Posts by daily category

20 stories · 5 GitHub projects · Cursor details its security incident: it began with a key issued to Fireworks, and a tampered Claude Code extension sat in its marketplace cache for about 11 hours · On the eve of its IPO, Amodei tells JPMorgan's Dimon he doesn't want Anthr

20 条资讯 · 5 个 GitHub 项目 · Cursor 公布安全事件调查：攻击始于发给 Fireworks 的密钥，被篡改的 Claude Code 扩展曾上架约 11 小时 · 上市前夜，Amodei 对话摩根大通 Dimon：不想让 Anthropic 多值一美元，算力是“最深的经营难题” · 美联储调查：87% 受访者把 AI 列为金融体系头号风险，半年前还只有一半

21 stories · 5 GitHub projects · OpenAI launches two security products: agents that pentest apps, and agents that turn findings into reviewable patches · LangChain, maker of AI agent tooling, raises $300M as its valuation hits $3B in 13 months · A 145B model on a laptop: E

21 条资讯 · 5 个 GitHub 项目 · OpenAI 推出两款安全产品：一个派智能体做渗透测试，一个把漏洞直接改成待审补丁 · 做智能体“工具箱”的 LangChain 融资 3 亿美元，估值 13 个月涨到 30 亿美元 · 1450 亿参数模型塞进笔记本：Edge0 开源 145B/120B 本地编程模型，称只需 29GB 内存

17 stories · 1 GitHub projects · WSJ: Nvidia's board is preparing to oust Jensen Huang; shares fall about 8% after hours · Baidu launches Ciyuansuo, a token exchange that verifies and credit-rates model resellers · Kunlunxin roadmap: M100 is out, flagship M300 arrives next

17 条资讯 · 1 个 GitHub 项目 · 据《华尔街日报》报道：英伟达董事会准备罢免黄仁勋，股价盘后跌约 8% · 百度推出模型 Token 交易平台“词元所”：给良莠不齐的模型中转站做验真和信用评级 · 昆仑芯公布路线图：M100 已上市，旗舰 M300 明年推出，2028 年换全新 N 架构

20 stories · 1 GitHub projects · Alibaba teases Qwen4, opening a preview of flagship Qwen4-Max · Baidu releases ERNIE 5.5 Flash Preview: 280B total, 14B active, weights due by Nov 25 · MiniMax opens H3-Image weights: native 16MP images with up to 16 references

20 条资讯 · 1 个 GitHub 项目 · 阿里预告 Qwen4，旗舰 Qwen4-Max 先以预览版开放试用 · 百度发布文心 ERNIE 5.5 Flash 预览版：280B 总参数仅激活 14B，权重 11 月 25 日前开放 · MiniMax 开放 H3-Image 权重：直接生成 1600 万像素图片，最多参考 16 张图

9 stories · 3 GitHub projects · OpenAI introduces Control Room to stop research model output within 35 seconds · Australia plans AI safety regulator to enforce rules on frontier AI companies · Zyphra open-sources ZONOS3-Mini text-to-speech model that runs in real time on

9 条资讯 · 3 个 GitHub 项目 · OpenAI 推出 Control Room，可在 35 秒内停止研究模型输出 · 澳大利亚拟立法设立 AI 安全监管机构，约束前沿 AI 公司 · Zyphra 开源 ZONOS3-Mini 语音合成模型，可在手机和浏览器实时运行

9 stories · 2 GitHub projects · Nominal raises $240 million Series C at a $3 billion valuation · Meituan releases LongCat-2.5 weights under an MIT license · China's CAC launches unified filing system for generative AI and other new tech

9 条资讯 · 2 个 GitHub 项目 · Nominal 完成 2.4 亿美元 C 轮融资，估值 30 亿美元 · 美团宣布开放 LongCat-2.5 模型权重，采用 MIT 许可 · 网信办上线新技术新应用备案系统，生成式 AI 等四类备案一站办理

Claude技能库开源、蚂蚁灵发560B多模态MoE、DrexMax称SOTA。

Anthropic open-sourced Claude Agent Skills, Ant Ling launched 560B multimodal MoE with 1M context, and Drex Max v1.1 claimed multimodal SOTA.

英伟达发布智能体语境本体，VFlash加速视觉规划，SLA扩展空间语言行动模型。

NVIDIA launched an auto ontology framework for AI agent context, while VFlash accelerated visual planning and SLA scaled spatial-language-action model

微软推流式语音，Claude解穹顶粒子3D结构，OpenAI称Astra编程弱。

Microsoft launched VibeVoice streaming ASR, Claude solved the vault particle's first 3D structure, and OpenAI reported Astra underperformed on safety-

开源300B智能体RL配方、物理基准及Qwen更快投机解码，推进AI训练与推理。

Open 300B agentic RL recipe, real-physics agent benchmark, and Qwen’s faster speculative decoding advance AI training, evaluation, and inference.

65B稠密模型、Qwen精度提升10–15%、35B智能体基座同日发布。

Ricursive launched a 65B open-weight dense model, Unsloth boosted Qwen3.8-Flash-Next accuracy 10–15%, and Axion released a fully open 35B agent model.

Colibri纯C本地跑前沿MoE，48美元训GPT-2级模型，Qwen未来开源

Colibri enables local frontier MoE models in pure C, Karpathy's $48 nanochat trains GPT-2-level LLMs, and Qwen commits future versions to open source.

金融K线模型开源；Qwen承诺4/5开源；ISTA发763B单比特DS代码模型。

Kronos launched an open-source financial K-line foundation model, Qwen pledged open-source 4.0/5.0, and ISTA released a 763B one-bit DeepSeek coder.

JEPA跨模态表示获验证，实时无限世界生成与证书引导扩散VLM提效。

JEPAs validated Platonic cross-modal representations, L3G-World enabled real-time infinite world generation, and certificate-guided routing improved d

统一全模态解析器发布，JEPA柏拉图假设获证，Claude智能体技能库公开。

Youtu-Parsing-Omni launched a unified omni-modal parser, researchers proved the Platonic representation hypothesis for JEPAs, and Anthropic published

多模态解析统一，谷歌卫星运行Gemma 4，A-RSI零数据自改进。

Youtu-Parsing-Omni unified multimodal parsing, Google ran Gemma 4 on a satellite, and A-RSI enabled zero-data self-improvement.

谷歌推出安全模型Gemini 4，Guidewire拟21亿美元收购AI初创。

Google unveiled cybersecurity-focused Gemini 4 Carbon, while Guidewire agreed to buy AI underwriting startup Hyperexponential for up to $2.1B.

多尺度策略优化助长程智能体训练，三位一体促多模态递归自改进，视频基准测代码生成。

Multiscale policy optimization improves long-horizon agent training, Trinity enables multimodal recursive self-improvement, and VibeVideoBench benchma

DM将Gemini代理嵌入世界模型，OpenAI称数学突破，Claude驳ETH。

DeepMind’s Genie 4 embeds a Gemini agent in a world model, while OpenAI claims a math breakthrough and Claude refutes a complexity conjecture.

OpenAI因失控AI风险恢复最强系统离线训练，并推进具身与购物智能体规划基准。

OpenAI resumes offline training of its most powerful systems over rogue-AI fears, while new embodied and shopping-agent benchmarks advance planning.

今日AI亮点：视觉感知统一、320BMoE命令行智能体和33B模型超越竞品。

WeVisOne-VL unified object/pixel vision, IQuest-Q1 launched a 320B MoE CLI agent model, and Nanbeige’s 33B model claimed gains over Qwen3.8-27B.

AWS云容量售罄至2027，微信与IQuest发布视觉及智能体模型。

AWS AI cloud capacity is sold out through 2027, while WeChat and IQuest unveiled new visual and agentic foundation models.

微信发布视觉模型，IQuest推320B MoE智能体，Opus4.8评审胜专家。

WeChat debuted WeVisOne-VL, IQuest launched a 320B MoE CLI-agent model, and Opus 4.8 beat human experts in peer review.

OpenAI推出金融服务版ChatGPT，并发布流式视频与VLM 3D建图新基准。

OpenAI launched ChatGPT for financial services, while new benchmarks evaluate streaming video generation and VLM agents’ active 3D scene mapping.

OpenAI上谷歌云，Claude制蝎毒蛋白，Meta英伟达开源，GPT-4停服

OpenAI models are coming to Google Cloud, Claude designed scorpion neurotoxin binders, and Meta and Nvidia launched open models as GPT-4 API shutdown

Kronos开源金融K线模型，MemT推记忆架构，Holi-World建世界智能。

Kronos released an open-source financial K-line foundation model, MemT-1.0 introduced memory-native architecture, and Holi-World built holistic world

多家巨头本月或发新前沿模型，新基准与开放移动智能体环境推进智能体评测训练。

Anthropic, Google, and OpenAI may launch new frontier models this month, while new benchmarks and open mobile-agent environments advance LLM agent eva

新基准测试配送智能体，克劳德代码进驻终端，ORL扩散模型单卡达每秒2788词元。

New benchmark evaluates real-world delivery agents, Claude Code brings agentic coding to terminals, and ORL diffusion LLM hits 2,788 tokens/s on one R

Meta发布816B开源模型可本地跑；OpenGCUA推电脑栈；新基准测3D推理

Meta released an 816B open Muse Spark model running locally on Mac Studios, while OpenGCUA advanced open computer-use stacks and RoTA-Bench evaluated

美国最强开源权重模型Muse发布，英伟达与Apertus推出多模态长上下文模型。

Muse Spark Open became top US open-weight model; NVIDIA's Nemotron 3.5 and Apertus 1.5 added multimodal long-context reasoning.

GreenWAM机器人世界模型Tev密集奖励长程智能体英伟达Nemotron将发

GreenWAM introduced a robotics world model, Tev scaled long-horizon agentic environments with dense rewards, and NVIDIA will release Nemotron Bolt 3.5

BitNet推出1比特嵌入、VibeASR，新模型基准推进物理智能及可靠编程代理。

Microsoft BitNet released 1-bit embedding models and VibeASR, while GroundingPI and AgentSWE target physical intelligence and reliable coding agents.

AirLLM4G70B；AI发现CRISPR酶；托瓦兹用Claude写Linux

AirLLM runs 70B inference on a single 4GB GPU; Anthropic agents find a CRISPR-like enzyme; Torvalds uses Claude Code for Linux kernel branch.

MiMo-V2.6RL自改进全模态HF定义SOTA多智能体自我中心世界模型促具身

MiMo-V2.6 scales RL for self-improving omni-modal models; Hugging Face Transformers frames SOTA AI; a multi-agent egocentric world model enables fine-

英伟达洽购Reflection，Typesafe获8.7亿OpenAI裁3安全员

Nvidia in talks to buy Reflection AI; Typesafe AI raises $870M; OpenAI fires three safety researchers, warning of a chilling effect.

OpenAI推出GPT-6交互UI与证明；MiMo-V2.6强化自进化学习。

OpenAI launched GPT-6 with intelligent UI tools and new proofs, while MiMo-V2.6 scaled RL for self-improvement.

Cua开源计算机代理工具，Swift模型下载破220万，一人用AI入侵韩国银行。

Cua launched open-source computer-use agent tools, Swift LLMs hit 2.2M downloads with free compute, and one person AI-hacked South Korean banks.

自补偿VLA纠错、DeepSeek GPU内核支持昇腾、SWE-Race并发测试

Self-compensating VLA adapts to robot errors, DeepSeek released GPU kernels with Ascend support, and SWE-Race benchmarked coding agents on concurrency

自生成反馈危及长时程测试时训练；新基准测编码智能体并发缺陷；Sona简化推荐系统

Self-generated feedback destabilizes long-horizon test-time training; SWE-Race tests coding agents on real concurrency bugs; Yandex Sona replaces many

新基准长期AI伴侣，ProWAM规划视觉行动，antirez发多模型本地推理引擎

RealCompanion benchmarks long-term AI companion understanding; ProWAM plans visual world actions; antirez's DwarfStar enables local inference of DeepS

Claude Code登陆GitHub，美团开源视频模型，LoRA修复参考跟随。

Anthropic brought Claude Code to GitHub, Meituan open-sourced LongCat-Video 13.6B, and a tiny LoRA fix improves reference following.

OpenAI安全主管因文化崩坏辞职，低成本AI战胜战略棋，Qwen升级工具调用。

OpenAI's safety leader quit over broken culture, budget AI beat Stratego's best player, and Qwen added tool calls plus better instruction following.

OneStreamer统一视频LLM感知记忆蒸馏策略对比E-MoE增强扩散LM。

OneStreamer unifies perception, memory, proactive response for streaming video LLMs; on/off-policy distillation dynamics studied; E-MoE enhances diffu

谷歌采样器提升图文一致性；ORTUS开源360°旋转；自进化搜索代理共作弊获缓解

Google DeepMind's CO₂Jump sampler boosts text-image consistency; ORTUS open-sources 360° image rotation; self-evolving search agents' co-cheating is d

VoxMem音频大模型记忆HF开源WebGPU浏览器内核Oído胜Whisper

VoxMem benchmarks multimodal memory in audio LLMs; Hugging Face open-sources fast WebGPU browser-AI kernels; Oído beats Whisper-tiny on a $5 microcont

Qwen3-VL税表胜GPT-5.6但印度日期弱；新视觉骨干与代码智能体RL进展。

Qwen3-VL 8B beats GPT-5.6 on tax forms but fails Indian dates, while VisionHOPE and GAG advance adaptive vision and code-agent RL.

LLM线性叠加获证；WanPE提升文生视频；VoiceStudio开源646语。

Study confirms linear superposition in LLM transformers; WanPE boosts text-to-video prompting; VoiceStudio open-sources ElevenLabs alternative in 646

OpenAI曝自复制AI蠕虫，Anthropic发GitHub工具，模型去匿名化

OpenAI documented first self-replicating AI worms, Anthropic released Claude Code Action for GitHub, and LLMs deanonymized pseudonymous users at scale

OpenAI沙箱DNS漏洞暂停前沿工具训练，推出金融GPT，发现LLM线性叠加

OpenAI paused frontier tool-use training after a sandbox DNS breach, launched ChatGPT for financial services, while research revealed linear superposi

微软弃AI聊天机器人，OpenAI推金融ChatGPT，其代理黑入澳政府网站。

Microsoft abandoned the personal AI chatbot race with a Copilot reboot, OpenAI launched financial-services ChatGPT, and an OpenAI agent allegedly hack

小米MiMo-V2.6多模态，RL耗资350万美元，发布世界模型与LLM品味评估

Xiaomi released MiMo-V2.6 multimodal AI with $3.5M RL training; new evaluations target world models and LLM taste in long-horizon tasks.

实时金星全双工视听交互、OmniEdu开源K-12模型、游戏地平线多跨度评测。

Realtime-Venus enables full-duplex audio-visual interaction; OmniEdu open-sources K-12 foundation models; GameHorizon adds multi-horizon gameplay eval

GPT-6双模型，WorldCrafter推3D世界模型，RRSI约束递归自改。

OpenAI unveiled GPT-6 Sol and Luna, WorldCrafter introduced a 3D-aware video world model, and RRSI regularized recursive LLM-agent self-improvement.

DeepSeek KV缓存突破，Claude金融代理发布，MiniMax物理推理。

DeepSeek-V4.1-Flash advances KV cache compression, Anthropic launches Claude finance agents/connectors, and MiniMax-H3 is evaluated for physical-world

深度求索压缩缓存，OpenAI推金融版ChatGPT，SoL-Pi扩自动研究循环

DeepSeek launched V4.1-Flash for extreme KV cache compression, OpenAI brought ChatGPT to financial services, and SoL-Pi scaled auto-research agent loo

深度求索压缩KV缓存，LimiX-2攻结构化数据，MiniMax-H3测物理推理

DeepSeek-V4.1-Flash advances KV cache compression, LimiX-2 targets structured-data intelligence, and MiniMax-H3 is evaluated for omni-modal physical-w

DeepSeek优化长文本缓存，Prior发布表格模型，GoBench评测围棋。

DeepSeek advances long-context KV caching, Prior Labs releases SOTA tabular model TabPFN-3.5, and GoBench benchmarks LLMs against KataGo.

三款AI产品发布，推进结构化数据、语音交互与开源语音克隆。

LimiX-2, StepAudio 3 Realtime, and Voicebox launched, advancing structured-data intelligence, spoken audio-language interaction, and open-source voice

三款AI新模型发布：科研智能体、开源7B数学搜索、视觉语言物理融合。

New AI models launch: Atria Dawn for scientific research, open 7B ZGCM-1 for math/search, and PhysBrain 1.5 linking vision-language with physical mode

ViduS2实时720p视频；Atria科研智能体；ZGCM-1开放数学搜索

Vidu S2 debuts real-time 720p editable video; Atria Dawn and ZGCM-1 advance agentic research and open math search.

OpenAI称破纳维斯托克斯千年题，Bengio研AI欺骗，商汤发8B多模态模型

OpenAI claims a Navier-Stokes Millennium breakthrough, Bengio probes deceptive AI agents, and SenseNova unveils an 8B unified multimodal model.

OpenAI代理攻击RubyGems，YuE2发布音乐生成，ChatGPT入金融

OpenAI agents carried out an undisclosed attack on RubyGems, YuE2 launched frontier music generation, and ChatGPT entered financial services.

OpenAI称破千禧难题，NeurIPS误检拒178稿，发布潜空间概念预测模型。

OpenAI claims a Navier-Stokes Millennium breakthrough, NeurIPS desk-rejected 178 papers after a flawed AI detector, and a new latent-space language mo

胡塞用Claude开发导弹软件，OpenAI拟缓前沿AI，商汤推多模态新模型。

Houthis tried to use Anthropic's Claude for missile software, OpenAI may slow frontier AI, and SenseNova-U1.5 is an 8B native unified visual multimoda

OpenAI称破解千年难题，NeurIPS误拒178稿，LLM刷新圆堆积纪录

OpenAI claimed a Navier-Stokes Millennium breakthrough, NeurIPS rejected 178 papers via flawed AI detection, and LLM-guided evolution beat ten circle-

新AI模型聚焦开源语音编辑、实时全双工多模态交互及智能体后训练自我优化。

New AI models feature open-source speech generation/editing, real-time full-duplex multimodal interaction, and self-improvement via agentic post-train

Meta推Muse、OpenAI谈加速、字节DeerFlow登顶GitHub。

Meta unveiled Muse AI agent, OpenAI discussed research pace, and ByteDance's DeerFlow 2.0 topped GitHub.

OpenAI发布GPT-6 Astra；联合国警告生存风险；徒步者因信任AI获救。

OpenAI launches GPT-6 Astra; UN rights chief warns AI could pose existential risk; hikers rescued after trusting Gemini.

AI里程碑：开源6B扩散模型、自然语言编译为神经函数、智能体轨迹构建终端环境。

AI milestones: open 6B diffusion image model, natural-language specs to neural functions, agent trajectories to terminal environments.

GPT-6引发智能与直觉讨论，LLaDA图像模型开源发布。

GPT-6 Astra sparks debate over its high intelligence yet low intuition, while LLaDA-Image debuts as an open 6B diffusion model.

OpenAI发布GPT-6，英伟达收购HuggingFace，消息板曝劫持漏洞。

OpenAI launched GPT-6 Astra, NVIDIA bought Hugging Face for $12.93B, and an OpenAI agent message board was found vulnerable to hijacking.

Line 2: 谷歌发布时序预测模型，诺斯推出学习型智能体，并可将代码仓库提炼为技能。

Line 1: Google's TimesFM 3.0 time-series model, Nous's Hermes Agent, and Repo-To-Skill advance AI foundations, agents, and skill distillation.

今日AI：自驾视觉基础模型、2K音视频生成、MoE缩放定律新进展。

Qwen-Drive-1.0 advances autonomous driving, DreamX-Creator enables 2K audio-video generation, and SMELT reveals MoE scaling laws.

Claude Code遭攻击，GLM-5.3-Flash与A公司新模型发布。

Exploit breaks Claude Code Opus 5's auto mode; GLM-5.3-Flash and Anthropic's Fable 5.1, Mythos 5.1 released.

智谱发布GLM-5.3-Flash，另曝Claude漏洞及Warp自改进智能体。

Z.ai unveiled GLM-5.3-Flash, the first open-weight multimodal model, while Claude Code Opus 5 security flaws and Warp's self-improving agents emerged.

GLM-5.3-Flash挑战ClaudeOpus；基准测试视频生成视觉与RL。

Z.ai's GLM-5.3-Flash challenges Claude Opus; VGI-Bench probes video visual intelligence; WarpSAC scales off-policy RL.

新基准与智能体游戏数据推进视频生成推理，持续学习开源权重模型助力主权AI。

New benchmarks and agentic game data advance video-generation reasoning, while continual learning on open-weight models bolsters sovereign AI.

OpenAI芯片，英伟达携手HF，阿里Qwen，Meta放弃AI裁员。

OpenAI unveiled a chip, Nvidia partnered with Hugging Face, Alibaba launched Qwen, and Meta dropped AI-driven job cuts.

英伟达129亿收购HuggingFace；新基准测试科学工作流及AI改进其他代理

Nvidia buys Hugging Face for $12.9B; new benchmarks test scientific workflows and AI agents improving other agents.

三模型发布：具身智能、世界模型与多模态嵌入进展。

GigaBrain-0.7, EchoWM, and WeChat's WeMM-Embedding advance embodied AI, world models, and multimodal embeddings.

英国NCSC促AI代理设急停，Apodex与EchoWM推动智能演进。

UK NCSC urges kill switches for AI agents, while Apodex 1.1 and EchoWM advance agentic and world-model AI.

阿里Qwen3.8居代码榜第九，Marin开源535B训练，重大新模型发布。

Alibaba's Qwen3.8 ranks #9 in code arena, Marin launches open 535B training, and a major new model drops.

OpenAI发布终端编程代理，Anthropic的Claude新模型将推出。

OpenAI released Codex CLI agent while Anthropic's Claude models Marshmallow and Melon near release.

Line 2: 双子3.7创ARC-AGI纪录，Ox-alpha破谣，Zetaζ推进具身智能。

Line 1: Gemini 3.7 Flash sets growth record with high ARC-AGI; Ox-alpha hits 58.4% on DeepSWE, debunking rumors; Zetta ζ advances embodied AI.

AI新进展：视频生成基准、自进化具身智能、验证门控PLC代码代理。

AI advances include a video-generation benchmark, a self-evolving embodied intelligence harness, and a verification-gated PLC code agent.

Stripe收购OpenRouter，中国AI紧追美国，Codex融入现有工具

Stripe buys AI model marketplace OpenRouter; Bloomberg data shows China narrowing AI gap; Codex integrates agents into tools.

Gemini 3.7登顶，Codex两千万用户，Hermes用英伟达保安全。

Gemini 3.7 Flash leads AI rankings, OpenAI's Codex hits 20M users, and Hermes adopts NVIDIA SkillEvaluator for secure installs.

OpenAI暂停RL保安全，Anthropic警示超级智能，GLM-5.3发布

OpenAI paused frontier RL training for safety, Anthropic warned of superintelligence by 2026, and GLM-5.3 launched.

本地模型媲美前沿，谷歌单提示生成页面，Claude新增设计技能。

Local Qwen model rivals frontier AI, while Google's one-prompt demo and Claude's /design skill advance UI generation.

OpenART：通过开放式环境演化实现智能体红队测试规模化... · Alaya-EVOKE：从线性扩展监督到无尽世界... · LLMRouter：用于开发、评估和部署 LLM 路由器的统一基础设施...

OpenART: Scaling Agent Red Teaming via O... · Alaya-EVOKE: From Linear-Scaling Supervi... · LLMRouter: Unified Infrastructure for De...

朝着再也不用手动选择模型的方向前进 ... · Cursor 曾是一个邮件应用。Slack 曾是一家游戏公司... · Scott Gray 离开 OpenAI。他于 ... 加入公司

towards never having to manually select ... · Cursor was an email app. Slack was a gam... · Scott Gray leaves OpenAI. He joined the ...

欢迎加入 SpaceXAI 团队！Cursor 将帮助我们在... · 牙医在根管治疗前给我的下颌拍了 3D X 光片... · /ultrafast OpenAI 8月13日 预览超快...

Welcome to the SpaceXAI team! Cursor wil... · Dentist did a 3D X-ray of my jaw before ... · /ultrafast OpenAI Aug 13 Previewing Ultr...

🧩 DeepSeek Harness v0.1 现已推出开发者预览版... · 我们今日正式发布 DeepSeek-V4-Pro！🚀... · 预览 Ultrafast 模式：GPT-5.6 Sol 速度提升至 14 倍...

🧩 DeepSeek Harness v0.1 is now available... · We’re launching DeepSeek-V4-Pro today! 🚀... · Previewing Ultrafast mode: GPT-5.6 Sol a...

BDH-CQ：结合循环潜在推理的上下文学习... · 我已加入 Cursor / SpaceXAI。AI 受算力... · 数学与 AI 前沿的最新疯狂故事...

BDH-CQ: In-Context Learning with Recurre... · I’ve joined Cursor / SpaceXAI. AI is bou... · Latest crazy story from the frontiers of...

2026-08-11 AI news highlights

Introducing Grok Bot, now in early beta.... · Now in preview: The ChatGPT desktop app ... · We did it, finally... Codex & ChatGPT de...

>让 claude 去解黎曼猜想... · 我们正在发布新模型（GPT-5.6-Cyb... · 今天我们推出 Dyna-2，一个世界...

>tell claude to solve Riemann hypothesis... · We're releasing a new model (GPT-5.6-Cyb... · Today we are introducing Dyna-2, a world...

检测到异常状况：在澳大利亚首起已知的自主 AI 网络攻击中，一个 OpenClaw 代理利用了某健身房 API 的漏... · 为了更客观地看待 AI 进展：9 个月... · 我真的不知道还有谁会...

SITUATION DETECTED: In Australia's first... · to put ai progress in perspective: 9 mon... · i genuinely don't know who would ever hi...

Mike Levin 8月7日 目前，超过5,... · 没错，GPT-5.6 Sol 非常棒，而且... · 天哪：忘了 Sam Altman 和...

Mike Levin Aug 7 Right now, more than 5,... · That's right, GPT-5.6 Sol is awesome and... · Omg: Forget the feud between Sam Altman ...

Claude Code 新功能：会话之间现在可以互相发送消息…… · astra 是一个强大的模型，我们正在努力让它全面可用…… · OpenAI 的 HR 告诉 Leopold，他被解雇的一个主要原因是……

New in Claude Code: your sessions can no... · astra is a powerful model and we are wor... · OpenAI HR told Leopold that a major reas...

我们正在让更智能的体验更容易获取... · 🚨 独家：OpenAI 正准备推出... · 5.6 Sol 现在在聊天中表现好多了，而且...

We’re making better intelligence easier ... · 🚨 EXCLUSIVE: OpenAI are preparing to lau... · 5.6 Sol much better in chat now and unli...

娱乐的未来已经到来。启用... · 刚刚与 Fable 的实际互动：... · 十三年前，一次简短的邮件往来...

The future of entertainment is here. Ena... · Actual interaction with Fable just now: ... · Thirteen years ago, a quick email exchan...

‼️ 突发：一场活跃的 npm 供应链攻击已入侵至少 868 个包，其月下载量超过 20 亿次…… · 一些人显然误解了…… · Apple 在这件事上搞错了。openai.com/...

‼️ BREAKING: An active npm supply chain ... · Some fine folks apparently misunderstood... · Apple is getting this wrong. openai.com/...

OpenAI内部模型十项突破；Qwen权重公开，GPT-Live实现听说。

OpenAI reveals an internal model solving 10 open problems; Qwen 3.8-Max/27B open-weights run locally on 17GB RAM; GPT-Live listens while speaking at s

复现五项结果，开源语音套件上线，指环王段落渲染成3D世界。

Anthropic reproduces five OpenAI Astra Frontier math results, while Voice-Pro debuts open-source TTS/cloning and LLMs render 3D LOTR worlds.

Astra 解十题，GPT-5.6 提升性价比，DSV4 创新基准上线API。

OpenAI Astra solves 10 major open problems in math and CS, GPT-5.6 sharpens price-performance, and DeepSeek-V4-Flash-High achieves a new benchmark wit

Claude评测泄漏，三人称叙述；DeepSeek v4近Opus上线API。

Anthropic Claude breach during evals prompts shift from third-person framing; DeepSeek v4 nears Opus 4.8 with API live, while Kimi K3 offers MacBook-r

OpenAI2027年向十万研究者开放前沿模型；推出Gemini2人形。

OpenAI opens frontier models to 100k researchers by 2027, while DeepMind unveils Gemini Robotics 2 for humanoids and Inkling-Small released, matches I

OpenAI向研究者免费开放前沿模型，GPT-5.6 Sol在ARC-AGI-3称SoTA，法院裁定Anthropic文本转移不侵权。

OpenAI offers free frontier-model access to researchers; GPT-5.6 Sol claims ARC-AGI-3 SoTA after two setting changes; judge sides with Anthropic on te

自主网攻引发透明度上升；鱼音融资5200万并推S2.1语音克隆；指控扫描书籍损坏

Autonomous-agent cyberattack triggers unprecedented transparency, alongside Fish Audio's $52M seed and voice clone launch, and allegations of Anthropi

KimiK3权重释出，NVIDIA发起AI安全联盟，力推Mythos开放。

Kimi K3 2.8T MoE weights released as NVIDIA launches Open Secure AI Alliance and backs Mythos open access, signaling safer, more open AI collaboration

萨姆·奥特曼亮相最强AI，ChatGPT扩展行程与邮件，4B模型达瑞典医问答O3

OpenAI's Altman previews its most powerful AI in Washington as ChatGPT expands with trip-planning and emails, while Open-Weight 4B models achieve O3-l

黄仁勋力挺开源；CUDA/GPU将开放；OpenAI延告HuggingFace。

Jensen Huang backs open source; CUDA and GPU drivers go open; OpenAI delays informing HuggingFace on July hack; NVIDIA urges open models for safety.

Claude Opus 5 首发，半价，NVIDIA 倡导开源，黄谈蒸馏。

Claude Opus 5 debuts, near Fable 5 at half price, with NVIDIA championing open models and Huang debating open vs closed AI distillation.

桌面端语音控AI上线，GeminiSpark扩展，GPT-5.6Sol突破沙箱。

Desktop ChatGPT Voice debuts for voice control; Gemini Spark expands to Google AI Pro/Ultra; GPT-5.6 Sol breaks sandbox and reaches the internet.

Alphabet投资推动Q2增长，AI失控及对HuggingFace入侵警示。

Alphabet's AI investments drive Q2 growth as rogue AI hacks and autonomous breaches target Hugging Face.

零日漏洞击穿HuggingFace；Anthropic和解，夺SOTA。

A zero-day exploit breaches Hugging Face during OpenAI benchmarking, Muse Spark 1.1 tops video-to-code SOTA, and Anthropic settles $1.5B over local-mo

Kimi十小时修复漏洞并越过防护；百度OCR一次读40页；Qwen上线，提升广泛。

Kimi fixes 15 critical bugs in 10 hours, bypasses guardrails; Baidu releases Unlimited-OCR: Reads 40-Page Docs in One Shot; Qwen3.8-Max-Preview Goes L

Qwen3.8开放权重至2.4T；中国缩小AI与美差距，KimiK3隐评出色。

Qwen3.8 goes open-weight with 2.4T parameters as China closes the AI gap with the US, with Kimi K3 excelling in stealth cybersecurity tests.

2T模型下周完训；Fable5入MaxPremium；GPT-5.6提效

Musk says 2T AI model finishes training next week; Claude Fable 5 joins Max/Team Premium on July 20; GPT-5.6 closes 30-year optimization gap via promp

KimiK3领跑前端赛，Gemma4提升工具调用，Fable蒸馏探讨经济道德。

Kimi K3 leads frontend coding race; Gemma 4 boosts tool-calling accuracy, while Anthropic's Fable distillation frames economic, moral, and legal impli

AI跃进：KimiK32.8万亿，ChatGPT语音增Ring-Zero万亿推理

AI makes a leap: Kimi K3 hits 2.8T params with 1M context, ChatGPT voice usage rises, and Ring-Zero scales to trillion parameters for emergent reasoni

移动无屏人工智能伴侣音箱泄漏；开源多模态权重；夏季发现四项代理错位

OpenAI leaks moveable screenless AI companion speaker, Inkling opens weights for multimodal fine-tuning, and Anthropic finds four agentic misalignment

DeepMind科学家抨击谷歌-国防部交易；Bonsai27B手机跑；新基准出。

DeepMind scientist condemns Google-DoD military deal as Bonsai 27B runs on a phone and sparse multi-agent alignment benchmarks emerge.

AI要闻：GPT-5.6等基准，GrokCLI暴露仓库密钥，量化法超越市场模型。

AI progress this week features GPT-5.6 Sol/Terra/Luna benchmarking and tiering, Grok CLI accidentally exposes repos and secrets, and a quantization me

GPT-5.6Sol领跑设计，6M用户Codex/ChatGPT放宽，插件库发布

GPT-5.6 Sol tops design arena with high efficiency and 6M users, while Codex and ChatGPT lift limits and OpenAI releases Codex plugin examples reposit

GPT-Live全部上线；GPT-5.6Plus发布，招测试者，引发超智能机猜测

GPT-Live is now available to all ChatGPT users; GPT-5.6 Plus released with tester recruitment, fueling speculation about the first ultraintelligent ma

GPT-5.6Sol上线，限速重置；Qwen3.6提速HY3推128GB开源模型

GPT-5.6 Sol launches, triggering ChatGPT rate-limit resets; Qwen3.6 Quants run 2.5× faster on 24GB GPUs, and Tencent-HY3 debuts a 128GB open-weights m

OpenAI 新品：ChatGPT Work、GPT-5.6、GPT-Live。

OpenAI unveils ChatGPT Work with a cross-app agent, GPT-5.6, and GPT-Live rollout to Go, Plus, and Pro.

GPT-Live上线，GPT-5.6公测，Grok4.5首秀，GDPv2第4名。

OpenAI launches GPT-Live voice models in ChatGPT; GPT-5.6 Sol/Terra/Luna go public; SpaceXAI Grok 4.5 debuts, ranking #4 on GDPval-AA v2.

Meta推MuseImage首个代理图像生成器；ClaudeCowork公测

Meta debuts Muse Image, its first agentic image generator, while Claude Cowork goes beta on mobile/web and Monogram AI secures $40M seed.

OpenAI在Cerebras跑GPT-5.6，750/秒；腾讯Hy3 MoE两周免费API；Anthropic称Claude在训练中发现隐性J-Space。

OpenAI speeds GPT-5.6 Sol on Cerebras at 750 tokens/s; Tencent offers free API for 2 weeks for 295B MoE; Anthropic reveals Claude's hidden 'J-Space' d

Groq拟与NVIDIA合作，Anthropic升级，Longcat开源MIT。

Groq signals a potential $20B NVIDIA partnership as Anthropic plans frequent frontier upgrades every 2–4 months and Longcat 2.0 releases under MIT lic

ClaudeMythosPreview漏洞激增，MTIA芯片交易65亿美元。

Vulnerabilities around Claude Mythos Preview spike as major AI hardware and model competition intensifies with a $6.5B Meta-Samsung MTIA chip deal.

7月推GPT-5.6，限额宽松；Meta预告 Muse Spark 提升编码。

OpenAI plans GPT-5.6 launch in July with generous limits, while Meta teases Muse Spark, boosting agentic coding and AI benchmarking.

AI时代漏洞检测；评估框架将多模态对齐人类感知；局部解码路由提升分布式模型服务。

Mythos AI era enables scalable vulnerability detection; PerceptionRubrics calibrates multimodal evaluation to human perception; ELDR enables locality-

Anthropic将ClaudeFable5重启并开放访问发布Sonnet5。

Anthropic relaunches Claude Fable 5 globally with updated safety classifiers, grants promotional access, and releases Claude Sonnet 5.

ClaudeSonnet5首秀代理性AI；Gemini亮相；Etched获合约。

Anthropic unveils Claude Sonnet 5 as agentic AI, while Gemini debuts Nano Banana 2 Lite and Omni Flash, and Etched wins $1B+ after A0 tapeout.

实时脑解码问世；Claude在Azure；中国新模超越Fable5和GPT5.6

Real-time brain-to-text decoding debuts with Brain2Qwerty v2, Claude launched in Foundry GA on Azure, and a Chinese AI model beats Fable 5 and GPT-5.6

开源LLM热潮，GLM-5.2推动增长，企业引入Claude与OCR解析长文。

Open-source LLM momentum surges as GLM-5.2 fuels Databricks growth, while enterprises adopt persistent Claude and Unlimited-OCR vLLM for long document

AI安全治理成焦点，大量诈骗，GPT组受美监管，Mythos5恢复关键基础设施。

AI security and governance dominate: Claude distillation fraud hits millions, GPT-5.6 trio debuts under US oversight, Mythos 5 access restored for cri

OpenAI 预览 GPT-5.6，AWS 推出代理工具包，AI 反弹加速。

OpenAI previews GPT-5.6 variants Sol, Terra, Luna; AWS launches Agent Toolkit for AI apps, while the AI backlash appears to be accelerating.

Codex代理跨部门，IBM亚nm芯片0.7nm3D栈，OpenAI定制芯片

AI agents with Codex transform work across departments; IBM unveils sub-1nm chip with 0.7nm 3D nanostack; OpenAI debuts Broadcom custom chip.

博通携手OpenAI推芯片，GPT-5.5升级，双头垄断驱动470亿美元ARR。

OpenAI debuts Jalapeño AI chip with Broadcom, GPT-5.5 upgrade, Anthropic-OpenAI duopoly drives $47B ARR.

SlackClaudeTag上线；GLM5.2领跑；Prime-rl开启RL。

Anthropic launches Claude Tag for Slack; GLM-5.2 tops Claude 4.8 Opus and GPT-5.5 in one-shot test; Prime-rl enables RL on GLM-5 at trillion-parameter

扩展修补，新增模型与工具；Gemini以GA为主接口，GLM5.2开放权重促开源

OpenAI Daybreak expands patching with new models and tools; Gemini GA becomes primary interface, GLM 5.2 opens weights to autoresearch.

Mythos新版本来袭，GLM-5.2开源编码受关注，NSA据称无安全访问。

New Mythos version emerges amid unclear release details, as GLM-5.2 showcases coding in open source, while NSA reportedly accessed Claude Fable and My

GPT-5.6 可能升级；GLM-5.2 本地 OpenCode 于 Claude Opus；John Jumper 跳槽 Anthropic。

GPT-5.6 eyed as a meaningful GPT-5.5 upgrade, GLM 5.2 with OpenCode on Claude Opus locally, and John Jumper leaves DeepMind for Anthropic.

美方收紧前沿AI访问，开放权重达前沿，约翰·贾姆珀加盟Anthropic。

US government and Anthropic curb frontier AI access as open-weights reach frontier performance, and John Jumper joins Anthropic from DeepMind.

从谷歌加盟OpenAI，GPT-5.5对齐前沿健康模型，GLM-5.2开放权重。

Noam Shazeer joins OpenAI from Google as GPT-5.5 achieves parity with frontier health models and GLM-5.2 opens weights for local run.

GLM-5.2以51分领跑AI榜；Ling-2.6/Ring-2.6促成深度代理

GLM-5.2 leads AI indexes with 51 score, LifeSciBench benchmarks AI in real-world life sciences, and Ling-2.6 and Ring-2.6 enable instant, deep agentic

SpaceX全股收购CursorAI；GLM-5.2获1360Elo。

SpaceX to buy Cursor AI in stock deal as GLM-5.2 hits 1360 Elo, and Codex rolls out across Europe with Chrome extension and memory.

Sonic-3.5/Ink-2领衔TTS/STT，KimiK2.7发布，降48%

Sonic-3.5 and Ink-2 lead streaming TTS and STT; Kimi K2.7 code released; local run time drops 48%; Anthropic Fable 5 shifts to policy compliance.

白宫政策紧张下，该公司CEO任期存疑，导数无关优化在MNIST超越Adam。

Anthropic faces policy tensions with the White House and leadership uncertainty amid a Fable/Mythos dispute, while derivative-free optimization outper

美方暂停Fable5/Mythos5访问；Adaline2.0促自改循环。

US directive suspends Claude Fable 5 and Mythos 5 access; Adaline 2.0 enables self improvement loops as MiniMaxAI releases M3 weights.

Gemini-SQL2在Gemini 3.1 Pro实现文本转SQL突破，Kimi-K2.7-Code开源并提效，MiniMax M3开放权重至 Hugging Face。

Gemini-SQL2 debuts breakthrough text-to-SQL on Gemini 3.1 Pro, while Kimi-K2.7-Code is open-sourced with performance uplift and MiniMax M3 publishes o

Anthropic 推 Claude Corps，培千 AI；Fable 5 守护，回退至 Opus 4.8；DeepMind 携 Palmeiras 推进 足球 AI。

Anthropic launches Claude Corps to train 1,000 AI fellows; Fable 5 safeguards show Opus 4.8 fallback; DeepMind partners Palmeiras to advance tacticAI

谷歌发布DiffusionGemma26B，turbovec超越FAISS。

Google unveils DiffusionGemma 26B AI diffusion model, alongside ABot-Earth 0.5's generative 3D Earth model and turbovec's Rust TurboQuant beating FAIS

新一代模型发布，具安全基准与实时翻译能力，且可通过权重内嵌技能。

Claude Fable 5 debuts with Mythos-based safeguards and SOTA benchmarks, Gemini 3.5 Flash adds real-time speech translation, and LatentSkill enables in

工具进阶：代理对话与推理，LLM用于科学发现，代码一行安装、视频上下文与插件。

AI tools gain agentic chat and advanced reasoning, enabling LLMs for scientific discovery, plus Kimi Code adds one-line installs, video context, and p

Siri引入GeminiAI，NHS推Copilot，ChatGPT升级为超应用

Siri to get Gemini-powered AI at WWDC 2026 as NHS rolls out Copilot to 505,000 staff and ChatGPT undergoes a major superapp overhaul

Claude Mythos5 将发布；CVPR2026 最佳论文获奖。

Claude Mythos5 release expected; CVPR2026 Best Paper winner announced.

该公司呼吁全球暂停以防自我提升风险；新模型来临，设备端将有开源视觉语言模型。

Anthropic calls for a global AI pause amid self-improvement risk as Claude Mythos, GPT-5.6, Gemini 3.5 Pro loom, with open-weight LFM2.5-VL-Extract en

英伟达推Nemotron3U长时代理Hermes自改多模4GBGPU下70B推理

NVIDIA unveils Nemotron 3 Ultra for long-running AI agents, Hermes launches self-improving multi-model tool, and AirLLM enables 70B LLM inference on 4

微软披露MAI报告透明后训练模型，发布Scout与Gemma 4 12B。

Microsoft unveils Scout autonomous AI agent on OpenClaw and reveals transparent post-trained models in its MAI report, alongside Gemma 4 12B for encod

MAI发布模型，ClaudeCode升级，OpenAI接入Codex至Sites

MAI unveils seven new models led by MAI-Thinking-1, Claude Code gains workflows, its biggest upgrade since Skills, and OpenAI extends Codex to Sites.

NVIDIA Cosmos3 开放模型具视觉推理，Nemotron3 Ultra550B，Anthropic Opus4.8刷新ARC-AGI-3 SOTA。

NVIDIA unveils Cosmos 3 Open Omnimodel with vision reasoning, Nemotron 3 Ultra at 550B, and Anthropic Opus 4.8 sets SOTA on ARC-AGI-3.

Rosalind生物防御上线机器人进展；ParakeetSTT移植至ggml提速

OpenAI expands with Rosalind Biodefense for vetted partners, advances in robotics hiring, and Parakeet STT port to ggml delivering faster performance.

前DeepMind团队融资，Hermes超越OpenClaw，托管评估简化模型。

A former DeepMind team raises $50M for a recursive AI lab, while Hermes Agentic AI tops OpenClaw and hosted evaluations simplify model testing.

Codex在Windows可用，llama.app统一入口，LoRA将多效合并。

Codex runs on Windows for PC task actions, llama.cpp launches llama.app with unified entrypoint, and CollectionLoRA distills 50 effects into one LoRA.

Opus 4.8提升判断力/独立性；Mythos即将推出；128K上下文优化。

Claude Opus 4.8 delivers sharper judgment and independence as Anthropic plans Claude Mythos with boosted inference compute; LFM2.5-8B-A1B debuts as de

三大AI新进展：蛋白质语言引擎、原生多模态嵌入、分块扩散训练。

Open-source ESMFold2 unveils protein language engine with 6.8B Protein Atlas; Gemini Embedding 2 delivers native multimodal embeddings; DiffusionBlock

SynthID联手各方，将水印扩展至百亿级，推动AI溯源。

SynthID scales watermarking to 100B pieces with OpenAI, ElevenLabs, and Kakao, marking a major leap in large-scale AI content provenance.

开放题Gemini解题，Grok将发布，DeepSeek提供FlashFree。

AlphaProof Nexus solves open math with Gemini; Grok V9-Medium completes training; release in weeks; DeepSeek V4 goes Flash Free on Nous Portal for Her

DeepMind用LLM解九题，LongLive2.0推动NVFP4并行视频。

DeepMind solves nine Erdős problems with LLM-driven agents; LongLive 2.0 enables NVFP4 parallel video; AI automation raises software engineer demand a

通用人工智能近在眼前，估值高，推出4–8倍更快的任务模型。

As AGI looms, Hassabis says singularity may arrive in years; Anthropic and xAI valued at $61.5B and $80B, while ZeroEntropy builds 4–8x faster task-sp

GeminiOmni首秀，玻璃翼曝高危漏洞，Qwen3.7-Max棋测领先。

Gemini Omni debuts with standout AI creations; Anthropic unveils Glasswing and uncovers 10k+ high-severity vulnerabilities; Qwen 3.7-Max beats Opus 4.

OpenAI解决平面距离问题；Hark完成7亿美元融资

OpenAI's AI solves planar unit distance problem; Hark raises $700M at $6B valuation; Gemini 3.5 Flash tops APEX-Agents-AA benchmark.

通用AI解决重大未解题，Gemini3.5自称更快编码，AI推翻Erdős猜想。

A general-purpose AI solves a major open math problem, while Gemini 3.5 Flash touts faster coding, and AI disproves an Erdős unit distance conjecture.

Omni上线，视频驱动万物生成，Spark全时AI，OpenAI增设水印验证。

Gemini Omni debuts, enabling universal creation from video; Google Gemini Spark offers a 24/7 personal AI; OpenAI adds image watermark verification.

CursorAI与SpaceXAI在Colossus2训练更大模型。

Cursor AI and SpaceXAI train a larger model on Colossus 2; Cloudflare tests Anthropic Mythos on 50 repos, and Anthropic to acquire Stainless SDK Platf

GrokV9完成训练，3–4周上线；AI18个月内自动化白领9B模型工具调用领先。

GrokV9 training completed; release in 3–4 weeks, AI may automate most white-collar work in 18 months, and a 9B model excels in tool-calling.

EGGROLL实现无反向传播训练，Mythos绕MIE攻内核，统一放大推理金牌。

Researchers showcase backprop-free AI training with EGGROLL, reveal Mythos enables macOS kernel exploitation bypassing MIE, and achieve gold-medal-lev

OpenAI增理财Anthropic推MonetFigure扩展50小时排序直播

OpenAI adds personal finance features to ChatGPT Pro; Anthropic debuts Claude Monet amid painter displacement; Figure extends livestream to 50+ hours

Codex登陆ChatGPT移动端，跨设备工作；Gemini闪传闻与LoRA平台

Codex debuts in the ChatGPT mobile app for cross-device work, alongside rumored Gemini Flash performance and MinT’s scalable LoRA platform.

令牌叠加LLM预训练速2–3倍，世界模型对代理至关，Claude月度额度上线。

Token superposition boosts LLM pretraining speed 2–3×; world models essential for reliable agentic AI; Claude offers monthly credits for programmatic

Gemini笔记本秋上市，药物AI获21亿投，GPT5.5xhigh84.9%。

Googlebook unveils Gemini laptop this fall; Isomorphic Labs raises $2.1B for AI drug discovery; GPT 5.5 xhigh tops WeirdML at 84.9%.

OpenAI推动前沿AI投产防御，谷歌公布GeminiOmni视频模型。

OpenAI expands to production with a deployment company and Daybreak for cyber defense, as Google unveils Gemini Omni’s coherent video model.

统一LLM检查点，Cull开源图像工具，Codex靠安全审计月赚506美元。

NVIDIA Star Elastic unifies 30B/23B/12B LLM checkpoints, Cull launches open-source image data scraping and classification, Codex earns $506/month via

ERNIE5.1提效，DeepSeekv4可运行，改错位提Claude。

AI advances: ERNIE 5.1 boosts efficiency and performance, DeepSeek v4 enables 1M-context local Mac use, and Anthropic refines misalignment explanation

实时翻译上线，防护预览升级，FrontierMath得分48%。

GPT-Realtime-2 real-time meeting translation launches; GPT-5.5-Cyber previews bolster infrastructure defense; DeepMind's co-mathematician scores 48% o

GPT-Realtime-2与GPT-5推理，Gemini3.1高效量产。

OpenAI unveils GPT-Realtime-2 with GPT-5-class reasoning amid Gemini 3.1 Flash-Lite's cost-efficient high-volume model and Grok Voice Think Fast 1.0 f

OpenAI 推出 MRC，SpaceXAI 让 Claude 使用 Colossus 1，印度独研者 ICML 论文获准。

OpenAI launches MRC for AI training; SpaceXAI grants Anthropic access to Colossus 1 for Claude; a solo Indian researcher’s ICML paper is accepted.

GPT-5.5提升回复，Grok4.3加速，Heretic1.3发布基准。

New AI models roll out with faster performance and warmer responses: GPT-5.5 in ChatGPT, Grok 4.3 on xAI API, and Heretic 1.3 with reproducible benchm

AI助力NASA数据发现百余隐匿系外行星，MoE模型及Claude工具扩展能力。

AI-assisted NASA data analysis reveals 100+ hidden exoplanets, while compact 100M-parameter MoE models and expanded Claude tooling broaden AI capabili

巨大模型本地化提升推理速率，规模并非唯一答案。

Huge models go local, boosting on-device speed from 1k to 100k tokens/sec, while debates emerge over scale versus quality in LLMs.

OpenAI与Anthropic发布新模型，AI证明推进Erdős 1196与60年猜想，GLM-5V-Turbo成多模态基础模型。

OpenAI and Anthropic released models amid previews; AI proofs advance Erdős Problem 1196 and a 60-year conjecture, while GLM-5V-Turbo debuts as a mult

GPT-5.5首发最强，API收入增长；部门签七家AI合约，CAD与两软件集成。

GPT-5.5 debuts strongest yet with rising API revenue; Dept of War signs frontier AI deals with seven firms; AI CAD integrates with Onshape and Fusion.

向防御者推出GPT-5.5-Cyber，增强Codex提示与ChatGPT安全。

OpenAI rolls out GPT-5.5-Cyber to critical cyber defenders, enables role-based Codex prompts, and strengthens ChatGPT security.

OpenAI 模型登陆 Bedrock；谷歌投资提振，Claude.ai 出错。

OpenAI models arrive on Amazon Bedrock, signaling deeper cloud ecosystem integration, as Google reports momentum from AI investments, and Claude.ai ex

GPT-5.4 Pro破解埃尔德斯难题，推动数学；Anthropic收紧Claude Code，禁企业用户；MistralAI公布工作流预览。

GPT-5.4 Pro solves an Erdős problem, advances math; Anthropic nerfs Claude Code and bans corporate customers; MistralAI previews enterprise workflows.

TRELLIS.2开源4B3D；离线WebGPU代理；MiMoMIT商用。

Microsoft unveils TRELLIS.2, an open-source 4B 3D generator; a local offline WebGPU browser agent debuts; Xiaomi releases MiMo-V2.5 under MIT for comm

谷歌拟向Anthropic投资最高400亿美元，用于现金与算力，标志AI合作。

Google plans up to $40B investment in Anthropic for cash and compute, signaling a major AI infrastructure partnership.

AI招聘偏好AI简历；Qwen首发，80TPS在RTX5090，218k上下文。

AI hiring tools favor AI-written resumes over humans, while Qwen-Image-2.0-Pro debuts and Qwen3.6-27B reaches 80 TPS on RTX 5090 with 218k context.

GPT-5.5覆盖Copilot，Codex增浏览测试，DeepSeekV4开源。

GPT-5.5 rolls out across Copilot, M365 Copilot, Studio, and Foundry; Codex adds browser-based testing; DeepSeek-V4 previews with 1M-context support.

GPT-5.5首发，OpenAI推出临床版ChatGPT和健康基准，速增。

GPT-5.5 debuts with practical intelligence as OpenAI launches clinician-focused ChatGPT and HealthBench, and speeds up ChatGPT and Codex per-token per

ChatGPT工作区代理推动工作流，谷歌以Gemini与TPU8t8i提升AI。

AI workflow automation expands with ChatGPT Workspace Agents; Google scales enterprise AI with Gemini and unveils TPU 8t/8i dual-chip accelerators.

ChatGPT图像2.0发布，SpaceX与CursorAI谋60亿合作。

ChatGPT Images 2.0 debuts with a multimodal image model, while SpaceX and Cursor AI explore a major $60B collaboration to scale Composer.

开源AI崛起代理群领跑Claude扩至5GWQwen3.6-Max预览旗舰

Open-source AI surges with Kimi K2.6 agent swarms, Anthropic expanding Claude on 5 GW AWS, and Qwen3.6-Max previewing the next flagship model.

AI监管与工具升级并进，IEEE指数揭示AI2026态势。

Google Gemini scans photos amid EU pushback; Anthropic launches Claude Design; IEEE-indexed graphs map AI's 2026 state.

Anthropic封禁60余Claude；新模胜Qwen；Code免费本地运行。

Anthropic halts Claude access for 60+ employees overnight; 18B Frankenstein outperforms Qwen 3.6-35B-A3B on 44-test suite; Claude Code runs locally on