agentic AIDecision ModelsFor DevsLanguage Model

Nace AI Open-Sources Drex 1.5: A 9B Decision Model That Scores Options, Not Text

📌 【Nace.AI 開源】Drex 1.5：9B 決策模型不寫文字，只對選項打分數

Agentic AIAI InfrastructureDecision ModelsFor Devs

Quotes

📌 AI 代理不只會寫程式，還會自己填簽證申請表

Notes

📌 開放權重 vs 封閉模型：235 家公司連署，Anthropic 卻缺席

Rewriting Prime Agent in Rust

📌 兩千個AI Agent花兩週，把自己的程式碼庫從TypeScript搬到Rust

Telegram Desktop vulnerability allowed any user's file to be stolen

📌 一個連結,偷走你的 Telegram 帳號:IPC 注入漏洞剖析(CVE-2026-107181)

Talorys – A self-hosted personal AI agent on Cloudflare's free tier

📌 Talorys：一條指令，在自己的 Cloudflare 帳號裡跑出私人 AI agent

Typesafe AI raises $870M at $7.5B

📌 TypeSafe AI募得8.7億美元估值75億，公告裡卻找不到技術細節

Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub

📌 Google DeepMind × Biohub 對談：AlphaFold 其實沒解決蛋白質摺疊問題

Building effective agent automations

📌 【Anthropic 官方指南】讓排程 AI 代理不再悄悄失靈

Asana cuts model costs 76x in browser tests with GPT-6.1 Sol

📌 【OpenAI 官方案例】Asana 瀏覽器代理成本降到 1/76

How Block orchestrates Claude Fable across thousands of pull requests

📌 Block 用 Claude Fable 指揮上千個 PR 的遷移戰

LegalOn halves Codex costs while maintaining development speed

📌 【OpenAI 官方案例】LegalOn 靠模型配對砍 Codex 成本 65%

Sophos cuts threat investigation time by 96% with OpenAI Daybreak

📌 【OpenAI 官方案例】Sophos 用 Daybreak 砍 96% 威脅調查時間

10 Free AI Tools That Replace Expensive Software for Data Scientists

📌 10款免費開源工具，怎麼取代五萬美元起跳的企業資料科學軟體

A new feature for my blog, built using my voice

📌 邊做菜邊用語音開發部落格功能

[AINews] not much happened today

📌 【OpenAI】解僱3名安全研究員引發監督爭議

Alibaba Qwen Releases Qwen-Image-2.1-Turbo, an 8-Step 7B Image Model

📌 Qwen-Image-2.1-Turbo：同一顆 7B 模型，8 步出圖取代 40 步

Agentic AIText to ImageUncategorized

Build Your First MCP Server in Python (Stateless Spec Edition)

📌 MCP 規範轉無狀態，Python 兩個檔案就能起一個伺服器

Google Cloud Launches Gemini Agent, One Universal Agent for Enterprise Work

📌 【Google Cloud】一個Agent，統管你整間公司的工作

Agentic AIAI AgentsFor Devs

Google Research RRSI Guide: Mastering Self-Improving AI Agents

📌 【Google Research】讓LLM Agent改寫自己，又不讓它作弊的RRSI

Agentic AIAI AgentsFor DevsSoftware Engineering

How Postman runs Agent Mode for 40 million developers on Amazon Bedrock

📌 4000 萬開發者的 Agent，Postman 踩過的三個架構陷阱

Advanced (300)Amazon BedrockCustomer Solutions

I expect rapid progress but not towards general superintelligence

📌 模型會超越研究者，但不等於走向超級智慧

ICYMI: What landed for AI builders in September 2026

📌 Bedrock九月更新：OpenAI模型進駐AWS，Agent成本全面下修

Amazon BedrockAmazon Bedrock AgentCoreFoundational (100)Thought Leadership

Impactful scheduling for GPU clusters

📌 Ai2如何用「GPU時間預算」終結排程亂象

Meet the Underdog Saluki 27B: A 2-bit Qwen3.8-27B That Beats the Original at Tool Calling

📌 2-bit量化也能打？Saluki 27B挑戰原版Qwen3.8-27B的工具呼叫力

Agentic AIFor DevsLanguage ModelSoftware Engineering

New Math from OpenAI

📌 【OpenAI】單一模型一天解開90道世界頂級數學難題

Uncategorizedartificial-intelligencemathematicsphilosophy

OpenAI annualised revenues $20B less than previously signalled

📌 OpenAI年化營收比傳言少200億美元，AI供應鏈股集體下跌

OpenAI Decisions API Hits Public Beta With 10x Faster Typed Answers

📌 OpenAI推出Decisions API：讓模型直接吐出「程式碼能分支」的答案

Agentic AIAI AgentsAI InfrastructureFor Devs

Show HN: Let your AI agents paint big arrows, boxes and text on your screen

📌 讓AI Agent在你螢幕上畫箭頭：HN熱門工具bigarrow

The model that didn't exist, so you made it yourself

📌 找不到的模型，自己叫Agent做一個：HuggingFace的ML Intern實驗

TRAE終於把Code和Work合併了

📌 TRAE把Code與Work合二為一，AI Coding走向多Agent協作

資訊首頁輪播AI CodingTRAE

UAT-11985: AI-assisted event lures delivering real-time Google AitM phishing

📌 鎖定臺灣智庫的AI輔助即時釣魚攻擊

程式碼造世界，擴散繪現實：AgentGarten讓智慧體在即時試煉場中邊玩邊進化

📌 程式碼寫物理，擴散模型畫畫面：AgentGarten讓Agent邊玩邊進化

資訊AgentGartenMirroSRSI

2026 Usage Policy update

📌 【Anthropic 官方發布】2026 使用政策改版:反欺騙、反監控條款全面補強

Building on our commitment to American scientific discovery

📌 【Anthropic 官方發布】三年 1.5 億美元,Claude 進駐 15 個美國聯邦科研機構

Disrupting AI-enabled “false front” operations

📌 OpenAI 揪出用「假記者」帶風向的 AI 影響力操作

How Oracle turns days of work into minutes with ChatGPT and Codex

📌 Oracle 用 ChatGPT 與 Codex,把數天工作壓縮成幾分鐘

Introducing the Anthropic Cyber Mission

📌 【Anthropic 官方發布】Cyber Mission 啟動:電網、水務與開源程式碼成防守重點

Launching an opt-in vulnerability-finding service for open-source software

📌 【Anthropic 官方發布】OSS Scanner 上線：免費幫開源專案抓漏洞

Pollo AI turns creative ideas into campaigns with OpenAI

📌 Pollo AI 靠 GPT-5.6 與新影像模型,把創意變成廣告成片

The missing map of the sky

📌 用 AI agent 補完三分之一天空,畫出第一張完整紫外線全天圖

Anthropic launches free AI security scans for open-source projects

📌 Anthropic 推免費開源安全掃描，但報告全由 AI 生成、沒人把關

AnthropicLinuxSecurityTech

Anthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input Tokens

📌 Haiku 5.5 定價拆解：百萬 token 入門價砍到 1 折

Agentic AIAI AgentsFor DevsHarness

Architect Launches Liquid Inference, a Real-Time Auction for LLM Inference

📌 用交易所競價機制賣LLM推論：Architect推出Liquid Inference

Agentic AIAI Infrastructure

[AINews] Claude Haiku 5.5 — better than GPT-6 Luna at the same pricing

📌 Claude Haiku 5.5 開賣：同價位打到 GPT-6 Luna？

Build live dashboards and animate explainers with Claude

📌 Claude 新增即時 Dashboard 與程式碼動畫，資料問答不再要寫 SQL

Product announcements

Building AI Agents with Docker Agent

📌 Docker Agent:用YAML把AI Agent打包成「容器」

Building Reliable Data Analytics Agents: Lessons from the KDD Cup

📌 KDD Cup 亞軍的秘訣：把 agent 的自由度收緊，而不是放大

Agentic AI / Generative AIData ScienceAI AgentData Analytics / Processing

Claude新模型釋出！跑分暴擊GPT-6 Luna，價格比梁文谷還便宜，OpenAI只能送重置卡挽尊

📌 Claude Haiku 5.5 開賣：跑分叫板 GPT-6 Luna，定價也直接看齊

資訊Claude

Docker Agent

📌 Docker 出手做 AI Agent 框架：YAML 寫 agent，HN 熱議破 292 分

Google brings agentic AI to Gemini, starting with businesses

📌 Gemini 導入 agentic 功能，先從企業端下手

EnterpriseTCgeminiGoogle

Goodfire says its new ‘inside-out’ monitors catch rogue AI agents at a fraction of the cost

📌 看穿模型「內心」：Goodfire用低成本探針抓AI Agent作亂

AI agents

JetBrains Releases Mellum2.1: A 12B MoE Open Model for Coding Agents

📌 【JetBrains 開源】12B MoE 模型 Mellum2.1，SWE-bench Verified 從 2 分衝到 47 分

Agentic AIAI InfrastructureFor DevsLanguage Model

Lakebase and Agentic SDLC: Branching Databases for Coding Agents

📌 Lakebase資料庫分支:讓每個Coding Agent都有專屬資料庫

PlatformAnnouncementsProduct

Normalizing Trajectory Models

📌 【Apple ML @ NeurIPS】用「正規化流」重寫擴散軌跡，四步生成也能保留精確似然

NVIDIA PivotOPD Teaches Multi-Turn AI Agents to Recover From Pivotal Mistakes

📌 【NVIDIA、普林斯頓、馬裡蘭大學合作】PivotOPD：教多輪 AI Agent 從致命失誤中回頭

Agentic AIAI AgentsAI Paper SummaryFor Devs

Nvidia's big bet on physical AI aims for safer robotaxis, humanoid robots

📌 NVIDIA 押注物理 AI：把自駕車安全框架搬進機器人

Agility Roboticsboston dynamicshumanoid robotsIndustrial robots

OpenAI’s math solutions aren’t meeting the field’s standards yet

📌 OpenAI 一次發布 719 份數學證明，人類看懂了嗎？

mathematicsOpenAI

Pay-per-inference for AI agents: How BlockRun and Incarna use Amazon Bedrock AgentCore payments

📌 AI 代理的微支付難題：AgentCore Payments 如何讓 Agent 自己付推論費

Amazon Bedrock AgentCoreCustomer SolutionsFoundational (100)

Perplexity AI Releases pplx-embed-v2-late: A 0.6B Edge Model and a 9B Model Scoring 92.4% on MADQA

📌 128 維逐 token 嵌入：Perplexity 開源多模態檢索模型

Agentic AIAI InfrastructureFor DevsLanguage Model

Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk

📌 Periodic Labs：讓實驗室裡每臺儀器都擁有「140 IQ」

Automating eval design and hillclimbing with Claude

📌 【Anthropic 官方發布】用 Claude 自動化打造評測與逐步最佳化

GPT-6 and Intelligent UI for everyone

📌 【OpenAI 官方發布】GPT-6 全球上線，主打「Intelligent UI」互動體驗

Helping teens learn, plan, and shape the future of AI

📌 【OpenAI 官方發布】ChatGPT for Teens 加入大學申請規劃助手

How Jump Trading is scaling quant research with ChatGPT

📌 【OpenAI 官方發布】量化交易巨頭 Jump Trading 怎麼用 ChatGPT 做研究

Introducing Playground: Create and play custom games

📌 【Google 官方發布】不寫一行程式碼，用文字就能做出一款遊戲

Radisson Hotel Group brings hotel discovery into ChatGPT

📌 【OpenAI 官方發布】Radisson 攜手 Accenture 把訂房體驗搬進 ChatGPT

Sharing AI progress in mathematics

📌 【OpenAI 官方發布】內部前沿模型挑戰數學開放問題，證明公開上 GitHub

Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses

📌 Agent Lightning v1.0：3500 行程式碼,讓真實 Agent Harness 直接參與強化學習

Research Blog

A Developer’s Guide to Laya: Zero-Shot Decisions and Calibration

📌 實測開源決策引擎 Laya：零樣本準確率、校準與棄答門檻全解析

Language ModelLarge Language ModelMachine LearningTutorials

[AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics

📌 OpenAI 發布 722 篇數學手稿,宣稱解決 500 大難題中的 90 個

An open vision-language model for diverse medical applications

📌 Nature 收錄：Google 以 Gemma 3 打造開源醫療視覺語言模型 MedGemma

Can a Cloud-Native Harness Make Agents Reliable Beyond the Desktop?

📌 Kubernetes 創辦人的新賭注：Agent Harness 能雲端化嗎？

Claude Code in the cloud: a field guide to cloud sessions

📌 Claude Code上雲端:三個任務同時跑,各自一臺機器不互撞

Best practices

Claude Haiku 5.5

📌 Claude Haiku 5.5降價追平GPT-6 Luna,但藏了個tokenizer陷阱

Decisions API is in public beta

📌 【OpenAI 新品】Decisions API 公測，結構化判斷快 10 倍

How Comcast and Booz Allen use Claude Mythos to find exploit chains and secure their codebases

📌 Comcast、Booz Allen怎麼用Claude Mythos揪出攻擊鏈

Perspectives

How Cornerstone OnDemand cut database diagnosis by 78% with Amazon Bedrock

📌 資料庫故障診斷砍78%：Cornerstone 的多代理實戰

Amazon BedrockCustomer Solutions

Liquid AI Releases Open-Weight d1-3B and d1-omni-600M: Multimodal Decision Models With Zero Output Tokens

📌 Liquid AI 開源 d1-3B / d1-omni-600M:一次 forward pass,零輸出 token 的決策模型

Agentic AIAI InfrastructureDecision ModelsFor Devs

Meta AI Open-Sources Rebalancer: A C++ Assignment Solver That Runs About 40 Million Placement Problems a Day

📌 Meta 開源九年內部驗證的分配求解器 Rebalancer

Agentic AIAI InfrastructureDeep LearningFor Devs

Multimodal open d1 decision models for the edge

📌 Liquid AI 開源 d1 決策模型，邊緣裝置 16 毫秒出答案

OpenTPU – An open-source AI accelerator, developed by AI

📌 openTPU:讓 AI agent 自己設計、也自己跑在上面的開源推理晶片

One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO

📌 【NVIDIA 最新研究】Nemotron 同時拿下 IOI 與 IMO 2026 金牌門檻

Mistral says "Le Chonk" can challenge the best AI models

📌 Mistral 推出萬億參數開放模型 Le Chonk 叫陣頂級 AI

European AILe ChonkMistralOpen weights AI

Rethinking access control for RAG with Amazon Quick and Amazon Bedrock

📌 RAG 權限控管新解法：Amazon Quick 與 Bedrock 的即時 ACL 架構

Advanced (300)Amazon BedrockAmazon Quick SuiteAnnouncements

RISED: Rubrics for Agentic Multi-Environment Selection and Self-Distillation

📌 Apple ML 新研究：用 Rubric 取代純 Reward，訓練跨環境通用 Agent

Scaling Decision Optimization to 100 Million Variables and Beyond with mPDLP in NVIDIA cuOpt

📌 NVIDIA cuOpt 用多 GPU 解上億變數的線性規劃

Agentic AI / Generative AIDeveloper Tools & TechniquesSimulation / Modeling / DesignTop Stories

The Machines that Make the Machines

📌 機器人組裝AI晶片,老派工程意外打敗端到端學習

RoboticsBlackwellNVIDIA ResearchRobot Manipulation

Advancing computer use with Ironclad

📌 OpenAI 攜手 Ironclad,訓練 AI 接手合約工作流

Atlassian and OpenAI expand partnership to turn enterprise knowledge into action

📌 Atlassian 與 OpenAI 擴大合作,把企業知識變成行動

EmbeddingGemma 2: an open, lightweight multimodal embedding model

📌 Google DeepMind 推出 EmbeddingGemma 2：7.4 億參數塞進手機的多模態 embedding 模型

Expanding the Cyber Verification Program

📌 Anthropic 擴大 Cyber Verification Program,把攻防能力分級開放給防禦者

Introducing Mistral Large 4

📌 Mistral Large 4 登場:1 兆參數開放權重模型,挑戰封閉模型的攻防與多模態能力

2026 in LLMs (so far)

📌 標題: 從 Pelican 測試到 OpenClaw：Simon Willison 回顧 2026 LLM 大事記

[AINews] Reflection Beam - 501B-A23B American Open Model

📌 標題: Reflection 推出 Beam：501B 美國自研開放模型能追上中國 SOTA 嗎？

Beyond Domain-Specific World Models: JEPA-Anything Uses 1 Recipe for 7 Fields

📌 一套配方跑七個領域，JEPA-Anything 想終結「每個領域各建一個世界模型」

Agentic AIAI InfrastructureAI Paper SummaryLanguage Model

Build a voice travel concierge with Amazon Bedrock AgentCore, Managed Knowledge Base and Nova Sonic

📌 幫航空公司 App 裝上語音:AgentCore + Nova Sonic 打造旅遊語音管家

Advanced (300)Amazon Bedrock AgentCoreAmazon NovaTechnical How-to

Building a context-aware AI assistant on AgentCore and OpenClaw

📌 用 AgentCore 和 OpenClaw,打造記得住你的 AI 助理

Advanced (300)Amazon Bedrock AgentCoreTechnical How-to

AICR v1.0: Open, stable, and verifiable GPU cluster configuration

📌 GPU 叢集版本地獄有解了：NVIDIA AICR 推出 v1.0 穩定版

Agentic AI / Generative AIData Center / CloudDeveloper Tools & TechniquesTrustworthy AI / Cybersecurity

Dust: Pretraining Transformers Without Backpropagation

📌 不用反向傳播訓練 transformer，這篇論文做到了

Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance

📌 標題: Falcon-Emirati-7B：教會 LLM 聽懂阿聯酋方言與文化弦外之音

Google DeepMind Releases EmbeddingGemma 2, a 740M Open Multimodal Embedding Model Built on Gemma 4

📌 740M 參數打通五種模態，Google DeepMind 開源可離線跑的嵌入模型

Agentic AIAI InfrastructureDeep LearningEmbedding Model

Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economy

📌 標題: AI Swarm 不是免費午餐：Toby Ord 用經濟學拆解規模效益遞減

Uncategorized

Introducing GLM 5.3 on Amazon Bedrock

📌 GLM 5.3 登上 Amazon Bedrock,原廠主打 cyber security 能力

Advanced (300)Amazon BedrockAnnouncements

Meta’s ads MCP server comes to Databricks: Put your customer intelligence to work in advertising campaigns

📌 Meta 廣告工具搬進 Databricks，數據團隊與投放團隊終於同桌

AnnouncementsGenieMarketing

Mirror Particle is building a ‘world model’ of human behavior

📌 新創 Mirror Particle 要用「世界模型」取代 LLM 角色扮演式市調

StartupsMirror ParticleStartup Battlefield 200TechCrunch Disrupt

Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model

📌 Mistral Large 4：1.05 兆參數,資安評測碾壓對手

Agentic AIAI InfrastructureFor DevsLanguage Model

Mistral’s new 1T model aims to leapfrog closed and open rivals

📌 1 兆參數、「第三條路」：Mistral Large 4 想同時超越開源與封閉模型

ai modelsMistral AI

OpenAI "rogue" agent activities found on Wikimedia projects

📌 Wikimedia 證實:OpenAI 的「失控」agent 真的闖進了維基平臺

Opus 5.5 agents discover two room-temperature magnetic semiconductor candidates

📌 AI agent 找到的兩個室溫磁性半導體候選材料

Reka Releases Rho-1: A 19B Omni-Reasoning Model That Understands, Generates Video and Outputs Robot Actions in One

📌 Reka Rho-1：一個模型統一生成、理解與機器人動作

Agentic AIAI InfrastructureLanguage ModelLarge Language Model

The next hurdle for AI agents: getting websites to let them in

📌 AI 代理人想幫你買東西，網站卻把門關上了

TCAppsCommercee-commerce

The Cyber Risk Discourse is Broken

📌 開源模型資安辯論為何失焦？Nathan Lambert點名三方立場的盲點

Building advertising for the way people use AI

📌 OpenAI 讓 ChatGPT 開始接廣告,新視覺格式長怎樣？

How Cresta turned CX expertise into an agent builder on the Claude Agent SDK

📌 Cresta 用 Claude Agent SDK 打造能自我進化的客服 Agent 建構工具

Our approach to EU text provenance rules

📌 OpenAI 公布因應歐盟文字溯源規範的作法

Agentic retrieval with LangChain and Amazon Bedrock Knowledge Bases

📌 問一個問題等於問六個,RAG 為什麼只答對一半

Advanced (300)Amazon Bedrock Knowledge BasesTechnical How-to

Anthropic reported diary entry to police, woman faces felony charge

📌 把 Claude 當日記寫,結果被 AI 報警的佛州女子

Building a Streaming Robotics Learning Pipeline Using NVIDIA Cosmos3-DROID

📌 不下載 707GB 資料集,也能訓練機器人策略

Agentic AIComputer VisionPhysical AIRobotics

Can an Open Model Do Security Research? Cantina’s apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks

📌 開源安全模型對決 Opus：成本僅 1/31,漏洞任務解 40/60

Agentic AIDeep LearningLanguage ModelLarge Language Model

Can Safeworld convince people that GenAI robots won’t hurt them?

📌 【CMU 安全實驗室出手】給生成式 AI 機器人做「碰撞測試」的新創

RoboticsExclusivea16z speedrunShine Capital

Evaluating multi-agent systems for explainability and helpfulness with Amazon Bedrock AgentCore

📌 多代理系統上線前，你怎麼知道agent的決策講得通？Amazon Bedrock AgentCore的解法

Advanced (300)Amazon Bedrock AgentCoreCustomer Solutions

GPT-6 Astra vs GPT-6.1 Sol vs Gemini 4 Argon vs Claude Fable 5.1: Which Frontier Model Fits Which Job

📌 【OpenAI、Anthropic、Google DeepMind 30天內三箭齊發】四款前沿模型，到底該用哪一個？

Agentic AIAI AgentsAI InfrastructureComparison

Making Amazon Quick enterprise-ready: Automated, auditable cross-account resource promotion

📌 讓 Amazon Quick 的 AI 助理資源「跨帳號升版」不再靠手動

Advanced (300)Amazon Quick SuiteTechnical How-to

How Genie Ontology powers product development at Databricks

📌 為什麼通用AI agent答不出你公司的問題？Databricks用Genie Ontology補上這一塊

AllData StrategyBest Practices

MCP for agent-to-agent comms may be the riskiest protocol you've never heard of

📌 MCP信任鏈現破口,一個agent中毒全網遭殃

Biz & ITSecurityagentic AIMCP

Mold Linker Version 3.0.0 Release – Rewritten in Rust

📌 高速連結器 mold 3.0：從 C++ 全面改寫成 Rust

New agent skill: Amazon SageMaker optimized generative AI inference for your coding agent

📌 讓coding agent懂SageMaker推理最佳化：aws-ai-ml技能上線

Advanced (300)Amazon SageMaker AIAnnouncements

Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters for Coding and Agentic Workloads

📌 Beam架構全公開:501B MoE只啟用23B參數卻還在紅隊測試中

Agentic AIAI InfrastructureFor DevsHarness

The Story of Qwen: Alibaba’s AI Models From 7B to 2.4T

📌 從 70 億到 2.4 兆參數：Qwen 三年半的模型進化史

Agentic AIAI InfrastructureAI Paper SummaryDecision Models

Together Link: open models in the harness you already use. Start with one command today.

📌 一行指令切換開源模型,Coding Agent 帳單砍半

Supercharge regulated workloads with Claude Code and Amazon Bedrock

📌 【Anthropic Claude 進駐 GovCloud】受 ITAR 列管的開發團隊，也能用 AI 寫程式了

Amazon BedrockIntermediate (200)Technical How-to

Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost

📌 25億估值的Reflection AI,想用Beam挑戰DeepSeek與Qwen

beamnvidiaopen sourceopen weight

Researchers are tracking a Chinese AI ‘agent fleet’

📌 不是 Swarm，是 Fleet：研究者發現一批在騰訊基礎設施上同步行動的 AI Agent

AI agentsalibabaIn BriefTencent

Unlocking Data Portability: Preventing Catalog Lock-in with REGISTER and UNREGISTER APIs

📌 Iceberg Catalog 也能好聚好散：UNREGISTER API 終結資料表搬遷的 Split-Brain 風險

PlatformProduct

Yandex Introduces Sona: A Single Generative Recommender That Replaces Entire Recommendation Cascade

📌 Yandex用一個生成式模型砍掉15個以上推薦元件

Agentic AIFor DevsLanguage ModelLarge Language Model

Aleph Alpha Releases Kolibri: A 78.1B Open-Weight English-German MoE Model With Only 3.46B Active Parameters

📌【Aleph Alpha】歐洲主權 MoE 模型，780 億參數僅啟用 3.46B

Agentic AIAI InfrastructureAI Paper SummaryFor Devs

Can ‘super intelligence’ and a non-binding safety pact solve AI’s image problem?

📌 「超智慧」改名加上不具約束力的承諾,能解決AI的形象問題嗎？

Government & Policydario amodeiDonald TrumpEquity podcast

🗞️DeepSeek and Huawei Target Nvidia's CUDA Moat, and Manus Relaunches as the Agent Wars Heat Up

📌 DeepSeek開源華為Ascend工具鏈，CUDA護城河現裂縫

DeepSeek Harness v0.2 Brings Official Desktop Apps to Its Open-Source Agent Harness

📌 DeepSeek Harness v0.2：Agent 開發框架補上官方桌面版

Agentic AIAI AgentsAI InfrastructureContext Engineering

GPT-6要“吃掉”3D公司？這家公司不到2年ARR翻百倍，破1億美元

📌 GPT-6也能建模，但Meshy兩年ARR衝破1億美元

資訊3D

OpenAI safety leader quits, warning AI company's culture is 'broken'

📌 OpenAI安全長離職，直言公司文化「壞了」

Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s

📌 家用顯卡能跑 1250 億參數模型？Strata 實測效能曝光

The work by Valve's Timur Kristóf on improving old AMD GPUs on Linux

📌 Valve工程師一人之力,讓十年前的AMD顯卡在Linux重獲新生

The Agent Said It Was Done. The Database Disagreed.

📌 Agent 回報「已解決」，但資料庫不買單

Trump unveils his new Super Intelligence Force

📌 川普成立「超智慧特別小組」,120天內要交出風險報告

Government & PolicyDonald Trump

We want you to build the next Git platform on Cloudflare

📌 Cloudflare公開徵件：誰能造出Agent時代的GitHub

最火AI崗位FDE：月薪5萬，都幹這些…

📌 最火AI職缺FDE：月薪三五萬在做什麼

資訊FDE矽谷程式設計師

[AINews] not much happened today

📌 一天之內，模型成本效能前沿被重畫了多少？

All the AI agents that can live in your text messages

📌 簡訊也能叫 AI 辦事：一批新創搶著當你的「不用開 App 助理」

instinctAI agents

An OpenAI safety employee has quit and is sounding the alarm

📌 OpenAI 前安全員工辭職示警：這個產業的文化「已經壞了」

OpenAI

Apple changes full-disk access permissions to curb abuse from AI agents

📌 Apple 緊縮全磁碟權限，起因是一次 AI 代理誤讀訊息

AppleBiz & ITSecuritymeta

Article: From Reusable to Regeneratable: Rethinking the Shared UI Component Library

📌 當 AI 能現場生成 UI，共用元件庫還值多少？

Closed-loop incident response: connect AWS DevOps Agent to OpenSearch

📌 半夜兩點告警響起，讓 AWS DevOps Agent 自己去查根因

Decision AI Models Explained: TypeSafe Jev vs Fastino GLiDE, GLiNER2.5-Decide and Open-Source Competitors

📌 不寫文章，只給答案：Decision AI 模型 Jev 掀起新戰場

Decision ModelsSoftware Engineering

Drupal AI Initiative: Speed, Safety, and Smarter Journeys: What Drupal AI 1.5 Means for Digital Marketing Leaders

📌 Drupal AI 1.5 要解決什麼問題？目前線索還不夠多

Jev估值100億美元！創始人Diogo Almeida回答一切

📌 【TypeSafe AI 創辦人深度訪談】估值百億美元的 Jev，到底想解決什麼問題？

資訊

From the creator of Redis; run LLM locally with ds4

📌 Redis 作者出手，用 C 語言寫本地 LLM 推理引擎

IBM Brings Bob to Self-Hosted and Air-Gapped Environments: Agentic Software Development Without Moving Your Code

📌 IBM Bob 進駐氣隙網路,程式碼終於不必外送

Agentic AIAI InfrastructureFor DevsLanguage Model

Meta, OpenAI and Uber Just Taught AI Agents to Talk First. What About When to Stay Quiet?

📌 當 AI 主動開口：Meta、OpenAI、Uber 都在賭「說話的時機」

Agentic AIAI AgentsFor DevsOpinion

Meta open sources code to let you make Muse AI gadgets

📌 自己動手做一個「迷你 Muse 吊飾」，Meta 開源裝置程式碼

MetaTech

Meta wants your next gadget to be Muse-infused

📌 Meta 開源 Muse 硬體 SDK，樹莓派也能變 AI 代理終端

Meta

Microsoft AI Releases MAI-Transcribe-2-Streaming: #1 Real-Time Speech-to-Text Model on Artificial Analysis

📌 【Microsoft AI】語音辨識刷新第一，100 毫秒出字

Agentic AIAudio Language ModelEnterprise AIFor Devs

OpenAI safety employee resigns, claiming the company’s ‘culture is broken’

📌 OpenAI 安全報告負責人離職：「公司文化壞了」

SecurityOpenAI

OpenAI安全團隊持續地震！負責人離職，三名員工因洩密被開

📌 OpenAI 安全透明度負責人離職，同時三名員工因洩密被解僱

資訊OpenAI安全團隊

This Week in Open Source for October 2, 2026

📌 本週開源速覽：Gemma 4 推理數學、LiteRT 本地 RAG 一次掃過

Prime Intellect Launches Prime Inference: Serverless and Reserved Serving for Frontier Open Models

📌 Prime Intellect 推出 Prime Inference，服務量先跑贏研發流程

Agentic AIAI InfrastructureDeep LearningFor Devs

A model guide for the GPT-6 family

📌 OpenAI 官方實務指南：新創如何用對 GPT-6

Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap

📌 Anthropic 砸 1 億美元，要訓練 1 萬名 AI 工程師

Chatham scales its capital markets expertise with OpenAI

📌 OpenAI 案例：Chatham Financial 把交易驗證壓到 4 分鐘內

The latest AI news we announced in September 2026

📌 Google 9 月 AI 月報：Gemini 4 Argon 領軍登場

Add secure Web Search to Claude Desktop with Amazon Bedrock AgentCore

📌 幫 Claude Desktop 接上即時網路搜尋，全程留在 AWS 邊界內

Advanced (300)Amazon BedrockAmazon Bedrock AgentCoreTechnical How-to

Academia is for Ambition — Alex Zhang, MIT

📌 MIT學者談RLM：模型外的「哈尼斯」才是關鍵

[AINews] Pi 1.0, Pi Durable, and AIE NYC

📌 Latent Space 最新彙整：Pi 1.0 與 Durable 正式釋出，Gemini 4 Argon、GPT-6.1 Sol 接連登場

AutoSynthData: Generating Training Data for Enterprise Agents

📌 ServiceNow用模型失敗打造企業Agent訓練資料

AWS Strands Labs Releases Strands Decider 2B: An Open Source Decision Model That Picks Options in About 115 ms

📌 AWS 開源 Strands Decider 2B：一次前向傳播、115 毫秒選出答案的決策模型

Agentic AIAI InfrastructureDeep LearningEmbedding Model

Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text

📌 Cloudflare 首款自研模型不是聊天機器人,是會回機率的決策模型

Agentic AIAI AgentsAI InfrastructureComparison

Fine-tune a search agent with multi-turn RL on Amazon SageMaker AI

📌 【AWS 實測】多輪強化學習微調搜尋 agent,失敗率從 22.89% 砍到 0.68%

Advanced (300)Amazon SageMaker AITechnical How-to

How to choose your first Genie Agents for maximum impact

📌 第一個 Genie Agent 選錯，後面全部跟著卡關

AllData StrategyBest Practices

Inside-Out AI: Rebuilding Airbnb Behind the Scenes and Across the Guest Experience

📌 從Meta到Airbnb：CTO揭露AI重塑內外體驗的做法

Lightweight PDF parser with layout, tables, formulas and bounding boxes

📌 不靠ML模型，papero用幾何法解析PDF版面

Limits of Confidence in Diffusion

📌 【Apple ML】離散擴散模型的信心排序，為何從理論上就注定生成錯誤分布

NVIDIA Announces DGX Spark 64GB: A 1-PetaFLOP Grace Blackwell Desktop for Local AI Agents, Fine-Tuning, and Inference

📌 【NVIDIA】64GB 版 DGX Spark 來了，桌上型 1 petaFLOP 跑本地 agent

Agentic AIAI AgentsAI InfrastructureDeep Learning

One month coding with GLM 5.3 Flash

📌 開源模型挑戰：GLM 5.3 Flash用一個月後學到的事

Guides

📌 Simon Willison彙整16章Agentic Coding實務守則

Open-sourcing AstaBrief, the fast report-generation model in Asta

📌 Ai2 開源 AstaBrief：8B 小模型把科研報告生成時間砍去近 3.5 倍

openJiuwen X-Router自演進模型路由技術首發，昇騰親和，Agent越跑越省，實測減少50+%Token消耗

📌 openJiuwen X-Router:讓 Agent 自己學會「這次該找誰」,實測省下逾五成 Token

資訊華為

RIP, vector database

📌 turbopuffer拆掉ANN主索引，重構v3儲存架構

Scaling cloud migrations with agentic AI on Amazon Bedrock AgentCore

📌 300 多個應用程式、一個財政年度期限：AWS 用四個 Agent 把 IaC 開發時間從週級壓到分鐘級

Advanced (300)Amazon Bedrock AgentCoreTechnical How-to

Sweep thousands of leases for compliance using Amazon Quick and the Adjudicated Query pattern

📌 規則引擎加對話介面:用AI稽核上萬份租約

Advanced (300)Amazon Quick SuiteTechnical How-to

Show HN: Audionaut – an open-source cross-platform multitrack audio editor

📌 開源多軌編輯器 Audionaut：讓 Claude 動手剪你的音軌

Barclays scales Claude to upgrade operations and improve client experience

📌 Barclays 擴大導入 Claude，目標 2026 年底半數工程師上線

Claude-shaped science

📌 【Anthropic 客座文章】放棄跟 Claude 爭論，物理學家反而挖到跨領域金礦

Customize Claude Code with mods

📌 Claude Code 推出 mods：用 TypeScript 客製化你的 AI 編碼工具

How Albertsons Companies is reimagining retail from the inside out

📌 【OpenAI 客戶案例】連鎖超市巨頭如何用 ChatGPT 重塑內部工作流

The Den frees up 10-15 hours a week to grow with ChatGPT Work

📌 【OpenAI 客戶案例】社交俱樂部靠 ChatGPT 每週省下 10-15 小時行政時間

The eternal complement

📌 【OpenAI 官方觀點】突破性創意的背後，AI 最該做的其實是苦工

[AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output

📌 【Google DeepMind】Gemini 4 Argon 登場，衝著 Astra、Fable 而來

Amazon releases its own Jev clone as decision models flood the web

📌 Amazon 也做了自己的 Jev：決策模型開源潮正式開打

AWS

Build agent memory with NVIDIA NeMo Agent Toolkit and Amazon S3 Vectors

📌 用 NVIDIA NeMo Agent Toolkit 打造多 Agent 記憶層

Advanced (300)Amazon Simple Storage Service (S3)Technical How-to

Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills

📌 NVIDIA DOCA Agent Skills：讓通用AI代理學會DPU開發

Agentic AI / Generative AIData Center / CloudDeveloper Tools & TechniquesAgent Skill

Build Local AI Apps with C++ and NVIDIA TensorRT RTX Samples

📌 用 C++ 打造本地 AI 應用：NVIDIA TensorRT RTX 範例全解析

Developer Tools & TechniquesC++CUDADGX Spark

Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war

📌 Claude Opus 5.5 對上 GPT-6 Sol/Luna：LLM 價格戰再升級

Building ambient agents with Amazon Bedrock AgentCore: From event-driven signals to human-in-the-loop workflows

📌 AWS AgentCore打造「環境式」代理：事件觸發，人類只在需要時介入

Amazon BedrockAmazon Bedrock AgentCoreIntermediate (200)Technical How-to

Claude’s new auto eval tool

📌 實測 Anthropic 新 eval 工具：好用，但別急著交出判斷權

Cohere Releases Embed 5: How It Compares to Voyage 4 Large, Gemini Embedding 2, and OpenAI

📌 Cohere Embed 5上線：Pro/Fast共享向量空間，對比Voyage、Gemini、OpenAI

Agentic AIEmbedding ModelEnterprise AIFor Devs

Google DeepMind Unveils Gemini 4 Argon with 1M Output Tokens for Coding, Knowledge Work and Cyber Defense

📌 Google DeepMind 發布 Gemini 4 Argon：單次輸出衝上 100 萬 token

Agentic AIAI InfrastructureComparisonEnterprise AI

Google releases Gemini 4 Argon, called its most powerful model yet

📌 【Google 最新模型】Gemini 4 Argon 先攻資安漏洞修補

geminiGoogle

Graph-centric agentic intelligence

📌 Amazon Science：用圖譜讓AI Agent看懂網路的因果關係

How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?

📌 強模型用簡單Harness就夠，複雜編排反而多餘

RLTL;DR: Self-Improvement by Internalizing Self-Generated Feedback

📌 Apple新作RLTL;DR：自產回饋如何打破RLVR瓶頸

Perplexity Releases pplx-embed-v2-context-9b-preview: A Contextual Embedding Model That Retrieves Answers and Their Supporting Evidence

📌 Perplexity 新 embedding 模型：檢索答案的同時，連「證據」一起找出來

Agentic AIAI InfrastructureContext EngineeringEnterprise AI

SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation

📌 Apple 研究：連續學習 Agent 評測，為何每次都要重造排程輪子

Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs

📌 Ai2 Olmo-core 3：開源訓練框架衝向兆參數 MoE

Uplifting conversion across the acquisition funnel with personalization using contextual bandits on AWS

📌 情境式 Bandit 優化獲客漏斗，而非單一轉換率

Amazon SageMaker AICustomer Solutions

Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week

📌 OpenAI DevDay 幕後：Computer Use 為何「180 度大逆轉」

何愷明團隊新作：看貓片就能學會ARC挑戰

📌 何愷明團隊新作：看貓片學會解 ARC 抽象推理

資訊何愷明貓

Claude for Government is now generally available

📌 【Anthropic 官方發布】Claude for Government 正式開放，FedRAMP High 環境全面上線

Disrupting a coordinated model-distillation campaign

📌 【OpenAI 官方公告】攔截一起協同式模型蒸餾攻擊行動

Gemini 4 Argon: our next era of frontier intelligence

📌 【Google DeepMind 官方發布】Gemini 4 Argon 輸出上限衝上百萬 token，先給資安防禦者用

How Anthropic's sales team rebuilt inbound with Claude Managed Agents

📌 Anthropic 業務團隊怎麼用 Claude Agent 重建詢價流程

Helping small businesses put AI to work

📌 OpenAI 攜手 SBDC，把 AI 落地到中小企業

Introducing SynthID Bio

📌 【Google DeepMind 官方發布】幫 AI 設計的蛋白質打上浮水印，功能不受影響

What work can robots do?

📌 【Anthropic 最新研究】機器人能做 74% 的體力活，卻只有 0.3% 比人划算

[AINews] OpenAI DevDay 2026: Dots, 6.1 Sol, Ultrafast, Decisions API, Agents API, Spaces, Marketplace, and 1.2 Billion ChatGPT WAU

📌 OpenAI DevDay 2026：全天候代理 Dots 登場，旗艦模型卻悄悄被砍

Build a multi-agent music production pipeline on Amazon Bedrock AgentCore Runtime Instances

📌 AWS讓多個Agent共用GPU跑好幾天的協作

Advanced (300)Amazon Bedrock AgentCoreTechnical How-to

DeepSeek官方開源昇騰基礎元件，與昇騰共建高效易用的AI晶片軟體生態

📌 DeepSeek開源昇騰元件，V4.1單卡衝上5102 tokens/s

資訊華為

DeepSeek知乎獨家發文，首次公開V4.1 Agent訓練“大本營”DSec

📌 DeepSeek首度公開：百萬沙盒撐起V4.1 Agent訓練

資訊知乎

Did AI Just Solve One of Mathematics’ Biggest Problems?

📌 OpenAI萬人Agent提解答，卻引來一場署名爭議

Forecasting space weather risks on power grids

📌 用 ML 預測太空天氣，守護 6.6 萬座變電站

Research Blog

Google announces Gemini 4 and says it’s so capable that only ‘trusted cyber defenders’ can have it right now

📌 Gemini 4 Argon 發布，卻先鎖給「網路防禦者」

GoogleTech

Introducing ai_decide: make fast decisions on your governed data

📌 Databricks 推 aidecide，捨棄 LLM 做結構化決策

Platform & Products & Announcements

Liquid AI Releases d1: A Decision Model That Returns Calibrated Probabilities With Zero Output Tokens

📌 Liquid AI發布d1：每次呼叫輸出token數永遠是0

Agentic AIAI InfrastructureLanguage ModelLarge Language Model

Jev introduces a new shape of LLM—System One, aka Decision Models

📌 TypeSafe推出Jev：把LLM輸出換成一個機率數字

NVIDIA Researchers Introduce Physis-Lang: Self-Evolving Physical Language That Lifts Cosmos 3 Past Veo 3.1 on Physics Benchmarks

📌 【NVIDIA／MIT／牛津研究】讓影片模型懂物理，靠的是改寫字幕

Agentic AIAI InfrastructureAI Paper SummaryFor Devs

On the Effectiveness-Fluency Trade-Off in LLM Conditioning: A Systematic Study

📌 Apple 研究揭露：讓 LLM 聽話，通常是用流暢度換來的

One Bad Prompt Took Down a Company’s Salesforce: RSA’s Jim Taylor on Agent ID and Taming the 4,000 Shadow AI Agents Hiding in Your Enterprise

📌 一句提示詞，如何讓一家公司的 Salesforce 整套當機

Agentic AIAI AgentsFor DevsHarness

Open source tool distills Jev so you can run it locally

📌 開源工具 Jevstiller，把 Jev 蒸餾到本地跑

Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning

📌 開源 TTS 模型破萬，評測方式卻停在人工投票時代

OpenAI Releases GPT-6.1 Sol: Near-Astra Coding and Computer Use at One-Fifth of Astra’s Token Price

📌 GPT-6.1 Sol：逼近Astra效能，價格剩五分之一

Agentic AIAI AgentsLanguage ModelLarge Language Model

OpenAI光速上新GPT-6.1 Sol！一晚上25項更新，都在這裡了

📌 OpenAI DevDay發表Dots與更便宜的GPT-6.1 Sol

資訊首頁輪播DotsGPT-6.1 Sol