【AI资讯】2026-10-23 20:00 — Grok接入手机日历与联系人，AI助手深入个人数据
本轮新增 73 条 AI 资讯 · 覆盖 15 个来源 · 第26轮
产业与商业 (12条)
Uber brings ChatGPT Work to all employees with OpenAI
来源：OpenAI
How AT&T is rethinking in-house legal work with OpenAI
来源：OpenAI
Anthropic Investing in the public benefit
来源：Anthropic
Meta Prometheus Gigawatt-Scale Supercluster
来源：Meta
Microsoft Maia 300 custom chip
来源：Microsoft
Only 13 percent of companies see real returns from AI, Gartner survey finds
来源：THE DECODER
OpenAI lowers valuation target, raises up to $30 billion
来源：THE DECODER
MiniMax M3.1 Flash shows China’s AI price race has no floor in sight
来源：THE DECODER
科大讯飞首发星辰TaaS平台，让Token真正成为千行百业的生产力
来源：量子位
特斯拉利润腰斩大跌43%，马斯克称FSD在中国也会遥遥领先
来源：量子位
谁在给OpenAI供应具身数据？挤进Physical AI核心供应链
来源：量子位
DeepSeek据悉完成1000亿元新一轮融资；特斯拉三季度利润大跌43%
来源：极客公园
研究与学术 (18条)
Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
来源：Hugging Face
Beyond tree cover: mapping the world’s forest types
来源：DeepMind
Open call for guest investigators in the life sciences
来源：Anthropic
Meta GenRecs Tokenization for Ads
来源：Meta
Meta Thin Orchestrator for AI Agents
来源：Meta
HuggingFace NVIDIA Kumo Relational
来源：Hugging Face
HuggingFace Vocabulary SLM Cost Cut 77%
来源：Hugging Face
Google Agent Compute Engine gives every Gemini agent its own hardware-isolated micro-virtual machine
来源：THE DECODER
OpenAI agents analyze mouse brain data in 32 minutes, a job that would take humans months
来源：THE DECODER
Masakhane AfrIFact Exposes AI Fact-Checking Failures Across 10 African Languages
来源：AlphaSignal
TU Delft FMER Trains Robot Policies 7x Faster on Sparse Tasks
来源：AlphaSignal
Idiap Abstract Compression Halves Audio Tokens Without Losing Conversation Context
来源：AlphaSignal
ETH Zurich MVMAE Beats Larger X-ray Models Using 10x Fewer Labels
来源：AlphaSignal
LeCun押注的世界模型，千诀团队隐变量可辨识性方向比LeCun早一年
来源：量子位
具身记忆赛道持续升温，长记忆迎来新SOTA
来源：雷锋网
Direct Pass@k Optimization for Thought Pruning in Test-Time Reasoning
来源：arXiv
EpiSEM-Bench: Evaluating Safety and Evidence Recovery in LLM Agents under Missing Evidence
来源：arXiv
Beyond Causal Masks: Auditing Hidden Future-Token Leakage in Adaptive Sequence Models
来源：arXiv
工具与产品 (30条)
NVIDIA Unveils Minerva to Bring End-to-End Observability for AI Factories
来源：NVIDIA
Build more responsive apps with the Decisions API
来源：OpenAI
Mistral Voxtral 3
来源：Mistral
Mistral Timestral
来源：Mistral
Anthropic Claude Opus 5.5
来源：Anthropic
Anthropic Claude Sonnet 5.5
来源：Anthropic
Anthropic Claude Haiku 5.5
来源：Anthropic
Anthropic Donating Agent Skills to Agentic AI Foundation
来源：Anthropic
Gemini 4 Argon for Robotics
来源：DeepMind
HuggingFace OlmoEarth v1.3
来源：Hugging Face
Canva launches Craft, its first in-house image model
来源：THE DECODER
Anthropic Claude writes code and proves it safe automatically
来源：THE DECODER
Recursive shows first result with Flint, 40B MoE local coding model
来源：THE DECODER
Anthropic takes on Grok Bot, Muse and Dots with always-on agent Claude Duet
来源：THE DECODER
xAI’s Grok Reaches Into Your Phone’s Calendar and Contacts
来源：AlphaSignal
Anthropic Claude Proves 239 Open-Source Libraries Crash-Free
来源：AlphaSignal
Anthropic Brings Claude Directly to Google Cloud With Same-Day Feature Access
来源：AlphaSignal
Alibaba Z-Image Turbo Tops Open-Weight Rankings on a 16GB GPU
来源：AlphaSignal
Alibaba Wan2.2 Animate Lets You Swap Characters Into Any Video
来源：AlphaSignal
Mistral Timestral Beats Larger Rivals at Zero-Shot Forecasting on CPUs
来源：AlphaSignal
Alibaba Qwen2.5-3B Shrinks to 2.5GB so Anyone Can Run AI Locally
来源：AlphaSignal
国产5B生图模型海外爆火，4天超百万下载
来源：量子位
企业AI需要两个世界模型，特赞让企业Agent真正懂业务
来源：量子位
一张5090跑起34B实时世界模型，国产玩家拿下英伟达大奖
来源：量子位
高德发布首个基于开放世界行为的全身基座模型λ0，推动机器人动作对齐
来源：量子位
蚂蚁数科发布决策模型LingDT-Decision-1.0，准确率超微软和Jev
来源：量子位
Kimi K3.1泄露：3D成本仅有GPT-6 Astra的1/14
来源：雷锋网
Anthropic被迫下场，推出个人AI助理Claude Duet
来源：极客公园
Adaptive Skill Granularity for Coding Agents
来源：arXiv
DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels
来源：arXiv
政策与安全 (8条)
OpenAI and Poland’s NASK partner to advance AI safety and security
来源：QuickOnlineTips | OpenAI
Broad access through consistent standards
来源：OpenAI
Anthropic Expanding Cyber Verification Program
来源：Anthropic
Gemini 4 Argon safety
来源：DeepMind
Major US banks are locking out AI agents that log in with customer credentials
来源：THE DECODER
AI agent in Skynet mode hijacks unprotected language models for self-sustaining cyberattacks
来源：THE DECODER
METR and Redwood Found Six Ways to Blind OpenAI’s Coding Agent Monitor
来源：AlphaSignal
Certified Execution Harness for Agentic AI: Petri-Net Supervisory Control
来源：arXiv
观点与评论 (5条)
What FDE means at OpenAI
来源：OpenAI
复旦贾萧松：VLA、WAM新架构层出不穷，但真正的瓶颈不在模型架构
来源：雷锋网
宇树10周年，王兴兴：我劝你相信AI
来源：极客公园
大模型内卷之后，AI竞争进入搜索基础设施时代
来源：极客公园
Who Is Sycophantic to Whom? Confident LLMs and the Collapse of Trust Calibration
来源：arXiv