AI TimelineAI 大事记

Microsoft AI announced MAI-Transcribe-1, a new speech-to-text model, and made MAI-Voice-1 and MAI-Image-2 broadly available for commercial use through Microsoft Foundry and MAI Playground. The three models cover transcription, voice generation, and image generation — positioning Microsoft as a full multimodal AI provider independent of OpenAI.



微软 AI 部门（由 Mustafa Suleiman 领导）于 4 月 2 日宣布 MAI-Transcribe-1 正式发布，同时将 MAI-Voice-1 和 MAI-Image-2 开放给开发者用于商业场景。三款模型均可通过 Microsoft Foundry 和 MAI Playground 调用，定位为高质量、高性价比的多模态 AI 能力，覆盖语音转文字、语音生成和图像生成三个方向。这标志着微软在 AI 模型层的自主研发能力进一步独立于 OpenAI。

OpenAI's latest funding round, led by Amazon, Nvidia, and SoftBank, values the AI lab at $852 billion as it nears an IPO. The round also opened $3B to retail investors.



OpenAI 最新一轮融资由亚马逊、英伟达和软银领投，将公司估值推至 8520 亿美元，向 IPO 迈进。本轮还向零售投资者开放了 30 亿美元份额，总融资规模达 122 亿美元。

Google DeepMind launched Gemma 4 on April 2, 2026 — a family of four open models (E2B, E4B, 26B MoE, 31B Dense) released under the Apache 2.0 license. The models are purpose-built for advanced reasoning and agentic workflows, designed to run on local hardware including mobile devices via AICore.



Google DeepMind 于 2026 年 4 月 2 日正式发布 Gemma 4，包含四个规格：E2B（有效 2B）、E4B（有效 4B）、26B MoE（混合专家）和 31B Dense，全系采用 Apache 2.0 开源协议。模型专为高级推理和 Agent 工作流设计，支持在本地硬件（包括移动端 AICore）上运行。这是 Google 开源模型系列一年来最大规模的更新，同时标志着授权协议从此前的限制性协议切换至更宽松的 Apache 2.0。

OpenAI closed a monster $122 billion funding round, raising $3 billion from retail investors while remaining private. The round, led by Amazon, NVIDIA, and SoftBank, values the AI lab at $852 billion — approaching the world's most valuable companies — as it accelerates toward an IPO.



OpenAI 尚未上市，但完成了 1220 亿美元融资轮，其中 30 亿美元来自散户投资者，整体估值达到 8520 亿美元，主要由 Amazon、NVIDIA 和 SoftBank 领投。这是人类历史上规模最大的科技公司私募融资之一，OpenAI IPO 路径正在加速。

Rebellions, an AI inference chip startup planning to go public later in 2026, raised $400M in a pre-IPO round at a $2.3B valuation. The company designs chips specifically for AI inference workloads, positioning itself as a challenger to Nvidia's dominance in the AI hardware market.



AI 推理芯片初创公司 Rebellions 完成 4 亿美元融资，融资后估值达 23 亿美元。公司计划于 2026 年晚些时候上市。Rebellions 专注于 AI 推理芯片的研发设计，直接挑战英伟达在 AI 硬件市场的主导地位。

Mistral AI secured $830M in debt financing to build a new data center near Paris, with operations expected to begin in Q2 2026. The move signals Mistral's ambition to build its own compute infrastructure as European AI labs race to gain hardware independence from US cloud providers.



Mistral AI 通过债务融资筹集 8.3 亿美元，计划在巴黎附近建设新数据中心，预计 2026 年第二季度投入运营。此举标志着 Mistral 正加速推进自有算力基础设施建设，欧洲 AI 实验室正在争相降低对美国云服务商的依赖。

Sanders and Rep. Alexandria Ocasio-Cortez introduced companion legislation to halt construction on new data centers until Congress passes comprehensive AI regulation.



美国参议员桑德斯（Bernie Sanders）与众议员 AOC（Alexandria Ocasio-Cortez）联合提出配套立法，要求在国会通过全面 AI 监管法案之前，暂停所有新数据中心的建设。这是迄今为止对 AI 基础设施扩张最强硬的国会层面立法提案之一。

OpenAI announced the discontinuation of Sora, its AI video generation app, with web and app experiences shutting down on April 26, 2026. The company stated it is reallocating compute resources toward other priorities, including world simulation for robotics. Sora had launched publicly in late 2024, just months before the shutdown announcement.



OpenAI 宣布将于 2026 年 4 月 26 日正式停止 Sora 视频生成应用的 Web 和 App 服务。公司表示将把算力资源转向其他方向，包括服务于机器人领域的世界模拟技术。Sora 于 2024 年底正式向公众开放，运营仅约 6 个月。

Gimlet Labs just raised an $80 million Series A for tech that lets AI run across NVIDIA, AMD, Intel, ARM, Cerebras and d-Matrix chips, simultaneously.



Gimlet Labs 完成 8000 万美元 A 轮融资，其技术能让 AI 推理同时跨 NVIDIA、AMD、Intel、ARM、Cerebras 和 d-Matrix 芯片运行，打破单一硬件供应商锁定。

Shortly after Amazon announced its $50 billion investment in OpenAI, AWS invited TechCrunch on a private tour of the Trainium chip lab at the heart of the deal. Trainium is Amazon's custom AI training chip, and it has won over major AI labs including Anthropic, OpenAI, and Apple as a cost-effective alternative to NVIDIA hardware.



亚马逊宣布向 OpenAI 投资 500 亿美元后，AWS 邀请 TechCrunch 对 Trainium 芯片实验室进行了独家参观。Trainium 是亚马逊自研 AI 训练芯片，已赢得 Anthropic、OpenAI、Apple 等顶级 AI 机构的采用，被视为英伟达硬件的成本效益替代方案。

Jeff Bezos is reportedly pursuing a new project: acquiring legacy industrial and manufacturing companies and revamping them with AI technology, with an ambition to deploy up to $100 billion in capital.



据报道，杰夫·贝索斯正在主导一个新项目：收购传统工业和制造业企业，并用AI技术对其进行改造升级，计划部署资本规模高达1000亿美元。这标志着贝索斯将个人投资重心从太空（蓝色起源）扩展到实体经济AI化改造赛道。

Microsoft announced MAI-Image-2, its second-generation text-to-image model, debuting at #3 on the Arena.ai text-to-image leaderboard globally. The model features enhanced photorealism with accurate lighting and skin tones, reliable in-image text generation for posters and infographics, and rich scene generation for cinematic or surreal compositions. Available immediately in the MAI Playground.



微软发布 MAI-Image-2 文生图模型，在 Arena.ai 全球文生图排行榜中位列第三。新模型主打三大能力：增强的照片级写实效果（自然光影、准确肤色）、可靠的图内文字生成（适合海报、信息图）、以及复杂场景生成能力。现已在 MAI Playground 中开放试用。

OpenAI announced the acquisition of Astral, the company behind widely-used Python developer tools uv, Ruff, and ty — totaling hundreds of millions of downloads per month. The Astral team will join OpenAI's Codex team, which has seen 3x user growth and 5x usage increase since the start of 2026 with over 2 million weekly active users. OpenAI plans to continue supporting Astral's open source tools after the deal closes, while integrating the toolchain into the Codex software development workflow.



OpenAI 宣布收购 Astral，即 Python 开发者工具 uv、Ruff、ty 的开发公司，这三款工具月下载量高达数亿次。Astral 团队将加入 OpenAI 的 Codex 团队。Codex 自 2026 年初以来用户数增长 3 倍、使用量增长 5 倍，每周活跃用户超 200 万。收购完成后，Astral 的开源工具将继续维护，并深度整合进 Codex 的软件开发工作流中。

MiniMax released M2.7, its first model to participate in its own development through self-evolution. The model achieves 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, approaching Claude Opus-level performance. M2.7 is capable of building complex agent harnesses, managing Agent Teams, and completing long-horizon productivity tasks. It maintains a 97% skill adherence rate with over 40 complex skills. Available via MiniMax Agent platform and API.



MiniMax 发布 M2.7，其首个深度参与自身进化的 AI 模型。在 SWE-Pro 基准测试中得分 56.22%，Terminal Bench 2 得分 57.0%，接近 Claude Opus 水平。M2.7 能够构建复杂 Agent 任务链，管理 Agent 团队，支持动态工具搜索，并在 40 多个复杂技能中保持 97% 的执行准确率。现已通过 MiniMax Agent 平台及 API 对外开放。

Mistral introduced Forge at NVIDIA GTC, a platform designed for enterprises and public sector organizations to build frontier-grade AI models grounded in their own proprietary data. Forge covers the full model lifecycle — creation, specialization, and continuous improvement — using Mistral's open-weight model library (including Mistral Small 4) without requiring customers to expose their data to the cloud.



Mistral 在 NVIDIA GTC 上发布 Forge，一个面向企业和公共机构的 AI 模型训练平台。企业可以在自有数据基础上构建、微调和迭代模型，覆盖完整模型生命周期，数据无需上传至云端。基础模型库包含 Mistral Small 4 等开源权重模型。

Under the guidance of NVIDIA CEO Jensen Huang, chip cooling startup Frore pivoted from solid-state cooling to liquid cooling technology, raising $143 million at a $1.64 billion valuation. Surging AI chip power consumption is driving demand for cooling infrastructure, making Frore the latest deep tech company to reach unicorn status.



在 NVIDIA CEO 黄仁勋的推动下，芯片散热初创公司 Frore 从固态散热转向液冷技术，完成 1.43 亿美元融资，估值达到 16.4 亿美元。AI 芯片功耗飙升带动散热基础设施需求，Frore 成为最新一家跨入独角兽行列的深科技公司。

Mistral releases Leanstral Comments, a lightweight model focused on code review and comment generation for developer workflows.



Mistral 发布 Leanstral Comments——专注代码审查和注释生成的轻量模型，面向开发者工作流优化。

OpenAI released GPT-5.4 mini and GPT-5.4 nano, bringing many capabilities of the GPT-5.4 flagship to faster, more cost-efficient models. Both are available in the API, Codex, and ChatGPT — mini is accessible to free users. The models are designed for high-volume, latency-sensitive workloads including coding assistants and multi-modal pipelines, running more than twice as fast as previous GPT-5 mini variants.



OpenAI 发布 GPT-5.4 mini 和 nano，将旗舰模型的能力下放到更快、更高性价比的版本。两款模型同步上线 API、Codex 和 ChatGPT，免费用户也可使用 mini 版本。主打高并发、低延迟场景（代理、编码助手、多模态工作流），速度较上代 GPT-5 mini 提升 2 倍以上。

Mistral released Mistral Small 4, a 119B-parameter Mixture-of-Experts model under the Apache 2.0 open-source license. The model unifies instruct, reasoning, vision, and coding capabilities in a single architecture (128 experts). It was announced alongside the Forge enterprise training platform at NVIDIA GTC and represents the next major release in the Mistral Small family.



Mistral 发布 Mistral Small 4，119B 参数混合专家（MoE）架构，Apache 2.0 开源协议。整合文本理解、图像识别、推理和代码能力于单一模型（128 个专家网络）。该模型在 NVIDIA GTC 上与 Forge 企业训练平台同步发布，是 Mistral Small 家族的重大升级。

Index Ventures Partner Shardul Shah says cybersecurity startup Wiz sits "at the center of three tailwinds: AI, cloud, and security spend." Those tailwinds powered what just became the largest venture-backed acquisition in history — Google's $32 billion deal for Wiz, finalized after Wiz declined Google's earlier $23B offer in 2024.

So what? Cybersecurity is now table stakes for enterprise AI. By acquiring Wiz, Google gets the leading cloud security platform and closes a major gap against Microsoft's security portfolio. This is a bet that the next wave of AI adoption runs through securing cloud infrastructure.



Index Ventures 合伙人 Shardul Shah 说，网络安全初创公司 Wiz 站在「AI、云计算、安全支出」三大顺风的交叉口。这三股力量共同推动了这笔创投史上最大并购案——Google 以 320 亿美元完成对 Wiz 的收购，此前 Wiz 曾于 2024 年拒绝了 Google 230 亿美元的报价。

网络安全已成为企业 AI 的标配。通过收购 Wiz，Google 获得领先的云安全平台，缩小了与微软安全产品组合的差距。这是一个押注：AI 下一波应用浪潮的关键在于保护云基础设施。

Nvidia-backed British AI infrastructure startup Nscale has raised another megaround of $2 billion, lifting its valuation to $14.6B. Former Meta COO Sheryl Sandberg and former UK Deputy Prime Minister Nick Clegg have both joined its board.

So what? AI infrastructure is the arms race that won't slow down. Nscale landing two of tech/politics' most prominent names signals it's positioning as a serious European alternative to US hyperscalers — and the "Stargate Norway" framing is a deliberate play for sovereign AI infrastructure contracts.



英伟达支持的英国 AI 基础设施初创公司 Nscale 完成新一轮 20 亿美元融资，估值升至 146 亿美元。前 Meta COO Sheryl Sandberg 与前英国副首相 Nick Clegg 双双加入其董事会。Nscale 被称为「挪威星际之门」，专注于为 AI 训练和推理提供算力基础设施。

OpenAI is acquiring Promptfoo, an AI security platform used by over 25% of Fortune 500 companies, with a widely-used open-source CLI for evaluating and red-teaming LLM applications. Once finalized, Promptfoo's capabilities will be integrated into OpenAI Frontier — the platform for building and operating AI coworkers — bringing native security testing, red-teaming, and oversight tooling to enterprise AI deployments.

So what? As agents get deployed into real workflows, security and eval infrastructure become table stakes. OpenAI buying Promptfoo signals they want to own that layer end-to-end inside Frontier rather than leaving it to third parties. The open-source project continues, but the enterprise roadmap now runs through OpenAI.



OpenAI 宣布收购 Promptfoo——一个被超过 25% 的财富 500 强公司使用的 AI 安全测试平台，拥有广泛使用的开源 CLI。收购后将把 Promptfoo 的安全测试和红队能力整合进 OpenAI Frontier 企业平台，帮助部署 AI Agent 的企业自动发现提示注入、数据泄露等漏洞。

AWS announces Amazon Connect Health, a purpose-built agentic AI solution that handles high-volume administrative tasks — appointment scheduling, clinical documentation, and medical coding — with real-time EHR access and insurance eligibility checks.

So what? AWS is betting that AI agents can eat healthcare admin — scheduling, billing, patient engagement. If this works at scale, it's a template for every vertical. The "AI agent platform" play is now AWS vs Google vs Microsoft, sector by sector.



AWS 发布 Amazon Connect Health，专为医疗行业打造的代理 AI 解决方案，处理高频行政任务——预约排期、临床文档记录、医疗编码——实时访问 EHR 和保险资格核查。AWS 赌 AI 代理能吃掉医疗行政——排班、计费、患者沟通。如果大规模跑通，将成为各垂直行业的模板。

ment of Defense has officially labeled Anthropic a supply-chain risk, making the AI firm the first American company with the label. Meanwhile, the DOD continues to use Anthropic's AI in Iran.



美国国防部正式将 Anthropic 列为供应链风险，使其成为首家获此标签的美国公司。同时 DOD 继续在伊朗行动中使用 Anthropic 的 AI。

Google releases Gemini 3.1 Flash-Lite, available in preview via Gemini API in Google AI Studio and Vertex AI. Positioned as their most cost-effective AI model to date.

So what? The race to the bottom on inference cost continues. Google is making the "good enough" tier cheaper and more accessible, which matters for agent architectures where you're chaining many calls.



Google 发布 Gemini 3.1 Flash-Lite，通过 Gemini API 在 Google AI Studio 和 Vertex AI 预览。定位为目前最具成本效益的 AI 模型。推理成本的底部竞赛持续进行。Google 把「够用」层做得更便宜更普及——对代理架构中链式调用场景很重要。

OpenAI releases GPT-5.4 in three variants: GPT-5.3 Instant (fast everyday tasks), GPT-5.4 Thinking (professional workflows with reasoning), and GPT-5.4 Pro. Features an opt-in 1M token context window. GPT-5.4 Thinking replaces GPT-5.2 Thinking, which sunsets in 3 months.

So what? Two days after GPT-5.3, OpenAI ships 5.4 — the iteration speed is insane. The Thinking variant signals reasoning models are now the default frontier, not a side experiment. 1M context opt-in is a quiet but significant infrastructure shift for agent builders.



OpenAI 发布 GPT-5.4 三个版本：GPT-5.3 Instant（日常快速任务）、GPT-5.4 Thinking（带推理的专业工作流）和 GPT-5.4 Pro。支持可选的 100 万 token 上下文窗口。GPT-5.4 Thinking 替代 GPT-5.2 Thinking（三个月后下线）。GPT-5.3 发布两天后就出 5.4——迭代速度疯狂。Thinking 版本说明推理模型已成为默认前沿方向。100 万上下文是代理开发者的重要基础设施升级。

Figure AI announces Figure 03, a consumer-grade humanoid robot designed for home assistance, priced under $50K.

So what? Humanoid robots go from factory floors to living rooms. The Jetsons timeline is real. This is a 10-year megatrend with massive investment implications.



Figure AI 公布 Figure 03，面向家庭辅助的消费级人形机器人，售价低于 $50K。人形机器人从工厂车间走进客厅。Jetsons 的时间线成真了。十年级别的大趋势，投资含义巨大。

Anthropic launches Claude Code Security, a new capability built into Claude Code on the web, in a limited research preview for Enterprise and Team customers (with expedited access for open-source maintainers). Unlike rule-based static analysis tools, it reasons about code like a human security researcher — tracing data flow, understanding component interactions, and catching complex vulnerabilities such as broken access control and business logic flaws. Findings go through a multi-stage verification process to filter false positives before reaching analysts.

So what? AI-generated code is flooding codebases faster than human reviewers can audit it. Anthropic is positioning Claude Code Security as the defense layer against AI-enabled attacks — using the same frontier reasoning capabilities that could help attackers to instead put power in defenders' hands. Releasing as a research preview first is deliberate: the dual-use risk is real, and they want to refine deployment guardrails before scaling.



Anthropic 推出 Claude Code Security——内置于 Claude Code 网页版的代码安全扫描能力，目前面向 Enterprise 和 Team 客户开放限量研究预览，开源仓库维护者可申请加速访问。与传统规则匹配的静态分析不同，它像人类安全研究员一样理解代码组件交互和数据流，能发现逻辑漏洞和访问控制缺陷。每条发现经过多阶段验证，自动过滤误报并分级严重程度。

Google released Gemini 3.1 Pro on February 19, 2026, a significant upgrade to the core intelligence behind Gemini 3, featuring a 1 million token context window and state-of-the-art performance on ARC-AGI-2 (77.1%) and other advanced benchmarks. The model significantly outperforms Gemini 3 Pro across reasoning and multimodal capabilities spanning text, images, audio, video, and code. Gemini 3.1 Pro rolled out immediately to the Gemini app as well as NotebookLM for Google AI Pro and Ultra subscribers. The release was part of Google's broader February 2026 AI drop that also included the Lyria 3 music generation model.



Google 发布 Gemini 3.1 Pro，100 万 token 上下文窗口，ARC-AGI-2 达 77.1%。推理和多模态能力大幅超越 Gemini 3 Pro，覆盖文本、图片、音频、视频和代码。立即登陆 Gemini 应用和 NotebookLM。这是 Google 2026 年 2 月 AI 大更新的一部分，同期还发布了 Lyria 3 音乐生成模型。

Anthropic released Claude Sonnet 4.6, the latest iteration in the Claude 4 Sonnet line featuring enhanced coding abilities, improved factual accuracy, and better performance on complex reasoning tasks. The model continued Anthropic's rapid iteration cadence on the mid-tier model that serves the majority of Claude API traffic. Claude Sonnet 4.6 maintained the balance of high capability and cost-efficiency that made the Sonnet family the default choice for most enterprise deployments.



Anthropic 发布 Claude Sonnet 4.6，Claude 4 Sonnet 系列最新迭代，编程能力增强、事实准确性提升、复杂推理表现更好。延续 Anthropic 在承载大部分 API 流量的中端模型上的快速迭代节奏。Sonnet 家族凭借高能力和高性价比的平衡，继续作为多数企业部署的默认选择。

ByteDance releases Doubao 2.0, an advanced model positioned for the "agent era" — designed to execute complex, multi-step real-world tasks rather than just answering questions.

So what? Doubao 2.0 is ByteDance's agentic pivot: from chatbot to AI operator. With TikTok-scale distribution behind it and Volcano Engine's enterprise contracts, ByteDance is moving up the value chain from model provider to task executor. If Doubao 2.0 can run workflows inside enterprise systems the way Doubao runs feeds on phones, ByteDance becomes a genuine enterprise AI threat.



字节跳动发布豆包 2.0，面向「代理时代」的高级模型——设计用于执行复杂多步真实任务而不只是回答问题。豆包 2.0 是字节的代理化转型：从聊天机器人到 AI 执行者。TikTok 级分发能力加持，火山引擎的企业合同做后盾——如果豆包 2.0 能在企业系统里跑工作流就像在手机上跑信息流一样顺畅，字节就成了真正的企业 AI 威胁。

Google DeepMind released a major upgrade to Gemini 3 Deep Think in February 2026, specifically designed for the complexities of science, research, and engineering where data is messy and solutions aren't black-and-white. The updated Deep Think mode uses iterative rounds of reasoning to explore multiple hypotheses simultaneously, enabling breakthroughs on hard scientific and mathematical problems. It became available to Google AI Ultra subscribers in the Gemini app and via the Gemini API for select researchers and enterprises. The update preceded Gemini 3.1 Pro by one week, establishing Deep Think as Google's premier reasoning capability for frontier research.



Google DeepMind 大幅升级 Gemini 3 Deep Think，专为科学、研究和工程中数据混乱、答案不确定的场景设计。升级版 Deep Think 用迭代多轮推理同时探索多个假设，在硬科学和数学问题上实现突破。面向 Google AI Ultra 订阅者和部分研究者及企业开放。在 Gemini 3.1 Pro 发布前一周上线。

OpenAI released GPT-5.3-Codex on February 5, 2026, advancing both frontier coding performance and reasoning/professional knowledge in a single model that is also 25% faster than GPT-5.2-Codex. The model set new benchmarks for agentic software engineering, enabling Codex to tackle longer, more complex real-world engineering workflows with improved judgment on ambiguous tasks. It launched three days after the Codex desktop app, cementing OpenAI's position in the rapidly growing AI coding agent market. GPT-5.3-Codex marked OpenAI's most direct challenge to Anthropic Claude Code and Google Jules.



OpenAI 发布 GPT-5.3-Codex，同时提升前沿编程能力和推理/专业知识，速度比 GPT-5.2-Codex 快 25%。在代理软件工程领域刷新基准，让 Codex 能处理更长更复杂的真实工程工作流。在 Codex 桌面应用上线三天后发布，巩固 OpenAI 在 AI 编程代理市场的地位。直接叫板 Anthropic Claude Code 和 Google Jules。

Anthropic published a statement committing to keeping Claude.ai an ad-free, distraction-free environment focused on genuine helpfulness. The announcement differentiated Claude from ad-supported AI services by emphasizing user trust and privacy. Anthropic articulated its philosophy that AI assistants should serve users' thinking rather than monetizing their attention. This positioning reinforced Anthropic's mission-driven approach to AI deployment distinct from advertising-dependent platforms.



Anthropic 公开承诺保持 Claude.ai 无广告、无干扰、专注真正帮助用户的环境。将 Claude 与广告支持的 AI 服务区分开，强调用户信任和隐私。Anthropic 表达了「AI 助手应服务于用户思考而非变现用户注意力」的理念。

OpenAI launched the Codex desktop app (macOS) on February 2, 2026, a focused development environment for running and managing Codex agent threads in parallel with built-in worktree support, automations, and Git integration. Built with Electron for cross-platform expansion, the app functions as a command center for developers working with Codex agents across multiple tasks simultaneously. It arrived alongside GPT-5.2-Codex and GPT-5.3-Codex, positioning the Codex ecosystem as a comprehensive AI-native IDE alternative. Windows support was announced as "coming very soon" at launch.



OpenAI 上线 Codex 桌面应用（macOS），专为并行运行和管理多个 Codex 代理线程设计，内置 worktree 支持、自动化和 Git 集成。用 Electron 构建，为跨平台扩展做准备。定位为全面的 AI 原生 IDE 替代方案。Windows 版即将推出。

NASA's Jet Propulsion Laboratory announced on January 30, 2026 that Anthropic's Claude had successfully planned and executed the first fully AI-directed drive of the Perseverance Mars rover, marking a historic milestone in autonomous space exploration. Claude analyzed the same imagery and terrain data that human mission planners use to generate waypoints, enabling the rover to safely navigate challenging Martian terrain without direct human planning. The initiative was coordinated from JPL's Rover Operations Center in collaboration with Anthropic. This made Claude the first AI model to autonomously operate a spacecraft on another planet, opening the door to faster, more efficient Mars science campaigns.



NASA 喷气推进实验室宣布 Anthropic 的 Claude 成功规划并执行了 Perseverance 火星车的首次全 AI 自主驾驶。Claude 分析与人类任务规划员相同的图像和地形数据来生成路径点，让火星车在无需人类规划的情况下安全穿越复杂火星地形。Claude 成为首个在另一颗星球上自主操控航天器的 AI 模型，为更快更高效的火星科学任务打开大门。

Unitree Robotics brought humanoid robots to China's CCTV Spring Festival Gala (春晚) — the most-watched TV event in the world with over 1 billion viewers. Their robots performed kung fu alongside martial arts students from Henan Tagou School, executing backflips and human-robot sparring sequences. The moment marked the mainstream cultural arrival of Chinese robotics and followed Unitree's announcement of 10,000+ unit annual production targets.



宇树科技的人形机器人登上央视春晚——全球收视超 10 亿的电视盛事。机器人与河南塔沟武校学生同台表演武术，完成后空翻和人机对打。中国机器人产业的主流文化亮相时刻。宇树同时公布年产万台以上的产能目标。

Alibaba releases Qwen 3.5, a frontier open-source model competitive with GPT-5 on reasoning and multilingual tasks.

So what? China's open-source AI keeps leveling up. The decoupling narrative weakens as capability gaps close. Open-source is the great equalizer.



阿里巴巴开源 Qwen 3.5，前沿模型在推理和多语言任务上与 GPT-5 竞争。中国开源 AI 持续升级。脱钩叙事随能力差距缩小而减弱。开源是最大的均衡器。

OpenAI released GPT-5.2-Codex on January 14, 2026, a version of GPT-5.2 further optimized for complex, real-world software engineering tasks. It achieved state-of-the-art results on SWE-Bench Pro (56.4%) and Terminal-Bench 2.0 (64.0%), excelling at long-horizon coding tasks, large-scale code transformations, and cybersecurity challenges. The model was immediately available to paid ChatGPT users in all Codex surfaces. GPT-5.2-Codex established the new standard for AI-assisted software engineering and signaled OpenAI's intent to dominate the developer tooling market.



OpenAI 发布 GPT-5.2-Codex，进一步优化真实世界复杂软件工程任务。SWE-Bench Pro 达 56.4%、Terminal-Bench 2.0 达 64.0%，擅长长周期编程、大规模代码变换和网络安全挑战。面向付费 ChatGPT 用户的所有 Codex 界面立即可用。为 AI 辅助软件工程立下新标杆，展示 OpenAI 主导开发者工具市场的意图。

Deepgram closes $130M Series C at $1.3B valuation, led by AVP with participation from BlackRock, Twilio, ServiceNow, and existing investors. The company ends 2024 cash-flow positive with 400+ enterprise customers.

So what? Voice AI infrastructure has its first unicorn. Deepgram's cash-flow positive status is rare in AI infrastructure — proving that owning the voice layer of the AI stack is genuinely monetizable. With 200,000+ developers and growing enterprise deployments, Deepgram is becoming the "voice API layer" for the AI economy, much like Stripe is the payment layer.



Deepgram 完成 $130M C 轮融资，估值 $1.3B，AVP 领投，BlackRock、Twilio、ServiceNow 参投。2024 年底已实现现金流为正，超 400 家企业客户。语音 AI 基础设施诞生首个独角兽。Deepgram 现金流为正在 AI 基础设施领域罕见——证明拥有 AI 技术栈的语音层是真正可变现的。20 万+ 开发者和不断增长的企业部署，Deepgram 正成为 AI 经济的「语音 API 层」。

President Trump signs an executive order rolling back Biden-era AI safety requirements and accelerating federal AI adoption.

So what? The US bets on speed over caution. AI companies get a green light, but the absence of guardrails means the market will self-regulate — for better or worse.



Trump 总统签署行政令，回滚拜登时代的 AI 安全要求并加速联邦 AI 采用。美国赌速度不赌谨慎。AI 公司获绿灯，但缺乏护栏意味着市场将自我监管——好坏参半。

CES 2026 became the "Year of the Humanoid Robot," with every major robotics company showcasing next-generation systems. Highlights included Boston Dynamics' Atlas commercial debut, NEURA Robotics' 4NE1 Gen 3 (cognitive humanoid), AgiBot's consumer-facing Genie, and dozens of other entrants. The sheer volume of humanoid demonstrations signaled that the robotics industry had crossed from research curiosity to commercial product cycle.



CES 2026 成为「人形机器人之年」，所有主要机器人公司都展示了下一代系统。亮点包括 Boston Dynamics 的 Atlas 商用首秀、NEURA Robotics 的 4NE1 Gen 3 认知人形机器人、AgiBot 面向消费者的 Genie 等。人形机器人演示的密度说明行业已从研究好奇心跨入商业产品周期。

Z.ai (Zhipu) releases GLM-4.7, an open-source AI coding model matching premium competitors on key benchmarks at $3/month (or free locally), directly challenging tools costing 60x more.

So what? China's AI price war hit developer tools. GLM-4.7 benchmarks against GitHub Copilot and Cursor while pricing at $3/month commercial or free for local deployment. As Chinese labs race toward Hong Kong IPOs, open-source as a distribution strategy makes the global coding AI market a commodity. The $200/month AI dev tool category is under existential pressure.



Z.ai（智谱）开源 GLM-4.7 AI 编程模型，在关键基准上追平高端竞品，商业版 $3/月或本地免费部署。中国 AI 价格战打到了开发工具领域。GLM-4.7 对标 GitHub Copilot 和 Cursor，定价却只是六十分之一。中国实验室冲刺港股 IPO 之际，开源成为分发策略——$200/月的 AI 开发工具品类面临生存压力。

Anthropic released Claude Opus 4.6, a significant upgrade to its flagship model featuring improved long-context reasoning, better tool use, and stronger performance on complex scientific and mathematical problems. The release demonstrated Anthropic's continued commitment to safety-focused frontier AI development. Claude Opus 4.6 pushed state-of-the-art on multiple benchmarks while maintaining Anthropic's Constitutional AI training approach.



Anthropic 发布 Claude Opus 4.6，旗舰模型重大升级，长上下文推理、工具调用和复杂科学数学问题的表现均有提升。在多项基准上刷新最优，同时保持 Constitutional AI 训练方法。

Xiaomi unveils MiMo-V2-Flash — a 309B parameter Mixture-of-Experts model (open-source) that rivals DeepSeek V3 on key benchmarks, announced alongside plans for RMB 200B in 5-year R&D investment.

So what? Xiaomi's 309B open-source MoE is proof that even hardware-first companies can compete at frontier AI scale. The pairing with a massive R&D commitment signals a strategic bet: Xiaomi's next decade is as much AI as smartphones. China's open-source AI competition is no longer between a handful of startups — it involves every major tech player.



小米亮相 MiMo-V2-Flash——309B 参数 MoE 开源模型，在关键基准上匹敌 DeepSeek V3。同时宣布 5 年 ¥2000 亿研发投入计划。小米的 309B 开源 MoE 证明硬件公司也能在前沿 AI 层面竞争。搭配大手笔研发承诺，小米下个十年的故事是 AI 和手机并重。中国开源 AI 竞争不再限于几家创业公司——每个大厂都在场。

OpenAI released GPT-5.2 on December 11, 2025, succeeding GPT-5.1 with a family of three models: GPT-5.2 Instant, GPT-5.2 Thinking, and GPT-5.2 Extended Thinking. The Extended Thinking variant allows the model to reason longer on difficult problems, making it competitive with specialized research-grade AI systems. GPT-5.2 introduced improved instruction following and better calibration on factual questions. The rapid succession from GPT-5.1 (released just a month prior) signaled OpenAI's accelerating iteration pace in the GPT-5 generation.



OpenAI 发布 GPT-5.2，包含三款模型：GPT-5.2 Instant、GPT-5.2 Thinking 和 GPT-5.2 Extended Thinking。Extended Thinking 版本可以在难题上深度推理，与专用研究级 AI 系统竞争。距 GPT-5.1 仅一个月就推出后续版本，OpenAI 在 GPT-5 世代的迭代速度惊人。

The White House issued an executive order establishing a unified national AI policy framework, signaling federal preemption of fragmented state-level AI regulations. The order explicitly criticized state algorithmic discrimination laws and conditioned federal funding on AI policy alignment.



白宫签署行政令建立统一国家 AI 政策框架，释放联邦将优先于各州碎片化 AI 监管的信号。明确批评州级算法歧视法律，将联邦资金与 AI 政策对齐挂钩。

Unitree's 5,000th general-purpose embodied robot (Lingxi X2) rolled off the production line. The company projected 10,000-20,000 shipments in 2026, and its robots were prominently featured at China's Spring Festival Gala in February 2026, demonstrating kung fu and dance routines.



宇树科技第 5,000 台通用具身机器人（灵犀 X2）下线。预计 2026 年出货 10,000-20,000 台。其机器人在 2026 年春晚上表演武术和舞蹈引发热议。

Mistral releases Mistral Large 3 (41B active / 675B total sparse MoE) and Ministral 3 family (3B, 7B, 14B) — small enough to run on laptops and drones with 4GB VRAM using 4-bit quantization.

So what? Mistral is positioning itself as the "runs anywhere" open AI stack: from frontier server inference down to embedded edge devices. As enterprise AI moves to on-premise and edge-first deployments, having a credible open model across the full compute spectrum is a moat no US-closed lab can match.



Mistral 发布 Mistral Large 3（41B 激活 / 675B 总参数稀疏 MoE）和 Ministral 3 家族（3B、7B、14B）——4GB 显存用 4-bit 量化就能在笔记本和无人机上跑。Mistral 定位「随处运行」的开放 AI 技术栈：从前沿服务器推理到嵌入式边缘设备。当企业 AI 转向本地优先和边缘部署时，覆盖全算力谱的可信开源模型是美国闭源实验室无法匹敌的护城河。

At AWS re:Invent 2025, Amazon announces Nova 2 (complete generational upgrade) and Kiro — an agentic IDE with browser automation achieving 90%+ reliability. Amazon doubles down on enterprise AI agents across Bedrock.

So what? AWS's agent-first posture is clear: they're not competing for creative AI or chatbot mindshare — they're competing for enterprise workflow automation budgets. Nova 2 powers Bedrock; Kiro takes aim at GitHub Copilot. Amazon's strategy is infrastructure + IDE + models: control all three layers and lock in enterprise developers.



AWS re:Invent 2025 上，Amazon 发布 Nova 2（全代升级）和 Kiro——带浏览器自动化的代理 IDE，可靠性超 90%。Amazon 全力押注企业 AI 代理。AWS 的代理优先姿态清晰：不争创意 AI 或聊天机器人心智——争的是企业工作流自动化预算。Nova 2 驱动 Bedrock，Kiro 瞄准 GitHub Copilot。基础设施 + IDE + 模型三层控制，锁定企业开发者。

BMW's Leipzig plant launched a pilot program deploying AEON humanoid robots for automotive manufacturing tasks. The trial tested humanoids on real production lines for assembly assistance, part handling, and quality inspection — tasks previously requiring specialized machinery or human dexterity. BMW's entry into humanoid robot deployment validated the automotive industry's embrace of general-purpose robot workers as a near-term manufacturing reality.



BMW 莱比锡工厂启动试点项目，部署 AEON 人形机器人执行汽车制造任务。在真实产线上测试装配辅助、零件搬运和质量检测。BMW 入局验证了汽车行业将通用机器人作为近期制造现实的决心。

Google launched Gemini 3 on November 18, 2025, its most capable AI model family to date, available from day one across Search, the Gemini app, AI Studio, Vertex AI, and the Gemini CLI. The flagship Gemini 3 Pro delivered state-of-the-art performance across reasoning, multimodal understanding, coding, and science benchmarks, with native support for text, images, audio, video, and code. Google also introduced Gemini 3 Deep Think, an enhanced reasoning mode pushing the limits of scientific and engineering problem-solving. The launch marked Google's most significant AI model release since Gemini 1.0, closing the perceived capability gap with OpenAI and Anthropic.



Google 上线 Gemini 3，第一天即覆盖 Search、Gemini 应用、AI Studio、Vertex AI 和 Gemini CLI。旗舰 Gemini 3 Pro 在推理、多模态理解、编程和科学基准上均达最优，原生支持文本、图片、音频、视频和代码。还推出 Gemini 3 Deep Think 增强推理模式。这是 Google 自 Gemini 1.0 以来最重大的 AI 模型发布，缩小了与 OpenAI 和 Anthropic 的感知差距。

Baidu unveils ERNIE 5.0 at Baidu World 2025 — a natively omnimodal foundation model that jointly models text, images, audio, and video. Baidu announces global expansion push.

So what? Baidu's ERNIE 5.0 natively trains on all modalities simultaneously rather than bolting them together post-hoc. In China's enterprise AI market, Baidu retains significant distribution through its search and cloud businesses. ERNIE 5.0 signals Baidu is still competing at the frontier, even as DeepSeek and ByteDance eat its consumer share.



百度在百度世界 2025 上亮相 ERNIE 5.0——原生全模态基础模型，联合建模文本、图片、音频和视频。同时宣布全球扩张计划。ERNIE 5.0 在所有模态上同时训练，而非事后拼接。在中国企业 AI 市场，百度靠搜索和云业务保有可观的分发能力。尽管 DeepSeek 和字节跳动侵蚀消费端份额，百度仍在前沿竞争。

OpenAI released GPT-5.1 on November 12, 2025, an upgrade to the GPT-5 series featuring warmer conversational tone, 8 customizable personality options, shopping research capabilities, and enhanced multimodal features. The model family launched in two primary modes: GPT-5.1 Instant for fast everyday tasks and GPT-5.1 Thinking for more complex reasoning. A week later, OpenAI added GPT-5.1-Codex-Max, an agentic coding model designed for extended autonomous software engineering tasks. GPT-5.1 marked OpenAI's first major upgrade cycle to GPT-5, refining the base model based on user feedback.



OpenAI 发布 GPT-5.1，GPT-5 系列的升级版，对话更温暖、8 种可自定义人格、购物调研功能和增强的多模态特性。两种主要模式：GPT-5.1 Instant（日常快速任务）和 GPT-5.1 Thinking（复杂推理）。一周后又加了 GPT-5.1-Codex-Max，面向长时间自主软件工程任务的代理编程模型。这是 GPT-5 的首次重大迭代升级。

OpenAI completed its restructuring as a public benefit corporation (PBC) on October 28, 2025, after years of debate about its unusual nonprofit-controlled hybrid structure. The nonprofit entity was renamed the OpenAI Foundation and became a significant shareholder rather than the controlling entity. Microsoft received approximately 27% of OpenAI Group PBC under the new structure. The conversion cleared the path for a potential future IPO and aligned OpenAI's governance with its commercial scale, while maintaining its stated mission to develop AI for the benefit of humanity.



OpenAI 完成重组，转型为公益公司（PBC）。非营利实体更名为 OpenAI Foundation，成为重要股东而非控制方。Microsoft 在新结构下获得约 27% 股份。转型为未来 IPO 铺路，让治理结构与商业规模匹配，同时维持为人类开发 AI 的使命。

Anthropic released Claude Sonnet 4.5, bringing the hybrid reasoning capabilities first introduced in Opus 4.5 to the mid-tier model tier. Sonnet 4.5 offered improved coding performance, better instruction following, and enhanced multi-turn conversation quality at a lower cost than Opus. The release continued Anthropic's strategy of iterating rapidly across the full model family. Claude Sonnet 4.5 quickly became the most widely used Claude model due to its balance of capability and affordability.



Anthropic 发布 Claude Sonnet 4.5，将 Opus 4.5 引入的混合推理能力带到中端模型。编程能力增强，指令执行更好，多轮对话质量提升，成本低于 Opus。延续 Anthropic 全模型家族快速迭代策略。凭借能力与价格的平衡，迅速成为使用最广的 Claude 模型。

Google releases Gemini 2.5 Flash with native thinking, 1M token context, and aggressive pricing for high-throughput applications.

So what? The "good enough and cheap" model tier gets scary good. For 90% of use cases, you don't need Opus — and that changes the economics of AI products.



Google 发布 Gemini 2.5 Flash，原生思考能力、100 万 token 上下文、高吞吐场景的激进定价。「够好又便宜」的模型层变得可怕地好。90% 的场景不需要 Opus——这改变了 AI 产品的经济学。

Grayscale files Form 10 for a Bittensor Trust, creating a regulated investment vehicle for TAO. Signals institutional interest in decentralized AI infrastructure.



Grayscale 为 Bittensor Trust 提交 Form 10 文件，创建 TAO 的合规投资工具。去中心化 AI 基础设施获得机构投资者关注。

1X Technologies opened pre-orders for NEO, its home humanoid robot designed for everyday consumer use. Unlike industrial humanoids targeting factories, NEO is built for home environments — doing chores, assisting elderly residents, and learning household routines. Priced to reach mass market, NEO's launch represented the first serious attempt to bring a bipedal AI robot into the consumer home market.



1X Technologies 开放 NEO 预购——面向家庭环境的消费级人形机器人。不同于瞄准工厂的工业机器人，NEO 做家务、辅助老人、学习家庭日常。定价面向大众市场。这是首次有公司认真尝试把双足 AI 机器人带进普通家庭。

Virtuals Protocol integrated Coinbase's x402 micropayment protocol, enabling AI agents on the Virtuals platform to autonomously pay for services and receive payments without human intervention. The x402 standard (inspired by HTTP 402 "Payment Required") allows AI agents to transact natively on Base L2, creating the foundation for a machine-to-machine economy where AI agents can independently monetize their actions.



Virtuals Protocol 集成 Coinbase 的 x402 微支付协议，让平台上的 AI 代理无需人工干预即可自主收付款。x402 标准（灵感来自 HTTP 402「需要付款」）让 AI 代理在 Base L2 上原生交易，为机器对机器经济奠定基础——AI 代理可以独立变现自己的行为。

OpenAI released Sora 2 on September 30, 2025, a major upgrade to its video generation model featuring more physically accurate and realistic output, synchronized dialogue, and coherent sound effects. Sora 2 introduced a companion iOS app and significantly improved controllability over camera movements and scene composition. The model showed dramatic improvements in maintaining temporal consistency and generating complex multi-character interactions. Alongside the model launch, OpenAI introduced storyboard features allowing frame-by-frame creative control over generated videos.



OpenAI 发布 Sora 2，视频生成模型大升级——物理效果更真实、支持同步对话和连贯音效。推出 iOS 伴侣应用，镜头运动和场景构图的可控性大幅提升。时间一致性和多角色复杂交互均有显著改进。还推出分镜功能，支持逐帧创意控制。

Figure AI begins mass production of Figure 02 humanoid robots for BMW manufacturing facilities.

So what? Physical AI goes from demo to deployment. The labor market disruption timeline just accelerated. Robotics + AI = the next trillion-dollar intersection.



Figure AI 开始量产 Figure 02 人形机器人，供应 BMW 制造工厂。物理 AI 从演示走向部署，劳动力市场颠覆的时间线加速。机器人 + AI = 下一个万亿美元交汇点。

ByteDance's Doubao surpasses all competitors to become China's most-used AI chatbot, reaching 155 million weekly active users according to QuestMobile data from late 2025 — while Volcano Engine, ByteDance's enterprise tech arm, targets RMB 25B+ revenue in 2025.

So what? ByteDance already cracked distribution once with TikTok/Douyin. Applying the same playbook to AI — deep TikTok-native integration, aggressive monetization through Volcano Engine, and relentless product iteration — works. Doubao's user lead reflects ByteDance's systemic advantage: it doesn't just have an AI product, it has China's largest content distribution network.



字节跳动的豆包超越所有竞品，成为中国使用最多的 AI 聊天机器人，周活达 1.55 亿。火山引擎 2025 年目标营收 ¥250 亿+。字节用 TikTok/抖音的分发套路做 AI 同样奏效——深度抖音原生集成、激进的火山引擎变现、和不停的产品迭代。豆包的用户领先反映了字节的系统性优势：它不只有 AI 产品，还有中国最大的内容分发网络。

OpenAI's Agents SDK reaches general availability, enabling developers to build production-grade AI agents with tool use, handoffs, and guardrails.

So what? The agent framework wars heat up. OpenAI, Anthropic, LangChain all competing — whoever wins the developer platform wins the next decade.



OpenAI Agents SDK 正式发布，让开发者构建生产级 AI 代理，支持工具调用、任务交接和安全护栏。代理框架战争升温，OpenAI、Anthropic、LangChain 全面竞争——谁赢得开发者平台，谁赢得下一个十年。

President Trump signed the GENIUS Act (Guiding and Establishing National Innovation for US Stablecoins Act) into law, creating the first comprehensive federal regulatory framework for US dollar-pegged stablecoins. The law established licensing requirements, reserve standards, and consumer protections — a landmark step in integrating digital assets into the mainstream financial system and directly relevant to AI payment infrastructure.



Trump 总统签署 GENIUS Act，建立美国首个全面的联邦稳定币监管框架。该法律制定了牌照要求、储备标准和消费者保护——将数字资产纳入主流金融体系的里程碑，也与 AI 支付基础设施直接相关。

OpenAI launches GPT-5, unifying the o-series reasoning with GPT fluency into a single model with native tool use and image generation.

So what? The convergence model — reasoning + creativity + tools in one. OpenAI reclaims the throne, at least temporarily. The bar keeps rising.



OpenAI 上线 GPT-5，将 o 系列推理能力与 GPT 的流畅性统一到单一模型中，原生支持工具调用和图像生成。融合模型——推理 + 创造力 + 工具合为一体。OpenAI 至少暂时夺回王座。标杆持续攀升。

OpenAI launched ChatGPT agent mode on July 17, 2025, integrating Operator's web browsing and computer-use capabilities directly into ChatGPT for Pro, Plus, and Team subscribers. ChatGPT can now proactively choose from a toolbox of agentic skills — browsing the web, running code, managing files — to complete complex multi-step tasks autonomously. The release effectively retired Operator as a standalone product, folding its capabilities into the main ChatGPT interface. This marked OpenAI's decisive move to transform ChatGPT from a chatbot into a fully capable autonomous AI agent.



OpenAI 上线 ChatGPT 代理模式，将 Operator 的网页浏览和电脑操控能力直接集成到 ChatGPT 中。ChatGPT 现在能主动从工具箱中选择技能——浏览网页、跑代码、管理文件——自主完成复杂多步任务。Operator 作为独立产品退役，能力并入 ChatGPT 主界面。OpenAI 将 ChatGPT 从聊天机器人转变为全能 AI 代理的决定性一步。

Moonshot AI releases Kimi K2, a 1 trillion parameter open-weight MoE model — one of the largest openly available models in existence. Achieves strong agentic task completion and tool-use benchmarks.

So what? Trillion-parameter open-weight models were once thought impossible outside of hyperscaler infrastructure. Kimi K2 rewrote that assumption. For the open-source community, it's the equivalent of Meta's Llama 3.1 405B moment — but orders of magnitude larger. Any researcher or startup can now run trillion-parameter intelligence locally or on cheap cloud infra.



月之暗面发布 Kimi K2，1 万亿参数开源 MoE 模型——现有最大的开放模型之一。代理任务和工具调用基准表现强势。万亿参数开源模型曾被认为在超大规模基础设施之外不可能实现。Kimi K2 改写了这个假设。

Anthropic raises additional funding from Amazon, pushing its valuation to $60B as the AI safety lab scales.

So what? Safety sells — Anthropic proves you can be responsible AND commercially massive. The safety premium is real for enterprise deals.



Anthropic 获得 Amazon 追加投资，估值推至 $60B。安全能卖钱——Anthropic 证明负责任和商业大成功可以兼得。安全溢价在企业订单中是真实存在的。

World (formerly Worldcoin) launches World ID 2.0 with verified human credentials, expanding to 50+ countries.

So what? Proof of humanity becomes critical as AI-generated content floods the internet. Crypto's answer to the deepfake problem — the intersection thesis plays out.



World（原 Worldcoin）上线 World ID 2.0，验证人类身份凭证，扩展至 50+ 个国家。AI 生成内容泛滥的时代，人类身份证明变得至关重要。Crypto 给深度伪造问题的答案——交叉主题论点在兑现。

Anthropic released Claude Opus 4.5, introducing hybrid reasoning — the ability to seamlessly switch between fast response mode and extended thinking mode within the same context. This allowed the model to reason deeply on complex problems without requiring separate "thinking" prompts. Claude Opus 4.5 achieved state-of-the-art scores on graduate-level reasoning benchmarks and demonstrated significant improvements in multi-step agentic tasks. The hybrid reasoning architecture became a new paradigm for frontier AI model design.



Anthropic 发布 Claude Opus 4.5，引入混合推理——在同一上下文中无缝切换快速响应和扩展思考模式。不需要单独的「思考」提示就能深度推理复杂问题。在研究生级推理基准上达到最优，多步代理任务显著提升。混合推理架构成为前沿 AI 模型设计的新范式。

Anthropic releases Claude 4 Opus, their most capable model, excelling at creative writing, complex analysis, and multi-step reasoning.

So what? Opus 4 is the first model that genuinely feels like a peer, not a tool. The ceiling for AI-human collaboration just got much higher.



Anthropic 发布 Claude 4 Opus，最强模型，在创意写作、复杂分析和多步推理上表现卓越。Opus 4 是第一个让人感觉像同伴而非工具的模型。AI 与人协作的天花板又高了一大截。

The SB 1047 controversy catalyzes global debate on frontier model liability, open-source carve-outs, and safety threshold enforcement.



SB 1047 争议引爆全球关于前沿模型责任、开源豁免和安全阈值执行的大讨论。

Tesla began deploying Optimus humanoid robots internally across its manufacturing facilities for autonomous factory work. With thousands of units in Fremont's pilot production line and Elon Musk projecting 1 million Optimus units per year at full scale, this marked the first time a major automaker deployed general-purpose humanoid robots for real production tasks. Tesla simultaneously announced Gen 3 design improvements and aggressive 2025-2026 production targets.



Tesla 开始在自家工厂内部部署 Optimus 人形机器人执行自主生产任务。Fremont 试产线上数千台，Elon Musk 预计全面量产后年产百万台 Optimus。这是首次有大型车企将通用人形机器人投入真实生产任务。同时公布 Gen 3 设计改进和激进的 2025-2026 生产目标。

Anthropic releases Claude 4 Sonnet, optimized for agentic workflows and sustained coding tasks with dramatically improved tool use.

So what? The "vibe coding" era gets its best engine yet. AI agents that can actually ship code reliably — this is when dev productivity really compounds.



Anthropic 发布 Claude 4 Sonnet，针对代理工作流和持续编程任务优化，工具调用能力大幅提升。「Vibe Coding」时代迎来最佳引擎。能可靠交付代码的 AI 代理——开发效率真正开始复利增长。

Google positions Gemini as an operating layer with agent workflows in Search, Android, and Workspace. Product strategy shifts from features to autonomous completion.



Google 将 Gemini 定位为操作层，在 Search、Android 和 Workspace 中引入代理工作流。产品策略从功能导向转为自主完成任务。

At Build 2025, Microsoft announces AI agents across Windows, GitHub, Azure, and Microsoft 365 — positioning the company around an "Agentic Web" where AI operates autonomously across enterprise workflows. GitHub Copilot crosses 180 million developer users.

So what? Microsoft is translating its OpenAI partnership into systemic OS-level advantage. When Copilot is in your IDE, your email, your browser, and your cloud — you don't switch tools, you switch modes. The enterprise lock-in flywheel is spinning at full speed.



Build 2025 上，Microsoft 宣布 AI 代理覆盖 Windows、GitHub、Azure 和 Microsoft 365，围绕「Agentic Web」定位——AI 在企业工作流中自主运作。GitHub Copilot 突破 1.8 亿开发者用户。Microsoft 把 OpenAI 合作关系转化为系统级操作系统优势。当 Copilot 在你的 IDE、邮箱、浏览器和云里时——企业锁定飞轮全速运转。

OpenAI releases Codex, a cloud-based software engineering agent powered by codex-1 model. Runs in sandboxed environments, handles multi-file edits, tests, and PRs.



OpenAI 发布 Codex，云端软件工程代理，跑 codex-1 模型。在沙盒环境中运行，能处理多文件编辑、测试和 PR。

The EU AI Act's first obligations take effect, requiring transparency for AI systems and banning certain high-risk applications.

So what? Regulatory fragmentation begins — EU goes strict, US goes light. Companies need two playbooks. Compliance becomes a competitive advantage for some, a moat for others.



欧盟 AI 法案首批义务生效，要求 AI 系统透明化并禁止特定高风险应用。监管碎片化开始——欧盟从严，美国从松。企业需要两套策略。合规成为某些公司的竞争优势，也成为另一些公司的护城河。

Xiaomi releases MiMo-7B, its first open-source reasoning model, which surpasses OpenAI o1-mini and QwQ-32B-Preview on AIME 2024-2025 and LiveCodeBench v5 — at 7B parameters.

So what? When a hardware company best known for affordable smartphones ships a best-in-class reasoning model, it signals AI has become table stakes for any serious tech player. Xiaomi's entry broadens China's AI competition beyond pure-play AI startups. With RMB 200B ($27.8B) in 5-year R&D commitments, hardware companies are becoming AI companies.



小米开源 MiMo-7B，首款推理模型，在 AIME 2024-2025 和 LiveCodeBench v5 上超越 OpenAI o1-mini 和 QwQ-32B-Preview——只有 7B 参数。当一家以平价手机著称的硬件公司能造出顶级推理模型，说明 AI 已成为所有严肃科技玩家的标配。小米的入局让中国 AI 竞争从纯 AI 创业公司扩展到全行业。

OpenAI released o3 and o4-mini on April 16, 2025, the first reasoning models with full tool access — including web browsing, Python execution, image generation, and file analysis — natively built in. o3 became OpenAI's most capable reasoning model, while o4-mini offered a strong balance of speed and performance at lower cost. Both models could "think with images" by cropping and zooming into uploaded visuals during their reasoning process. The launch represented a fundamental convergence of reasoning and tool-use capabilities in a single model.



OpenAI 发布 o3 和 o4-mini，首批原生支持全工具访问的推理模型——网页浏览、Python 执行、图像生成和文件分析全部内置。o3 是 OpenAI 最强推理模型，o4-mini 在速度和性能之间取得平衡。两者都能在推理过程中「用图像思考」——裁剪和放大上传的图片。推理与工具调用在单一模型中根本性融合。

Deepgram launches Aura-2, its enterprise text-to-speech model designed for real-time voice AI applications, with human-like expressiveness and ultra-low latency for production voice agents.

So what? With Nova-2 for STT and Aura-2 for TTS, Deepgram now offers a complete voice AI stack under one API. End-to-end speech-to-speech inference without stitching together multiple vendors means fewer points of failure, lower latency, and better cost economics. Voice AI agents went from prototype to production-ready.



Deepgram 上线 Aura-2，企业级文本转语音模型，为实时语音 AI 应用设计，拟人表达力强、延迟超低。Nova-2 做 STT、Aura-2 做 TTS——Deepgram 在单一 API 下提供完整语音 AI 技术栈。端到端语音推理无需拼接多个供应商。语音 AI 代理从原型走向生产就绪。

OpenAI released GPT-4.1 on April 14, 2025, a new model family in the API optimized for coding, instruction following, and long-context tasks with a 1 million token context window. GPT-4.1 outperformed GPT-4.5 Preview on many key benchmarks while being significantly cheaper and faster, leading to the deprecation of GPT-4.5 Preview. It came in three sizes — GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano — serving different cost-performance tradeoffs. GPT-4.1 quickly became the default model for developers building production applications on the OpenAI API.



OpenAI 发布 GPT-4.1，API 专用模型家族，针对编程、指令执行和长上下文任务优化，100 万 token 上下文窗口。GPT-4.1 在多项关键基准上超越 GPT-4.5 Preview，同时更便宜更快，导致 GPT-4.5 Preview 被弃用。三个版本——GPT-4.1、mini 和 nano——覆盖不同性价比需求。迅速成为开发者构建生产应用的默认选择。

OpenAI closes a $40B funding round led by SoftBank, valuing the company at $300B — the largest private tech raise ever.

So what? The AI arms race has a price tag: $300B. This is either justified by AGI potential or the biggest bubble in tech history. I lean toward the former.



OpenAI 完成 $40B 融资，SoftBank 领投，估值 $300B——史上最大私有科技融资。AI 军备竞赛的标价：$300B。这要么被 AGI 潜力证明合理，要么是科技史上最大泡沫。我倾向前者。

Meta releases Llama 4 Scout (17B active / 109B total MoE) and Maverick, both natively multimodal. Also previews Llama 4 Behemoth — a teacher model claimed to outperform GPT-4.5 on STEM benchmarks.

So what? Meta bets on MoE + multimodal from day one instead of retrofitting vision into text-only models. With Scout deployable on a single H100 GPU, open-source closes the gap further. The Behemoth preview signals Meta's intent to compete at the very top — and give away the fruits publicly.



Meta 发布 Llama 4 Scout（17B 激活 / 109B 总参数 MoE）和 Maverick，均为原生多模态。还预告了 Llama 4 Behemoth——一个在 STEM 基准上声称超越 GPT-4.5 的教师模型。Meta 从第一天就押注 MoE + 多模态，而不是往纯文本模型上硬加视觉。Scout 单卡 H100 就能跑，开源继续拉近差距。

Google launches Gemini 2.5 Pro with native thinking capabilities, 1M token context, and top-tier coding performance.

So what? Google goes all-in on reasoning. The three-way race (OpenAI/Anthropic/Google) is closer than ever. Competition = faster progress.



Google 上线 Gemini 2.5 Pro，原生思考能力、100 万 token 上下文、顶级编程表现。Google 全力押注推理。OpenAI/Anthropic/Google 三方赛跑比以往更胶着，竞争 = 更快进步。

OpenAI released native image generation capabilities within GPT-4o on March 25, 2025, enabling the model to generate and edit images directly in conversation without switching to a separate system. Unlike DALL-E 3, the 4o image generation is tightly integrated with the model's reasoning, allowing it to understand context, follow complex instructions, and use uploaded images as references. The feature went viral immediately for its ability to generate Studio Ghibli-style artwork and create coherent multi-panel scenes. It marked OpenAI's first truly multimodal generation model where vision understanding and image creation are unified.



OpenAI 为 GPT-4o 加入原生图像生成能力，在对话中直接生成和编辑图片，不用切换到单独系统。跟 DALL-E 3 不同，4o 图像生成与模型推理深度集成，能理解上下文、执行复杂指令、参考上传图片。Studio Ghibli 风格作品刷屏走红。这是 OpenAI 首个真正的多模态生成模型——视觉理解和图像创作合为一体。

DeepSeek releases V3.1, a hybrid model combining fast and slow thinking modes. Improved performance in agentic tasks and tool use. Continues China's open-source AI momentum.



DeepSeek 发布 V3.1，混合模型结合快思考和慢思考模式。代理任务和工具调用能力增强。中国开源 AI 势头不减。

At GTC 2025, Jensen Huang unveiled NVIDIA's next two GPU generations: Blackwell Ultra, an enhanced version of the Blackwell architecture with improved memory and compute for AI training workloads; and Vera Rubin, the follow-on architecture combining the Vera CPU with the Rubin GPU for a new era of AI supercomputing. NVIDIA also expanded its NIM microservices ecosystem and announced new Cosmos robotics simulation capabilities. The keynote solidified NVIDIA's roadmap dominance in the AI accelerator market for the coming years.



GTC 2025 上，Jensen Huang 公布 NVIDIA 下两代 GPU 架构：Blackwell Ultra（增强版 Blackwell，更强的内存和算力）和 Vera Rubin（Vera CPU + Rubin GPU，开启 AI 超算新时代）。同时扩展 NIM 微服务生态和 Cosmos 机器人仿真能力。NVIDIA 的路线图统治力持续巩固。

Baidu launches ERNIE 4.5 (native multimodal, starting at RMB 0.004 per 1K tokens) and ERNIE X1 (deep-thinking reasoning model) — while simultaneously making ERNIE Bot free for consumers.

So what? DeepSeek's low-cost shock forced every Chinese AI lab to reprice. Baidu's response was dual: a cheaper multimodal frontier model AND a reasoning model to match the o1/R1 wave. Going free on the consumer app is the right strategic move — ERNIE needs users to stay relevant in China's AI chatbot wars, where Doubao and Kimi are both growing fast.



百度上线 ERNIE 4.5（原生多模态，起价 ¥0.004/千 token）和 ERNIE X1（深度思考推理模型），同时让文心一言消费端免费。DeepSeek 的低价冲击迫使所有中国 AI 实验室重新定价。百度的应对：更便宜的多模态前沿模型加推理模型。消费端免费是正确的战略——ERNIE 需要用户才能在豆包和 Kimi 快速增长的中国 AI 聊天机器人大战中保持相关性。

Google DeepMind announced Gemini Robotics and Gemini Robotics-ER on March 12, 2025, vision-language-action (VLA) models built on Gemini 2.0 designed to give robots the ability to understand, reason about, and act in the physical world. Gemini Robotics can follow open-ended natural language instructions to perform complex manipulation tasks, while the ER (Embodied Reasoning) variant enables deeper spatial and physical understanding. The models demonstrated generalization across diverse robot morphologies and tasks without task-specific retraining. This marked Google DeepMind's first major foray into general-purpose robotics foundation models competitive with Physical Intelligence and Tesla Optimus.



Google DeepMind 发布 Gemini Robotics 和 Gemini Robotics-ER，基于 Gemini 2.0 的视觉-语言-动作（VLA）模型，让机器人能理解、推理并在物理世界中行动。Gemini Robotics 可以遵循开放式自然语言指令执行复杂操作任务，ER 版本具备更深的空间和物理理解能力。无需针对特定任务重新训练即可泛化到不同机器人形态。这是 Google DeepMind 在通用机器人基础模型领域的首次重大布局。

Variational draws attention for research-led AI×Crypto design, emphasizing model-native market infrastructure and autonomous systems.



Variational 凭借研究驱动的 AI×Crypto 设计引发关注，强调模型原生市场基础设施和自主系统。

Virtual Protocol crosses $1B in total value locked as AI agent tokens gain mainstream crypto attention.

So what? Crypto × AI isn't just a narrative anymore — there's real capital flowing. The tokenization of AI agents creates a new asset class worth watching.



Virtual Protocol TVL 突破 $1B，AI 代理代币获得主流 Crypto 关注。Crypto×AI 不再只是叙事——真金白银在涌入。AI 代理的代币化创造了一个值得关注的新资产类别。

GoKite announces toolkit updates for building and deploying crypto-native AI agents, focusing on automation + market workflows.



GoKite 更新工具包，用于构建和部署 Crypto 原生 AI 代理，聚焦自动化和市场工作流。

OpenMind AGI positions itself around decentralized agent coordination and open infrastructure principles, gaining attention in AI×Crypto circles.



OpenMind AGI 围绕去中心化代理协调和开放基础设施原则定位，在 AI×Crypto 圈子获得关注。

Zama expands fully homomorphic encryption tooling for on-chain and AI use cases, enabling private inference and computation withoutTLS-style confidentiality for AI workflows.



Zama 扩展全同态加密工具，覆盖链上和 AI 用例，实现隐私推理和计算——AI 工作流的机密性无需依赖 TLS 式方案。

Anthropic releases Claude Code, an agentic coding tool that lives in your terminal. Can read/edit files, run commands, create PRs. Competes with Cursor and GitHub Copilot.



Anthropic 发布 Claude Code，一款活在终端里的代理式编程工具。能读写文件、运行命令、创建 PR，直接对标 Cursor 和 GitHub Copilot。

Anthropic releases Claude 3.7 Sonnet — a hybrid model that can toggle between fast responses and deep reasoning with extended thinking.

So what? Best of both worlds: quick chat when you need it, deep analysis when you don't. The UX of reasoning models just got much better.



Anthropic 发布 Claude 3.7 Sonnet——混合模型，可以在快速响应和深度推理的扩展思考之间切换。两全其美：需要快聊就快聊，需要深度分析就深度分析。推理模型的用户体验大幅改善。

xAI released Grok-3 on February 17, 2025, trained on the 100,000-GPU Colossus supercluster with approximately 10x the computational resources of prior models. Grok-3 introduced "Think" mode, a chain-of-thought reasoning capability for complex problems, and "DeepSearch" for comprehensive research tasks. Benchmark results showed it competitive with or surpassing GPT-4.5 and Claude 3.7 Sonnet on math, science, and coding tasks. A premium SuperGrok subscription tier was introduced to access Grok-3's advanced features, with the model briefly made available to free X users as well.



xAI 发布 Grok-3，在 10 万块 GPU 的 Colossus 超算集群上训练，算力约为前代 10 倍。新增「Think」思维链推理模式和「DeepSearch」深度研究功能。基准结果显示与 GPT-4.5 和 Claude 3.7 Sonnet 竞争甚至超越。推出 SuperGrok 高级订阅。

OpenAI releases GPT-4.5, a large-scale model focused on reducing hallucinations and improving "EQ" in conversations.

So what? Incremental but meaningful — less hallucination = more enterprise trust. The reliability gap is closing.



OpenAI 发布 GPT-4.5，大规模模型专注于减少幻觉和提升对话中的「情商」。渐进但有意义——幻觉更少意味着企业信任度更高，可靠性差距在缩小。

Bittensor subnet count and specialization accelerate (inference, data, validation markets), making TAO ecosystem core infrastructure in decentralized AI.



Bittensor 子网数量和专业化加速（推理、数据、验证市场），TAO 生态正成为去中心化 AI 的核心基础设施。

OpenAI released o3-mini, a cost-efficient reasoning model delivering strong performance on STEM tasks including math, science, and coding at a fraction of the compute cost of o3. The model offered three reasoning effort levels — low, medium, and high — letting developers balance speed and accuracy. o3-mini outperformed o1 on many STEM benchmarks while being significantly cheaper, making advanced reasoning accessible to a much broader set of developers and applications.



OpenAI 发布 o3-mini，高性价比推理模型，在 STEM 任务上表现强劲，算力消耗远低于 o3。提供低/中/高三档推理力度，让开发者灵活平衡速度和准确性。o3-mini 在多项 STEM 基准上超越 o1，成本大幅降低，让高级推理触达更广泛的开发者和应用。

ai16z officially rebranded to ElizaOS after the real a16z (Andreessen Horowitz) requested the name change due to brand confusion. The project simultaneously announced plans to build its own L1 blockchain optimized for AI agents. The rebranding crystallized the project's identity around its open-source agent framework, which had by this point powered thousands of AI agent deployments across crypto.



ai16z 正式改名 ElizaOS——真正的 a16z（Andreessen Horowitz）因品牌混淆要求更名。项目同时宣布计划构建面向 AI 代理优化的 L1 区块链。更名让项目身份聚焦于其开源代理框架，该框架已驱动数千个 AI 代理部署。

DeepSeek R1's efficiency claims trigger a massive tech selloff — NVIDIA drops 17% in a single day, erasing nearly $600B in market cap.

So what? The "more compute = better AI" thesis gets challenged. If you can train frontier models for 1/10th the cost, the entire AI investment thesis needs re-pricing.



DeepSeek R1 的效率声明引发科技股大抛售——NVIDIA 单日跌 17%，市值蒸发近 $600B。「更多算力 = 更好 AI」的论点遭到挑战。如果前沿模型训练成本降到十分之一，整个 AI 投资逻辑都得重新定价。

Virtuals Protocol announced expansion to Solana, establishing the platform across both Base (Ethereum L2) and Solana ecosystems. The launch included a Meteora liquidity pool, a Strategic SOL Reserve (1% of trading fees converted to SOL), and new builder grants. Previously a Base-only AI agent launchpad that had reached $1B+ TVL, Virtuals' multi-chain expansion signaled the mainstreaming of AI agent tokens.



Virtuals Protocol 宣布扩展到 Solana，横跨 Base 和 Solana 双生态。上线 Meteora 流动性池，设立战略 SOL 储备（交易费 1% 转换为 SOL），并推出新的 builder 补助。之前仅在 Base 上运营、TVL 超 $1B 的 Virtuals 多链扩张标志着 AI 代理代币走向主流。

OpenAI launched Operator on January 23, 2025, a research preview of an AI agent that can browse the web and complete tasks autonomously using its own browser. Powered by the Computer-Using Agent (CUA) model, Operator can fill forms, place orders, and navigate websites without human hand-holding. It was initially available only to ChatGPT Pro users in the US, with a broader rollout planned later. Operator was later fully integrated into ChatGPT in July 2025 as "ChatGPT agent mode," signaling OpenAI's strategic shift toward agentic AI.



OpenAI 上线 Operator，AI 代理可以用自己的浏览器自主浏览网页并完成任务。跑 Computer-Using Agent (CUA) 模型，能填表单、下订单、导航网站。先向美国 ChatGPT Pro 用户开放。后来在 2025 年 7 月完全并入 ChatGPT，成为「ChatGPT 代理模式」。

Chinese lab DeepSeek releases R1, an open-source reasoning model rivaling OpenAI o1 at a fraction of the training cost.

So what? The "China can't compete" narrative dies. Open-source reasoning at 1/10th the cost changes the economics of AI. Every startup just got a gift.



中国实验室 DeepSeek 开源 R1，推理模型媲美 OpenAI o1，训练成本只是零头。「中国做不了」的叙事就此终结。十分之一成本的开源推理改写了 AI 经济学，每个创业公司都收到了一份大礼。

Moonshot AI releases Kimi K1.5, a reasoning-focused model that matches OpenAI o1 on mathematics, coding, and multimodal reasoning benchmarks — delivered just days before DeepSeek-R1 made global headlines.

So what? The week of January 20, 2025 became a proof point that China's AI labs weren't trailing by months — they were parallel. Kimi K1.5 and DeepSeek-R1 arriving simultaneously shattered the narrative that advanced reasoning models required trillion-dollar infrastructure. The reasoning capabilities race went fully global.



月之暗面发布 Kimi K1.5，推理模型在数学、编程和多模态推理基准上追平 OpenAI o1——比 DeepSeek-R1 全球刷屏只早了几天。2025 年 1 月 20 日这周证明中国 AI 实验室不是落后几个月——而是齐头并进。Kimi K1.5 和 DeepSeek-R1 同时到来，打破了「高级推理模型需要万亿美元基础设施」的叙事。

MiniMax open-sources MiniMax-01, a hybrid Mixture-of-Experts model with 1 million token context window, trained with custom reinforcement learning (CISPO). Trained at a fraction of typical frontier costs.

So what? MiniMax didn't just match DeepSeek on cost efficiency — they open-sourced a 1M context model when GPT-4 was limited to 128K. For any application needing full-book or full-codebase context, this was a paradigm shift. Chinese labs were no longer followers; they were setting the bar on long-context architecture.



MiniMax 开源 MiniMax-01，混合 MoE 模型，100 万 token 上下文窗口，用自研强化学习（CISPO）训练，成本远低于前沿水平。当 GPT-4 还限在 128K 时，中国实验室开源了 100 万上下文模型。中国 AI 不再是跟随者——他们在定义长上下文架构的标杆。

io.net expands distributed GPU network and integrations for AI teams needing flexible training/inference capacity without centralized cloud dependence.



io.net 扩展分布式 GPU 网络和集成能力，为 AI 团队提供灵活的训练/推理算力，摆脱中心化云依赖。

At CES 2025, NVIDIA announced two major initiatives: Cosmos, a world foundation model platform for training physical AI and robotics systems using synthetic video data; and Project DIGITS, a $3,000 personal AI supercomputer delivering 1 petaflop of AI performance using the GB10 Grace Blackwell Superchip. Cosmos aimed to accelerate autonomous vehicle and robotics development by generating physically accurate simulation data. Project DIGITS brought AI supercomputer capability to a desktop form factor, democratizing local AI model deployment.



CES 2025 上，NVIDIA 发布两大计划：Cosmos 世界基础模型平台（用合成视频数据训练物理 AI 和机器人系统）和 Project DIGITS（$3,000 个人 AI 超级计算机，搭载 GB10 Grace Blackwell 超级芯片，1 petaflop AI 算力）。Cosmos 加速自动驾驶和机器人开发，Project DIGITS 把 AI 超算搬上桌面。

Virtuals Protocol reached its all-time high with an 850%+ rally from October 2024, becoming the leading AI agent tokenization platform on Base. The protocol enables anyone to create, tokenize, and deploy AI agents, with over $1B in cumulative agent creation volume.



Virtuals Protocol 创下历史新高，从 2024 年 10 月起涨超 850%，成为 Base 上领先的 AI 代理代币化平台。任何人都可以创建、代币化和部署 AI 代理，累计代理创建量超 $1B。

DeepSeek releases V3 on Christmas day. Trained for ~$5.5M (vs hundreds of millions for US frontier models). MoE architecture with 671B parameters. Tops multiple benchmarks. Signals China can compete efficiently despite chip export controls.



DeepSeek 在圣诞节发布 V3，训练成本约 $5.5M（美国前沿模型动辄数亿美元）。671B 参数 MoE 架构，多项基准登顶。中国在芯片出口管制下依然能高效竞争。

OpenAI announces o3, achieving 87.5% on ARC-AGI benchmark (previous best: ~50%). Also 25.2% on Frontier Math (vs <2% prior). Significant step toward general reasoning.



OpenAI 公布 o3，在 ARC-AGI 基准上达到 87.5%（此前最高约 50%），Frontier Math 上 25.2%（此前不到 2%）。通用推理能力迈出重要一步。

DeepMind upgrades video and image generation stack with Veo 2 and Imagen 3 improvements, improving motion consistency, style control, and safety filtering.



DeepMind 升级视频和图像生成技术栈，Veo 2 和 Imagen 3 在运动一致性、风格控制和安全过滤上均有改进。

Google introduces Gemini 2.0 Flash, optimized for fast multimodal responses and agentic tool use. Marks transition from chat model to action model.



Google 推出 Gemini 2.0 Flash，针对快速多模态响应和代理式工具调用优化。从聊天模型向行动模型转型。

OpenAI releases Sora to the public — AI-generated video from text prompts, with surprisingly coherent results.

So what? Content creation gets another earthquake. Short-form video, ads, prototyping — all disrupted. The "fake vs real" problem intensifies.



OpenAI 向公众开放 Sora——用文本生成 AI 视频，效果出人意料地连贯。内容创作迎来又一场地震，短视频、广告、原型设计全被颠覆。「真假难辨」的问题加剧。

OpenAI launched ChatGPT Pro, a premium $200/month subscription tier offering unlimited access to its most advanced models including o1 Pro Mode. The Pro tier was designed for power users requiring the highest performance with no rate limits. It included access to OpenAI's extended thinking capabilities via o1 Pro, which used significantly more compute to solve harder problems. The launch signaled OpenAI's move toward a multi-tier commercial strategy targeting professionals and enterprise users.



OpenAI 上线 ChatGPT Pro，$200/月高端订阅，无限使用最强模型，包括 o1 Pro Mode。o1 Pro 调用大量算力解决更难的问题。Pro 套餐标志着 OpenAI 面向专业用户和企业的多层商业策略。

At AWS re:Invent 2024, Amazon launches the Nova family — Micro, Lite, Pro, and Canvas/Reel — its first proprietary frontier models trained in-house and offered via Bedrock.

So what? Amazon had been the "neutral platform" in AI, just selling Anthropic and others. Nova signals the transition: AWS is now a model builder, not just a cloud provider. With Bedrock as the distribution channel and Trainium as the silicon stack, Amazon is betting it can compete from cloud all the way down to silicon. That's a vertically integrated threat to OpenAI's enterprise moat.



AWS re:Invent 2024 上，Amazon 推出 Nova 家族——Micro、Lite、Pro 和 Canvas/Reel——首批自研前沿模型，通过 Bedrock 分发。Amazon 从「中立平台」变身「模型制造商」。Bedrock 做分发、Trainium 做芯片——从云到硅的垂直整合对 OpenAI 的企业护城河构成威胁。

Render ecosystem updates focus on GPU marketplace utility for AI rendering and training workloads, reinforcing decentralized compute narrative.



Render 生态更新聚焦 GPU 市场在 AI 渲染和训练工作负载上的效用，强化去中心化算力叙事。

AWS introduces Trainium2 generation to challenge Nvidia GPU dependency and attract large-scale foundation model training workloads.



AWS 推出 Trainium2，挑战对 Nvidia GPU 的依赖，吸引大规模基础模型训练工作负载。

Anthropic releases MCP, an open protocol for connecting AI assistants to external data sources and tools. Adopted by Cursor, Replit, and others. Becomes de facto standard for AI tool integration.



Anthropic 发布 MCP，一个连接 AI 助手与外部数据源和工具的开放协议。Cursor、Replit 等迅速采用，成为 AI 工具集成的事实标准。

Kimi continues long-context positioning in China market with upgrades focused on document intelligence, research tasks, and productivity use cases.



Kimi 继续在中国市场强化长上下文定位，聚焦文档智能、研究任务和生产力场景。

Donald Trump wins the 2024 presidential election, signaling a pro-innovation, deregulation approach to AI policy.

So what? Expect lighter AI regulation in the US. Good for AI companies' velocity, but the safety crowd loses their loudest ally in government.



Trump 赢得 2024 年大选，释放亲创新、放松监管的 AI 政策信号。美国 AI 监管预计会更宽松，利好 AI 公司的速度，但安全派在政府中失去了最大声音。

OpenAI introduced ChatGPT Search on October 31, 2024, integrating real-time web search directly into ChatGPT with cited sources and up-to-date results. The feature challenged traditional search engines by delivering conversational, context-aware answers with direct links. It initially rolled out to ChatGPT Plus subscribers, then expanded to all logged-in users in December 2024, and to everyone in February 2025. ChatGPT Search positioned OpenAI as a direct competitor to Google Search for the first time.



OpenAI 推出 ChatGPT Search，直接在 ChatGPT 里集成实时网页搜索，带引用来源和最新结果。先向 Plus 用户开放，12 月扩展至所有登录用户，2025 年 2 月向所有人开放。OpenAI 首次直接挑战 Google 搜索。

Physical Intelligence raises $400M at $2.4B valuation from Bezos, Thiel, OpenAI. Building pi0 — a foundation model for general-purpose robot control.



Physical Intelligence 完成 $400M 融资，估值 $2.4B，Bezos、Thiel、OpenAI 参投。在造 pi0——通用机器人控制基础模型。

US AI Safety Institute formalizes testing and evaluation direction, signaling that policy competition is moving from principles to measurable standards.



美国 AI 安全研究所确定测试和评估方向，信号表明政策竞争正从原则层面走向可量化标准。

GitHub extends Copilot from inline completion to multi-step planning/execution with Copilot Workspace.



GitHub 将 Copilot 从行内补全扩展到多步规划/执行，推出 Copilot Workspace。

ai16z — named as a parody of a16z — launched on Solana as the first AI-managed decentralized venture fund. Built by Shaw Walters, the project's AI agent "Eliza" makes investment decisions autonomously, using the open-source ElizaOS framework. Within weeks it became one of the largest AI tokens by market cap and spawned an ecosystem of thousands of AI agents. The ElizaOS framework became the de facto standard for building crypto-native AI agents.



ai16z（戏仿 a16z 命名）在 Solana 上线，首个 AI 管理的去中心化风投基金。AI 代理「Eliza」自主做投资决策，基于开源 ElizaOS 框架。数周内成为市值最大的 AI 代币之一，催生了数千个 AI 代理的生态。ElizaOS 成为构建 Crypto 原生 AI 代理的事实标准。

Anthropic released an upgraded Claude 3.5 Sonnet alongside the new Claude 3.5 Haiku, and introduced Computer Use — a groundbreaking capability allowing Claude to control a computer by viewing screenshots, moving the cursor, clicking, and typing. This marked the first time a widely available AI model could directly interact with a computer's graphical interface. Computer Use opened the door to autonomous agent workflows, though Anthropic cautioned it was an early, experimental feature. The update also included an improved Claude.ai interface.



Anthropic 发布升级版 Claude 3.5 Sonnet 和新的 Claude 3.5 Haiku，同时推出 Computer Use——让 Claude 可以通过查看截图、移动鼠标、点击和打字来操控电脑。这是首次有广泛可用的 AI 模型能直接操作图形界面。Computer Use 为自主代理工作流打开大门，Anthropic 标注这仍是早期实验性功能。

Zhipu introduces GLM-4-Plus with stronger coding and Chinese reasoning performance, expanding domestic alternatives for enterprise AI workloads.



智谱推出 GLM-4-Plus，编程和中文推理能力增强，为企业 AI 工作负载扩展国产替代方案。

An AI agent called Truth Terminal, created by researcher Andy Ayrey, inadvertently catalyzed the $GOAT memecoin on Solana, which surged to over $800M market cap. The event ignited the AI agent token narrative and is widely considered the starting point of the crypto×AI agent mania.



研究者 Andy Ayrey 创建的 AI 代理 Truth Terminal 无意间催生了 Solana 上的 $GOAT 迷因币，市值飙至 $800M 以上。这一事件点燃了 AI 代理代币叙事，被广泛视为 Crypto×AI 代理狂潮的起点。

The 2024 Nobel Prize in Chemistry was split between David Baker (for computational protein design) and Demis Hassabis & John Jumper (for AlphaFold's protein structure prediction). AlphaFold solved a 50-year grand challenge in biology and has accelerated drug discovery globally. Two Nobel Prizes for AI in two consecutive days underscored the technology's transformative scientific impact.



2024 年诺贝尔化学奖由 David Baker（计算蛋白质设计）和 Demis Hassabis、John Jumper（AlphaFold 蛋白质结构预测）分享。AlphaFold 解决了生物学 50 年的重大挑战，加速全球药物发现。连续两天两个诺贝尔奖颁给 AI——技术的变革性影响不言而喻。

The Royal Swedish Academy of Sciences awarded the 2024 Nobel Prize in Physics jointly to John Hopfield and Geoffrey Hinton "for foundational discoveries and inventions that enable machine learning with artificial neural networks." The award — surprising many physicists — validated decades of work on neural nets and marked AI's formal recognition by the scientific establishment. Hinton had left Google just months earlier warning of AI's dangers.



2024 年诺贝尔物理学奖授予 John Hopfield 和 Geoffrey Hinton，表彰他们「在利用人工神经网络实现机器学习方面的奠基性发现和发明」。这个奖让很多物理学家意外——AI 获得了科学最高殿堂的正式认可。Hinton 几个月前刚从 Google 辞职警告 AI 的危险。

OpenAI launched Canvas, a new collaborative interface within ChatGPT designed for writing and coding projects that go beyond simple chat. Canvas opens in a separate window alongside the chat, giving users a persistent workspace where they can make targeted edits, run Python code, and iterate on documents or code with AI assistance. The feature was initially rolled out to ChatGPT Plus and Team users before becoming available to all users in December 2024. Canvas represented OpenAI's push to make ChatGPT a full-fledged creative and development co-pilot.



OpenAI 推出 Canvas，ChatGPT 里的协作式创作界面，专为写作和编程项目设计。Canvas 在聊天旁打开独立窗口，用户可以在持久工作空间里做定向编辑、跑 Python 代码、和 AI 一起迭代文档或代码。OpenAI 在把 ChatGPT 打造成全功能的创作和开发副驾驶。

OpenAI closed a $6.6 billion funding round led by Thrive Capital, valuing the company at $157 billion — making it one of the most valuable private companies in the world. The round included investments from Microsoft, NVIDIA, SoftBank, and others. The capital was earmarked for compute infrastructure, research, and continued product development. The raise came as OpenAI was exploring a transition from its nonprofit-controlled structure to a for-profit benefit corporation, raising governance questions across the AI industry.



OpenAI 完成 $6.6B 融资，Thrive Capital 领投，估值 $157B——全球最有价值的私有公司之一。Microsoft、NVIDIA、SoftBank 等参投。资金用于算力基础设施、研究和产品开发。此轮融资正值 OpenAI 探索从非营利控制结构转型为营利性公益公司。

OpenAI launched the Realtime API in public beta on October 1, 2024, enabling developers to build low-latency, speech-to-speech conversational experiences powered by GPT-4o. The API supports six preset voices and handles interruptions natively, enabling more natural human-AI dialogue. It unlocks a new class of voice agent applications without requiring developers to stitch together separate ASR, LLM, and TTS pipelines. The Realtime API marked OpenAI's first developer-facing product for building persistent, interactive voice agents at scale.



OpenAI 公测 Realtime API，让开发者用 GPT-4o 构建低延迟语音对话体验。支持 6 种预设音色，原生处理打断。不用再拼接 ASR、LLM、TTS 三条管线——OpenAI 首次面向开发者推出构建持久交互式语音代理的产品。

Hugging Face surpasses 1 million public AI model listings on its Hub, cementing its role as the de facto repository of open-source AI. Growing from 300K models in March 2024 to 1M in September.

So what? GitHub moment for AI. When a platform has 1M models, it stops being a tool and becomes infrastructure — every company building AI touches Hugging Face. The exponential growth reflects open-source AI's commoditization: what required frontier labs in 2023 is now community-built and freely shared. Anyone building closed systems is fighting an increasingly large open tide.



Hugging Face Hub 上的公开 AI 模型突破 100 万个，从 3 月的 30 万到 9 月的 100 万。当一个平台有 100 万模型时，它不再是工具——它是基础设施。开源 AI 的指数级增长说明了一切：2023 年需要前沿实验室做的事，现在社区就能做。

Meta released Llama 3.2 at Meta Connect on September 25, 2024, introducing multimodal vision capabilities to the Llama model family for the first time. The release included 11B and 90B vision-enabled models alongside lightweight 1B and 3B models optimized for edge and mobile deployment. Llama 3.2 could process both text and images, enabling tasks like visual question answering and document analysis. The lightweight models were designed to run on-device, including on Meta's smart glasses and smartphones.



Meta 在 Meta Connect 上发布 Llama 3.2，首次为 Llama 家族加入多模态视觉能力。包含 11B 和 90B 视觉模型，以及优化边缘和移动端部署的 1B 和 3B 轻量模型。能处理文本和图片，支持视觉问答和文档分析。轻量版可以在手机和智能眼镜上本地运行。

Alibaba releases Qwen 2.5 family: 10 model sizes from 0.5B to 72B parameters, all open-weight. Strong multilingual performance, particularly in Chinese. Includes code and math variants.



阿里巴巴发布 Qwen 2.5 家族，10 个尺寸（0.5B 到 72B），全部开源。多语言表现强势，尤其中文。包含代码和数学专用版本。

Agility scales Digit deployments in warehouse/logistics scenarios, proving near-term enterprise humanoid use cases.



Agility 扩大 Digit 在仓储物流场景的部署，验证了人形机器人的近期企业应用场景。

OpenAI releases o1, a model trained with reinforcement learning to "think before answering," excelling at math, coding, and science.

So what? The paradigm shifts from "next token prediction" to "chain-of-thought reasoning." This is the beginning of System 2 thinking in AI — a step toward AGI.



OpenAI 发布 o1，用强化学习训练模型「先想再答」，在数学、编程和科学上表现突出。范式从「预测下一个 token」转向「思维链推理」——AI 的 System 2 思维起步，迈向 AGI 的一步。

Akash expands AI-focused cloud marketplace features and partner integrations, strengthening decentralized alternatives to hyperscaler GPU pricing.



Akash 扩展面向 AI 的去中心化 GPU 云市场功能和合作伙伴集成，为 AI 团队提供灵活的训练/推理算力，替代中心化云服务商的高价 GPU。

MiniMax launches video-01 under its Hailuo AI brand — a high-quality text-to-video generation model that goes viral globally for its cinematic output quality, especially character consistency.

So what? While the West debated Sora's restricted access, China shipped. Hailuo AI's viral demos showed that video generation quality was no longer a US-exclusive capability. MiniMax proved it could build multimodal creative AI at consumer scale — and ship it openly to the public while OpenAI kept Sora behind waitlists.



MiniMax 以 Hailuo AI 品牌上线 video-01 文生视频模型，电影级输出质量迅速全球走红，角色一致性尤其突出。西方还在等 Sora 开放时，中国直接出货了。Hailuo AI 证明视频生成能力不再是美国独有。

DeepSeek released DeepSeek-V2.5 in September 2024, merging the capabilities of DeepSeek-V2-Chat and DeepSeek-Coder-V2 into a unified model with improved general and coding abilities. The combined model delivered better alignment, stronger reasoning, and enhanced coding performance compared to running separate V2 variants. DeepSeek-V2.5 was subsequently revised in December 2024. The release demonstrated DeepSeek's rapid iteration pace in producing best-in-class open-source models.



DeepSeek 发布 V2.5，将 V2-Chat 和 Coder-V2 合并为统一模型，通用和编程能力同步提升。对齐效果更好，推理更强，编程表现增强。12 月又做了修订版。展现了 DeepSeek 在开源模型领域极快的迭代节奏。

1X presents NEO beta direction for home environments, emphasizing safe human-robot interaction and embodied AI.



1X 展示 NEO beta 版，面向家庭环境设计，强调安全的人机交互和具身 AI。

xAI released Grok-2 on August 14, 2024, a major upgrade featuring multimodal vision understanding and access to Aurora, xAI's image generation model, making it one of the first commercial chatbots with integrated text-to-image generation. Grok-2 demonstrated competitive performance with GPT-4o and Claude 3.5 Sonnet on reasoning and knowledge benchmarks. It was made available to X Premium and Premium+ subscribers, with a more efficient Grok-2 mini variant released alongside it. The launch also marked xAI's first push toward making Grok a standalone product beyond just an X platform feature.



xAI 发布 Grok-2，支持多模态视觉理解，集成 Aurora 图像生成模型——首批自带文生图功能的商用聊天机器人之一。在推理和知识基准上与 GPT-4o 和 Claude 3.5 Sonnet 竞争。面向 X Premium 和 Premium+ 用户开放，同时上线更高效的 Grok-2 mini。

NEAR Protocol announced a strategic pivot to become the chain for user-owned AI, launching NEAR AI with plans for AI assistants, agent infrastructure, and decentralized AI model training. Co-founder Illia Polosukhin (Transformer paper co-author) positioned NEAR as the natural bridge between AI and crypto.



NEAR Protocol 宣布战略转型为用户自主 AI 的底层链，推出 NEAR AI 计划，涵盖 AI 助手、代理基础设施和去中心化 AI 模型训练。联合创始人 Illia Polosukhin（Transformer 论文合著者）将 NEAR 定位为 AI 与 Crypto 之间的天然桥梁。

NEAR Protocol formalized its strategic pivot to become the AI layer for Web3, emphasizing chain abstraction and AI-native decentralized applications. The foundation announced a $20M fund to accelerate AI development on NEAR. The strategy repositioned NEAR from a general L1 blockchain to a specialized platform for building AI agents and autonomous on-chain applications.



NEAR Protocol 正式战略转向，定位为 Web3 的 AI 层，强调链抽象和 AI 原生去中心化应用。基金会宣布 $20M 专项资金加速 NEAR 上的 AI 开发。NEAR 从通用 L1 公链重新定位为 AI 代理和链上自主应用的专用平台。

Mistral AI releases Mistral Large 2 (123B), competing with GPT-4o and Llama 3.1 on coding and reasoning.

So what? Europe has a serious AI contender. Geographic diversification of AI capabilities matters for geopolitics and regulatory arbitrage.



Mistral AI 发布 Mistral Large 2（123B），在编程和推理上对标 GPT-4o 和 Llama 3.1。欧洲有了真正的前沿 AI 竞争者，AI 能力的地理多元化对地缘政治和监管套利意义重大。

Meta releases Llama 3.1 405B — 15 trillion tokens, 16,000 GPUs, and the first open-weight model to credibly match GPT-4 on real benchmarks. Available as API and for local deployment.

So what? This is the moment open-source caught up to the frontier. Any lab, researcher, or startup can now fine-tune a GPT-4 tier model without paying OpenAI. It didn't just open a door — it demolished a wall. The "open vs. closed" debate in AI permanently shifted, and every enterprise that cared about data privacy suddenly had a legitimate alternative.



Meta 开源 Llama 3.1 405B——150 万亿 token 训练、16,000 块 GPU，首个在真实基准上追平 GPT-4 的开源模型。开源追上前沿的时刻。任何实验室、研究者或创业公司都能微调 GPT-4 级模型而不用给 OpenAI 付费。「开源 vs 闭源」的辩论永久性转向。

OpenAI released GPT-4o mini, a highly cost-efficient small model designed to replace GPT-3.5 Turbo across most use cases. At $0.15 per million input tokens, it was far cheaper than previous models while outperforming GPT-4 Turbo on key benchmarks. GPT-4o mini supported both text and vision inputs and featured a 128K context window. The release democratized access to powerful AI capabilities and became the default model for high-volume API applications.



OpenAI 发布 GPT-4o mini，高性价比小模型，替代 GPT-3.5 Turbo。输入价格 $0.15/百万 token，在关键基准上超越 GPT-4 Turbo。支持文本和视觉输入，128K 上下文窗口。成为高流量 API 应用的默认选择。

Pittsburgh-based Skild AI emerged from stealth with a $300M Series A — the largest robotics AI round of 2024 — led by Lightspeed, Coatue, SoftBank, and Jeff Bezos. Skild is building the world's first general-purpose robotic foundation model: one brain that can control any robot across any task. At a $1.5B valuation, it signaled that the robotics AI race had entered a new capital-intensive phase.



匹兹堡 Skild AI 以 $300M A 轮融资走出隐身模式——2024 年机器人 AI 最大单轮。Lightspeed、Coatue、SoftBank 和 Jeff Bezos 参投。Skild 在造世界首个通用机器人基础模型：一个大脑控制任何机器人执行任何任务。$1.5B 估值标志着机器人 AI 竞赛进入烧钱新阶段。

Google released Gemma 2 on June 27, 2024, with significantly improved performance over the original Gemma models. Available in 9B and 27B parameter sizes, Gemma 2 demonstrated competitive performance with much larger models while remaining efficient enough to run on a single GPU or TPU. The models featured improvements in reasoning, code generation, and multilingual capabilities. Gemma 2 became the preferred open-weights baseline for researchers evaluating efficient AI architectures.



Google 发布 Gemma 2，9B 和 27B 版本，性能比初代大幅提升。单卡 GPU 或 TPU 就能跑，推理、代码和多语言能力全面增强。Gemma 2 成为研究者评估高效 AI 架构时的首选开源基线。

Anthropic releases Claude 3.5 Sonnet, outperforming Claude 3 Opus at Sonnet-tier pricing and speed.

So what? The mid-tier model beats the previous flagship — this pattern of rapid capability jumps becomes the new normal. Cost-per-intelligence is in freefall.



Anthropic 发布 Claude 3.5 Sonnet，以 Sonnet 级别的价格和速度超越了上一代旗舰 Claude 3 Opus。中端模型打败前代旗舰——能力跳跃式提升成为新常态，智能的单位成本在自由落体。

Nvidia surpasses Microsoft and Apple with $3.34T market cap, driven by insatiable AI chip demand. H100/A100 GPUs power virtually all frontier model training.



Nvidia 以 $3.34T 市值超越 Microsoft 和 Apple，登顶全球最有价值公司。AI 芯片需求无止境，H100/A100 驱动着几乎所有前沿模型训练。

Three major AI crypto projects merge tokens into the Artificial Superintelligence Alliance (ASI). Combined entity becomes largest decentralized AI network by market cap.



三大 AI Crypto 项目合并代币组成 Artificial Superintelligence Alliance (ASI)，成为市值最大的去中心化 AI 网络。

Sequoia Capital publishes analysis showing AI companies need to generate $600B in revenue to justify current infrastructure spend.

So what? The first serious "show me the money" moment for AI. Reminds me of the 2000 dot-com reality check — but this time the tech actually works.



Sequoia Capital 发布分析称 AI 公司需要产生 $600B 收入才能匹配当前基础设施投入。AI 的第一个「给我看到钱」时刻。像 2000 年互联网泡沫的现实检验——不过这次技术真的能用。

Stability AI released Stable Diffusion 3 Medium on June 12, 2024, making the open-weights version of its next-generation text-to-image model publicly available after months of enterprise preview. SD3 Medium uses a novel Multimodal Diffusion Transformer (MMDiT) architecture that dramatically improves text rendering within images and multi-subject scene composition. The open release democratized access to state-of-the-art image generation, with larger variants (SD3 Large and SD3 Ultra) available via the API. SD3 Medium was followed by Stable Diffusion 3.5 in October 2024, cementing Stability AI's position as the leading open-source image generation platform.



Stability AI 开源 Stable Diffusion 3 Medium，采用全新 Multimodal Diffusion Transformer (MMDiT) 架构，文字渲染和多主体场景构图大幅提升。更大版本（SD3 Large/Ultra）通过 API 开放。10 月又发布 SD 3.5，巩固了 Stability AI 在开源图像生成领域的领先地位。

Mistral AI closes €600M (~$645M) Series B, bringing valuation to €5.8B ($6.2B). Backed by Andreessen Horowitz, General Catalyst, and others — cementing it as Europe's most valuable AI lab.

So what? In just 12 months from founding, Mistral became Europe's flagship AI champion. The raise wasn't just money — it was a geopolitical statement: Europe has a credible alternative to US hyperscalers and a potential counterweight to US/China AI duopoly narratives. The €5.8B valuation set the benchmark for sovereign AI investment.



Mistral AI 完成 €600M（约 $645M）B 轮融资，估值 €5.8B。a16z、General Catalyst 等参投。成立仅 12 个月就成为欧洲最有价值的 AI 实验室。这不只是一轮融资——更是地缘政治信号：欧洲有了对抗美国超大规模厂商的可信替代方案。

Alibaba releases Qwen2 — a family of models (0.5B to 72B) that tops open-source leaderboards for multilingual performance, with strong coding and reasoning. Qwen2-72B outperforms Llama 3 70B and challenges frontier closed models on Chinese benchmarks.

So what? Qwen2 set the benchmark for what a Chinese open-weight model family could be: multilingual by design (29 languages), strong on global benchmarks, and fully open-weight. For any enterprise needing Chinese + English AI without vendor lock-in, Qwen2 became the default. Alibaba's cloud distribution made it the most widely deployed open-source model in China.



阿里巴巴开源 Qwen2 模型家族（0.5B 到 72B），在多语言基准上领先开源榜单。Qwen2-72B 超越 Llama 3 70B，在中文基准上叫板闭源前沿模型。29 种语言支持、全球基准强势，加上阿里云的分发能力——Qwen2 成为中国最广泛部署的开源模型。

Mistral launches Codestral (22B), its first code-specialized model. Designed for code completion and generation across 80+ programming languages, with fill-in-the-middle (FIM) support.

So what? Mistral signaled it's not just playing "open-source GPT" — it's carving out developer-specific vertical markets. Codestral integrates with VS Code, JetBrains, and Continue, making it competitive with GitHub Copilot for developers who prefer local or EU-hosted models. Europe's first serious contender in the coding AI race.



Mistral 上线 Codestral（22B），首款代码专用模型，支持 80+ 编程语言的代码补全和生成，带 fill-in-the-middle（FIM）功能。集成 VS Code、JetBrains 和 Continue，成为偏好本地或欧洲托管的开发者替代 GitHub Copilot 的选择。欧洲在编程 AI 赛道首次有了真正的竞争者。

xAI closed a $6 billion Series B funding round on May 26, 2024, reaching a post-money valuation of $24 billion with participation from Andreessen Horowitz, Sequoia Capital, Valor Equity Partners, Vy Capital, Fidelity, and Prince Alwaleed Bin Talal. The capital was earmarked for scaling xAI's infrastructure, including the massive Colossus supercomputer cluster in Memphis, Tennessee used to train Grok models. The round established xAI as one of the best-funded AI startups globally within just one year of its founding. A second $6B round (Series C) followed in December 2024, bringing total raised to over $12 billion.



xAI 完成 $6B B 轮融资，估值 $24B。a16z、Sequoia、Fidelity 等顶级机构参投。资金用于扩建 Memphis 的 Colossus 超级计算集群来训练 Grok 模型。成立仅一年就跻身全球资金最充裕的 AI 创业公司。12 月又完成 $6B C 轮，总融资超 $12B。

At the AI Seoul Summit, 16 leading AI companies including OpenAI, Google, Anthropic, Meta, and xAI signed voluntary Frontier AI Safety Commitments. The commitments included publishing safety frameworks, conducting red-teaming, and sharing safety information with governments.



AI 首尔峰会上，OpenAI、Google、Anthropic、Meta、xAI 等 16 家头部 AI 公司签署自愿性前沿 AI 安全承诺，包括公开安全框架、进行红队测试、与政府共享安全信息。

Microsoft introduces Copilot+ PCs with dedicated Neural Processing Units (NPUs). AI Recall feature, real-time translation, and image generation built into Windows.



Microsoft 推出 Copilot+ PC，配备专用 NPU（神经处理单元）。AI Recall 功能、实时翻译和图像生成直接集成进 Windows。

ByteDance extends Doubao family with lower-cost and multimodal variants, positioning it as consumer + enterprise model platform via Volcengine APIs.



字节跳动扩展豆包模型家族，推出更低成本和多模态版本，通过火山引擎 API 覆盖消费和企业场景。

Google announces Gemini integration across Search AI Overviews, Gmail, Docs, and Android at I/O 2024. AI becomes default interface layer across Google's product stack.



Google I/O 2024 上宣布 Gemini 全面整合进 Search AI Overviews、Gmail、Docs 和 Android。AI 正在成为 Google 产品栈的默认交互层。

Project Astra demonstrates low-latency, always-on multimodal interaction: camera + voice + memory + tool use. Prototype points to next-generation universal assistant.



Project Astra 演示低延迟、常驻式多模态交互：摄像头 + 语音 + 记忆 + 工具调用。原型指向下一代通用助手的方向。

OpenAI launches GPT-4o with native audio, vision, and text capabilities in a single model, with dramatically faster response times.

So what? The "Her" moment — real-time voice AI that feels human. This is when normies start to get it. The interface layer of AI just leaped forward.



OpenAI 上线 GPT-4o，单一模型原生支持音频、视觉和文本，响应速度飞快。《Her》时刻——实时语音 AI 让人感觉真的像人了。AI 的交互界面向前跨了一大步。

Chinese robotics company Unitree unveils G1 humanoid robot at $16,000 — a fraction of competitors. Makes humanoid robots accessible for research and development.



中国机器人公司宇树科技亮相 G1 人形机器人，售价 $16,000——只是竞品的零头。让人形机器人研发门槛大幅降低。

Google DeepMind and Isomorphic Labs unveiled AlphaFold 3 on May 8, 2024, a groundbreaking AI model capable of predicting the structure and interactions of all of life's molecules — proteins, DNA, RNA, ligands, and more. AlphaFold 3 dramatically improved upon its predecessor's accuracy, particularly for drug-target interactions. The model was made freely accessible to researchers via the AlphaFold Server. The release was hailed as a landmark in computational biology with profound implications for drug discovery and disease research.



Google DeepMind 和 Isomorphic Labs 亮相 AlphaFold 3，能预测所有生命分子的结构和相互作用——蛋白质、DNA、RNA、配体等。对药物靶点互作的预测精度大幅提升。通过 AlphaFold Server 免费开放给研究者。被誉为计算生物学的里程碑，对药物发现和疾病研究影响深远。

DeepSeek released DeepSeek-V2 on May 7, 2024, a powerful 236B parameter Mixture-of-Experts (MoE) model that activated only 21B parameters per token. The architecture delivered GPT-4-level performance at a fraction of the inference cost, with API pricing 75% cheaper than GPT-4 Turbo. DeepSeek-V2 achieved state-of-the-art performance among open-source models on reasoning and coding benchmarks. The release demonstrated that Chinese AI labs could build frontier-level models using innovative architectures, foreshadowing DeepSeek's later global impact.



DeepSeek 发布 V2，236B 参数 MoE 模型，每个 token 只激活 21B 参数。GPT-4 级性能，API 价格比 GPT-4 Turbo 便宜 75%。在开源模型推理和编程基准上达到领先水平。中国 AI 实验室用创新架构造出前沿模型，为 DeepSeek 后来的全球影响埋下伏笔。

Microsoft releases Phi-3-mini (3.8B), a small language model that achieves performance comparable to models 10× its size, capable of running fully on-device on a smartphone.

So what? AI isn't just about bigger and bigger — Phi-3 proved that quality data beats raw scale. For the first time, you could run a GPT-3.5 tier assistant locally without cloud. This kicked off the SLM race: on-device AI became viable, and every inference startup had to reckon with the edge. Microsoft's research team signaled: the future isn't only in the cloud.



Microsoft 发布 Phi-3-mini（3.8B），一个小语言模型，性能媲美大 10 倍的模型，手机就能跑。Phi-3 证明高质量数据比暴力堆参数更重要。首次实现本地运行 GPT-3.5 级别的助手，掀起了 SLM 竞赛——端侧 AI 正式可行。

Meta open-sources Llama 3 (8B and 70B), setting new benchmarks for open-weight models.

So what? Open source catches up fast. For builders, this means you can run near-GPT-4 quality locally. The commoditization of intelligence accelerates.



Meta 开源 Llama 3（8B 和 70B），刷新开源模型基准。对开发者来说，本地就能跑接近 GPT-4 品质的模型。智能的商品化在加速。

Meta launched Meta AI, its AI assistant powered by the Meta Llama 3 model, on April 18, 2024. Meta AI was embedded directly into Facebook, Instagram, WhatsApp, and Messenger, as well as Meta Ray-Ban smart glasses. Users could interact with Meta AI via text, voice, and the camera on smart glasses. At launch, Meta AI could search the web in real time and generate images through Instagram. The rollout made Meta AI one of the most widely distributed AI assistants globally.



Meta 上线 Meta AI 助手，跑 Llama 3 模型，直接嵌入 Facebook、Instagram、WhatsApp、Messenger 和 Meta Ray-Ban 智能眼镜。支持文本、语音和眼镜摄像头交互，还能实时搜索和在 Instagram 生成图片。Meta AI 由此成为全球分发最广的 AI 助手之一。

Boston Dynamics retires hydraulic Atlas and reveals fully electric version — stronger, more agile, designed for real-world commercial deployment with Hyundai.



Boston Dynamics 退役液压版 Atlas，亮相全电动版——更强、更敏捷，为与 Hyundai 合作的商业部署而生。

Amazon completes its $4 billion investment in Anthropic, the largest single AI investment to date. Anthropic will use AWS Trainium chips for training.



Amazon 完成对 Anthropic $4B 投资——截至当时 AI 领域最大的单笔投资。Anthropic 将使用 AWS Trainium 芯片进行训练。

NVIDIA announces Blackwell B200 GPU at GTC, claiming 30x inference speedup over Hopper for LLMs.

So what? Compute keeps scaling. NVIDIA's moat deepens. If you're investing in AI infrastructure, NVDA remains the picks-and-shovels play.



NVIDIA 在 GTC 上发布 Blackwell B200 GPU，声称 LLM 推理速度比 Hopper 快 30 倍。算力持续扩展，NVIDIA 的护城河越挖越深。

Hugging Face launches LeRobot, an open-source library for real-world robotics machine learning in PyTorch, led by former Tesla scientist Remi Cadene. Includes models, datasets, and simulation environments on the Hub.

So what? Hugging Face, which democratized NLP and then vision AI, turns its model-sharing flywheel toward physical intelligence. LeRobot is to robotics what Transformers.py was to language models: a single library that packages best-in-class algorithms for anyone to run. As physical AI heats up, whoever owns the open robotics stack controls the research default.



Hugging Face 推出 LeRobot，一个用 PyTorch 做真实世界机器人机器学习的开源库，由前 Tesla 科学家 Remi Cadene 主导。包含模型、数据集和仿真环境。LeRobot 之于机器人，就像 Transformers.py 之于语言模型：一个打包了最优算法让所有人都能跑的库。

NVIDIA launches GR00T foundation model initiative for humanoid robots, combining simulation, synthetic data, and embodied policy learning.



NVIDIA 推出 GR00T 人形机器人基础模型计划，融合仿真、合成数据和具身策略学习。

xAI open-sourced Grok-1 on March 17, 2024, releasing the weights and architecture of its 314 billion parameter Mixture-of-Experts language model under the Apache 2.0 license on GitHub and Hugging Face. Grok-1 was the raw base model checkpoint from the pre-training phase completed in October 2023, not fine-tuned for dialogue — making it one of the largest open-weight models ever released at the time. The release came shortly after the open-source community had been debating whether frontier labs would ever release such large models. It significantly raised the bar for what the open-source AI ecosystem could work with and build upon.



xAI 开源 Grok-1，将 314B 参数的 MoE 模型权重和架构以 Apache 2.0 协议放到 GitHub 和 Hugging Face。Grok-1 是 2023 年 10 月完成预训练的原始基座模型，未做对话微调——当时开源过的最大模型之一。这次开源大幅抬高了开源 AI 生态能触及的天花板。

Baidu announces ERNIE 4.0 Turbo upgrade with faster inference and lower cost. Continues push to deploy ERNIE across search and enterprise APIs.



百度宣布 ERNIE 4.0 Turbo 升级，推理更快、成本更低，继续推动 ERNIE 在搜索和企业 API 上的部署。

Anthropic released Claude 3 Haiku, the fastest and most compact model in the Claude 3 family, designed for near-instant responsiveness in enterprise applications. Despite being the smallest Claude 3 model, Haiku outperformed many competitor models on key benchmarks. At just $0.25 per million input tokens, it offered the best price-performance ratio in the Claude lineup. Haiku was positioned as the go-to model for high-throughput, latency-sensitive production deployments.



Anthropic 发布 Claude 3 Haiku，Claude 3 家族中最快最紧凑的模型，专为企业级高吞吐低延迟场景设计。尽管最小，Haiku 在关键基准上仍超越不少竞品。输入价格仅 $0.25/百万 token，性价比拉满。

Figure AI raises $675M at $2.6B valuation. Investors include Microsoft, OpenAI, Nvidia, Jeff Bezos. OpenAI provides multimodal AI models for Figure 01 humanoid.



Figure AI 完成 $675M 融资，估值 $2.6B。Microsoft、OpenAI、Nvidia、Jeff Bezos 均参投。OpenAI 为 Figure 01 人形机器人提供多模态 AI 模型。

Cognition unveils Devin, one of the first end-to-end software engineer agents publicly demonstrated with autonomous ticket completion workflows.



Cognition 亮相 Devin，首批公开演示的端到端软件工程师 AI 代理之一，能自主完成工单任务。

Anthropic releases Claude 3 family — Opus, Sonnet, and Haiku — with Opus matching or beating GPT-4 on key benchmarks.

So what? Three-horse race confirmed. Anthropic proves safety-focused labs can also build frontier models. The "responsible AI" narrative gets real teeth.



Anthropic 发布 Claude 3 家族——Opus、Sonnet 和 Haiku，其中 Opus 在关键基准上追平或超越 GPT-4。三方争霸格局确立，Anthropic 证明安全导向的实验室也能造出前沿模型。

Arweave launched AO Computer, a hyper-parallel computing environment built on top of its permanent storage layer. AO enables autonomous AI agents to run on decentralized infrastructure with verifiable computation, positioning Arweave beyond storage into AI compute.



Arweave 推出 AO Computer，一个构建在永久存储层之上的超并行计算环境。AO 让自主 AI 代理可以在去中心化基础设施上运行并进行可验证计算，Arweave 由此从存储扩展到 AI 计算领域。

Mistral AI launched Mistral Large on February 26, 2024, its most capable model competing directly with GPT-4 and Claude 3. Mistral Large featured top-tier reasoning, 32K context window, strong multilingual abilities in English, French, Spanish, German, and Italian, and native function calling. It was made available through the La Plateforme API and Microsoft Azure. Mistral Large ranked second on the MMLU benchmark among publicly available models, firmly establishing Mistral as a frontier AI lab.



Mistral AI 上线 Mistral Large，正面对标 GPT-4 和 Claude 3。32K 上下文窗口，多语言能力突出，支持原生函数调用。通过 La Plateforme API 和 Microsoft Azure 开放。在 MMLU 基准上排名公开模型第二，Mistral 正式跻身前沿 AI 实验室行列。

Google released Gemma on February 21, 2024 — a family of lightweight, open-weights AI models built from the same research and technology used to create Gemini. Available in 2B and 7B parameter sizes, Gemma models could run on laptops, workstations, or cloud infrastructure. Google provided pre-trained and instruction-tuned variants along with responsible use guidelines. Gemma became an immediate hit in the open-source AI community, offering Google-quality models that developers could freely use, fine-tune, and deploy.



Google 开源 Gemma，用造 Gemini 的同套技术打造的轻量级模型，2B 和 7B 两个版本，笔记本电脑就能跑。提供预训练和指令微调版本。Gemma 在开源社区迅速走红——Google 品质的模型，开发者可以自由使用、微调和部署。

Google releases Gemini 1.5 Pro with 1 million token context window — a 10x leap from GPT-4's 128K. MoE architecture enables processing entire codebases and hour-long videos.



Google 发布 Gemini 1.5 Pro，上下文窗口达 100 万 token——是 GPT-4 的 128K 的 10 倍。MoE 架构让它能处理整个代码库和一小时的视频。

OpenAI reveals Sora, capable of generating 60-second realistic videos from text. Demonstrates unprecedented temporal consistency. Not yet publicly available.



OpenAI 展示 Sora，能从文本生成 60 秒的逼真视频，时序一致性前所未有。暂未公开。

Google launches Gemini 1.0 Ultra, its most capable model, claiming to surpass GPT-4 on multiple benchmarks.

So what? Google finally shows up to the LLM race with real firepower. The duopoly narrative (OpenAI vs Google) solidifies — good for competition, good for users.



Google 上线 Gemini 1.0 Ultra，号称在多项基准上超越 GPT-4。Google 终于在 LLM 赛道拿出真火力，OpenAI vs Google 的双寡头格局开始成型。

Deepgram releases Nova-2, its most accurate real-time speech-to-text model, beating competitors on accuracy, speed, and cost across multiple languages and domains.

So what? Voice AI finally became a production-grade infrastructure primitive. Nova-2's combination of sub-200ms latency, 40+ language support, and lower price-per-minute than Whisper made it the default choice for developers building conversational AI products. Real-time transcription moved from "technically possible" to "obviously good enough for enterprise."



Deepgram 发布 Nova-2，实时语音转文字模型，在准确率、速度和成本上全面碾压竞品。低于 200ms 延迟、40+ 语言、比 Whisper 更便宜——语音 AI 从「技术上可行」变成了「企业级可用」。

Zhipu AI releases GLM-4, its flagship large language model competing with GPT-4 on Chinese-language tasks and general reasoning, with multimodal and code capabilities.

So what? China's top AI labs didn't just close the gap on English benchmarks — GLM-4 showed strength on Chinese-specific reasoning where Western models underperform. For enterprise buyers in China, a domestically trained frontier model isn't just a preference — it's a regulatory and data sovereignty necessity. Zhipu's GLM line is China's most viable GPT-4 substitute.



智谱 AI 发布 GLM-4，旗舰大语言模型，在中文任务和通用推理上对标 GPT-4，支持多模态和代码生成。中国头部 AI 实验室不只是在追赶英文基准——GLM-4 在中文推理上展现了西方模型难以企及的优势。对中国企业买家来说，国产前沿模型不仅是偏好，更是监管和数据主权的必需。

Tesla demonstrates Optimus Gen 2 with improved walking, hand dexterity, and the ability to fold shirts. Plans thousands of units for factory deployment in 2025.



Tesla 展示 Optimus Gen 2，行走、手部灵活性均有改进，还能叠衣服。计划 2025 年向工厂部署数千台。

OpenAI launched the GPT Store, a marketplace where users can discover and deploy custom GPTs built by creators worldwide. At launch, over 3 million custom GPTs had already been created by users since the feature was announced at DevDay in November 2023. The store featured categories including productivity, education, and programming. OpenAI also announced plans for a revenue-sharing program with top GPT creators, signaling a new creator economy around custom AI assistants.



OpenAI 上线 GPT Store，用户可以发现和使用全球创作者做的自定义 GPTs。上线时已有超过 300 万个自定义 GPTs 被创建。商店覆盖生产力、教育、编程等类别。OpenAI 还宣布将推出创作者收入分成计划，预示着围绕自定义 AI 助手的新创作者经济。

At CES 2024, NVIDIA showcased its Avatar Cloud Engine (ACE) technology enabling game developers to create AI-powered NPCs with dynamic, personalized conversations using natural language. NVIDIA also demonstrated digital human technologies through its Tokkio and Audio2Face platforms, showing photorealistic AI avatars capable of real-time conversation. The announcements highlighted NVIDIA's expansion beyond training chips into real-time AI inference for interactive applications. CEO Jensen Huang's keynote drew massive crowds as NVIDIA's AI dominance was at its peak.



CES 2024 上，NVIDIA 展示 Avatar Cloud Engine (ACE) 技术，让游戏开发者用自然语言给 NPC 注入 AI 驱动的动态对话能力。同时演示了 Tokkino 和 Audio2Face 平台的逼真 AI 数字人。这些动作凸显 NVIDIA 从训练芯片向实时 AI 推理交互领域扩张。Jensen Huang 的主题演讲人气爆棚，NVIDIA 的 AI 统治力正处巅峰。

The New York Times filed a landmark copyright lawsuit against OpenAI and Microsoft, alleging their AI models were trained on millions of NYT articles without permission. The case became the highest-profile AI copyright dispute and a bellwether for the industry.



《纽约时报》起诉 OpenAI 和 Microsoft，指控 AI 模型未经许可用数百万篇 NYT 文章训练。这成为 AI 版权领域最具标志性的诉讼案，也是行业风向标。

Mistral AI released Mixtral 8x7B on December 11, 2023, a sparse Mixture-of-Experts language model that uses only 2 of 8 expert networks per token, delivering GPT-3.5-level performance at a fraction of the inference cost. Released with open weights under the Apache 2.0 license, Mixtral 8x7B outperformed Llama 2 70B on most benchmarks despite using fewer active parameters. The model achieved strong multilingual performance in English, French, Italian, German, and Spanish, and excelled at code generation. Mixtral's efficient MoE architecture became one of the most influential open-source model designs of 2024, widely adopted and fine-tuned by the community.



Mistral AI 开源 Mixtral 8x7B，一款稀疏混合专家（MoE）模型，每个 token 只激活 8 个专家中的 2 个，用极低推理成本达到 GPT-3.5 水平。Apache 2.0 协议开放权重，在多数基准上超越 Llama 2 70B。多语言能力突出，代码生成优秀。Mixtral 的 MoE 架构成为 2024 年最有影响力的开源模型设计之一。

Google unveiled Gemini 1.0 on December 6, 2023 — its most capable and general AI model, designed to be natively multimodal across text, images, audio, video, and code. Gemini came in three sizes: Ultra, Pro, and Nano. Gemini Ultra outperformed GPT-4 on 30 of 32 academic benchmarks and was the first model to surpass human experts on the MMLU benchmark. The launch marked Google's most significant response to ChatGPT and its renewed commitment to AI leadership.



Google 亮相 Gemini 1.0——原生多模态 AI 模型，横跨文本、图片、音频、视频和代码。三个版本：Ultra、Pro 和 Nano。Gemini Ultra 在 32 项学术基准中的 30 项超越 GPT-4，首次在 MMLU 基准上超过人类专家。这是 Google 对 ChatGPT 最有力的回击。

Anthropic released Claude 2.1, doubling the context window to 200,000 tokens — the longest of any publicly available AI model at the time. The model also featured reduced hallucination rates and improved instruction following. Claude 2.1 could process entire codebases, financial reports, or legal documents in a single context. The 200K window set a new standard for enterprise long-document AI applications.



Anthropic 发布 Claude 2.1，上下文窗口翻倍至 200,000 token——当时公开可用 AI 模型中最长。幻觉率降低，指令执行更精准。Claude 2.1 可以一次处理整个代码库、财报或法律文件。200K 窗口为企业级长文档 AI 应用树立了新标准。

On November 17, 2023, OpenAI's board abruptly fired CEO Sam Altman, citing a loss of confidence in his candor. The move triggered immediate chaos: President Greg Brockman resigned, hundreds of employees threatened to follow Altman to Microsoft, and investors were blindsided. After four days of intense negotiations and near-collapse of the company, Altman was reinstated as CEO on November 21 with a reconstituted board. The saga exposed deep tensions between OpenAI's nonprofit governance structure and its commercial ambitions, accelerating the company's push toward a for-profit restructuring.



OpenAI 董事会突然解雇 CEO Sam Altman，引发巨震：总裁 Greg Brockman 辞职，数百员工威胁跟随 Altman 去 Microsoft，投资人措手不及。经过四天激烈博弈和公司几近崩溃，Altman 在 11 月 21 日复职，董事会重组。这场风波暴露了 OpenAI 非营利治理结构与商业野心之间的深层矛盾，加速了公司向营利性转型的步伐。

Ritual raised $25M in seed funding to build open, decentralized infrastructure for AI computation. The project aims to break the monopoly of large tech companies on AI access by enabling permissionless AI inference on-chain. Backed by Archetype and others, Ritual became an early pioneer of the AI×crypto infrastructure narrative.



Ritual 完成 $25M 种子轮融资，致力于构建去中心化的开放 AI 计算基础设施。项目目标是打破大科技公司对 AI 的垄断，实现链上无许可 AI 推理。Archetype 等机构参投，Ritual 成为 AI×Crypto 基础设施赛道的早期先驱。

At its first developer conference DevDay, OpenAI announced GPT-4 Turbo with a 128K context window and knowledge cutoff of April 2023, along with dramatically reduced pricing. The event also introduced custom GPTs — user-created versions of ChatGPT configured with specific instructions, knowledge, and tools. The GPT Store was announced as a forthcoming marketplace for sharing custom GPTs. DevDay marked OpenAI's transition from research lab to full-scale developer platform.



OpenAI 首届开发者大会 DevDay 上，发布 GPT-4 Turbo（128K 上下文窗口），价格大幅下降。同时推出自定义 GPTs——用户可以按需配置专属 ChatGPT，并预告 GPT Store 即将上线。DevDay 标志着 OpenAI 从研究实验室向全功能开发者平台转型。

The UK hosted the world's first global AI Safety Summit at the historic Bletchley Park — birthplace of modern computing. Representatives from 28 countries including the US, China, and EU signed the "Bletchley Declaration," acknowledging AI's existential risks. Major AI companies committed to sharing safety information with governments. A landmark moment for international AI governance.



英国在布莱切利庄园——现代计算的诞生地——举办首届全球 AI 安全峰会。28 个国家的代表（包括美国、中国和欧盟）签署《布莱切利宣言》，承认 AI 的存在性风险。各大 AI 公司承诺与政府共享安全信息。国际 AI 治理的里程碑。

President Biden signed the most comprehensive US executive order on AI, requiring safety testing for powerful AI systems, establishing AI safety standards through NIST, and addressing risks across national security, privacy, equity, and labor markets.



拜登总统签署美国最全面的 AI 行政令，要求对强大 AI 系统进行安全测试，通过 NIST 制定 AI 安全标准，覆盖国家安全、隐私、公平和劳动力市场等风险。

At Meta Connect 2023, Meta unveiled the Meta Quest 3 mixed-reality headset alongside a significant AI push. The event introduced AI Studio, enabling creators and businesses to build custom AI personas within Meta's social platforms. Meta also demonstrated AI-powered characters, real-time AI-generated backgrounds for video calls, and sticker generation in WhatsApp and Messenger. The announcements marked Meta's bet on the intersection of spatial computing and generative AI as the next computing platform.



Meta Connect 2023 上，Meta 发布 Quest 3 混合现实头显，同步推进 AI 战略。活动展示了 AI Studio、AI 角色、实时 AI 视频背景和 WhatsApp/Messenger 的贴纸生成功能。这些动作表明 Meta 押注空间计算与生成式 AI 的交汇点作为下一代计算平台。

Mistral AI released Mistral 7B on September 27, 2023, a 7-billion parameter open-source model that outperformed Meta's Llama 2 13B on every benchmark despite being half the size. Released via BitTorrent with an Apache 2.0 license with no usage restrictions, it became instantly available to the open-source community. Mistral 7B demonstrated that model efficiency could match brute-force scale, establishing Mistral as a leading open-source AI lab. The release was a watershed moment for small, highly capable language models.



Mistral AI 开源 Mistral 7B，一个 70 亿参数模型，在所有基准上都超过了大一倍的 Llama 2 13B。通过 BitTorrent 以 Apache 2.0 协议无限制分发。Mistral 7B 证明模型效率可以匹敌暴力堆参数，Mistral 由此跻身顶级开源 AI 实验室。小而强的语言模型时代开启。

Amazon announced a major strategic investment of up to $4 billion in Anthropic on September 25, 2023, with an initial $1.25 billion committed immediately. As part of the deal, Anthropic would use AWS as its primary cloud provider, and Amazon's customers would gain access to Claude models through Amazon Bedrock. Amazon also received a minority stake in Anthropic. The investment was Amazon's largest-ever venture bet and positioned AWS as a leading platform for enterprise AI — directly challenging Microsoft's OpenAI partnership.



Amazon 宣布对 Anthropic 战略投资最高 $4B，首期到账 $1.25B。作为交换，Anthropic 将 AWS 作为主要云服务商，AWS 客户可通过 Amazon Bedrock 使用 Claude 模型。这是 Amazon 有史以来最大的风投押注，直接叫板 Microsoft 和 OpenAI 的合作关系。

Hugging Face raised $235 million in a Series D funding round at a $4.5 billion valuation on August 23, 2023. The round was led by Salesforce Ventures and included investments from Google, Amazon, Nvidia, AMD, Intel, IBM, and Qualcomm — representing unprecedented backing from virtually all major tech companies. The investment validated Hugging Face's position as the central platform for open-source AI, hosting hundreds of thousands of models and datasets. The funding was earmarked for expanding the Hub, enterprise features, and compute infrastructure.



Hugging Face 完成 $235M D 轮融资，估值 $4.5B。Salesforce Ventures 领投，Google、Amazon、Nvidia、AMD、Intel、IBM、Qualcomm 齐参与——几乎所有科技巨头都投了。这验证了 Hugging Face 作为开源 AI 中心平台的地位，资金用于扩展 Hub、企业功能和算力基础设施。

China's Cyberspace Administration enacted the "Interim Measures for the Management of Generative Artificial Intelligence Services," becoming one of the world's first comprehensive generative AI regulatory frameworks. The rules require AI services to align with "socialist core values," register with authorities, and label AI-generated content. A milestone in global AI governance.



中国网信办施行《生成式人工智能服务管理暂行办法》，成为全球首批综合性生成式 AI 监管框架之一。要求 AI 服务符合社会主义核心价值观、向主管部门备案、标注 AI 生成内容。全球 AI 治理的里程碑事件。

Meta released Llama 2 on July 18, 2023, in partnership with Microsoft, making it freely available for research and commercial use. Llama 2 models (7B, 13B, 34B, 70B) were significantly better than Llama 1 and included instruction-tuned chat variants. The commercial license covered most businesses with under 700M monthly users, making it the most permissive frontier open model to date. Llama 2 became the dominant open-source base model for enterprise fine-tuning and local deployment, validating Meta's open strategy.



Meta 联合 Microsoft 开源 Llama 2（7B/13B/34B/70B），研究和商业免费使用。包含指令微调的聊天版本，商业许可覆盖月活 7 亿以下的企业。Llama 2 成为企业微调和本地部署的主流开源基座模型，验证了 Meta 的开放策略。

Google launched NotebookLM on July 12, 2023, initially to a limited group of US users. Originally previewed at Google I/O 2023 as "Project Tailwind," NotebookLM is an AI-powered research and note-taking tool that grounds its responses in user-provided documents, reducing hallucination. The product became broadly available in December 2023, gained AI-generated audio podcast summaries ("Audio Overviews") in 2024, and quickly became one of Google's most praised AI consumer products. It represented a new paradigm for AI-assisted reading and research.



Google 上线 NotebookLM，一款 AI 驱动的研究和笔记工具，回答基于用户提供的文档，有效减少幻觉。2024 年加入 AI 生成的播客音频摘要后迅速走红，成为 Google 最受好评的 AI 消费产品之一。它开创了 AI 辅助阅读和研究的新范式。

Anthropic launched Claude 2, a major upgrade featuring improved reasoning, longer context (100K tokens), and better coding capabilities. Claude 2 scored higher than its predecessor on academic benchmarks including bar exam, GRE, and coding tests. It was made available to the public via claude.ai and API, competing directly with GPT-4. The model also showed improvements in following complex multi-step instructions.



Anthropic 上线 Claude 2，推理能力升级、上下文拉到 100K token、编程能力增强。Claude 2 在律师考试、GRE 和编程测试中全面超越前代。通过 claude.ai 和 API 向公众开放，正面对标 GPT-4。多步骤复杂指令的执行也明显变好。

OpenAI CEO Sam Altman testified before the US Senate Judiciary Subcommittee on Privacy, Technology, and the Law, calling for government regulation of AI. He suggested the creation of a new federal agency to license AI companies. The hearing marked AI's arrival as a mainstream political issue and set the stage for global regulatory debates.



OpenAI CEO Sam Altman 在美国参议院听证会作证，呼吁政府监管 AI，建议设立新的联邦机构来给 AI 公司发牌照。这场听证标志着 AI 正式成为主流政治议题，为全球监管讨论拉开序幕。

Canadian robotics company Sanctuary AI unveiled Phoenix, its sixth-generation general-purpose humanoid robot, alongside Carbon — an AI control system designed to give Phoenix human-like intelligence. Standing 5'7" and weighing 155 lbs, Phoenix could perform over 100 industrial tasks. Sanctuary's vision: robots that can learn any job a human can do, marking an early milestone in the race for commercially viable humanoid robots.



加拿大机器人公司 Sanctuary AI 亮相第六代通用人形机器人 Phoenix，搭配 Carbon AI 控制系统。身高 5'7"、体重 155 磅的 Phoenix 能执行超过 100 种工业任务。Sanctuary 的愿景：造出能学会任何人类工作的机器人，在商用人形机器人竞赛中立下早期里程碑。

Anthropic announced that Claude now supports a 100,000-token context window — equivalent to roughly 75,000 words or an entire book. This was a breakthrough in long-context AI, enabling users to analyze large codebases, lengthy documents, and complex research papers in a single prompt. The 100K context window set a new industry benchmark and significantly expanded Claude's practical use cases.



Anthropic 宣布 Claude 支持 100,000 token 上下文窗口——约 75,000 词，差不多一本书。这是长上下文 AI 的突破，用户可以一次分析大型代码库、长篇文档和复杂研究论文。100K 上下文窗口刷新了行业标杆，大幅拓展了 Claude 的实用场景。

Google unveiled PaLM 2 at Google I/O on May 10, 2023, its next-generation large language model with significantly improved multilingual, reasoning, and coding capabilities. PaLM 2 powered over 25 Google products at launch, including Bard, Workspace AI features, and the new Med-PaLM 2 medical variant. It was available in four sizes (Gecko, Otter, Bison, Unicorn) allowing deployment across devices from mobile to enterprise scale. PaLM 2 represented Google's answer to GPT-4, demonstrating that training efficiency improvements could yield major capability gains without simply scaling parameters.



Google 在 I/O 大会亮相 PaLM 2，多语言、推理和编程能力大幅提升。PaLM 2 上线即驱动 25+ 款 Google 产品，包括 Bard 和医疗版 Med-PaLM 2。四个尺寸（Gecko、Otter、Bison、Unicorn）覆盖从手机到企业级部署。PaLM 2 是 Google 对 GPT-4 的回应——训练效率提升带来的能力飞跃，不只靠堆参数。

Geoffrey Hinton, the "Godfather of AI" and 2024 Nobel laureate, resigned from Google to speak freely about the dangers of AI. He warned that AI systems could become more intelligent than humans sooner than expected, marking a watershed moment in the AI safety debate.



「AI 教父」Geoffrey Hinton（2024 年诺贝尔奖得主）从 Google 辞职，为的是能自由谈论 AI 的危险。他警告 AI 系统可能比预期更早超越人类智能，这成为 AI 安全讨论的分水岭时刻。

Google announced the merger of DeepMind and Google Brain into a single unit called Google DeepMind on April 14, 2023. The move consolidated Google's two most prominent AI research organizations under a unified leadership structure, with Demis Hassabis serving as CEO. The merger was designed to accelerate AI progress by combining DeepMind's foundational research strengths with Google Brain's deep expertise in large-scale machine learning. Google DeepMind became one of the most powerful AI research organizations in the world.



Google 宣布将 DeepMind 和 Google Brain 合并为 Google DeepMind，由 Demis Hassabis 担任 CEO。此举把 Google 两大顶级 AI 研究机构整合到统一架构下，结合 DeepMind 的基础研究实力和 Google Brain 的大规模机器学习经验。Google DeepMind 由此成为全球最强大的 AI 研究组织之一。

Anthropic publicly launched Claude, its first AI assistant, on March 14, 2023. Claude was designed with safety as a core principle, trained using Constitutional AI (CAI) to be helpful, harmless, and honest. The launch marked Anthropic's entrance into the consumer AI market, positioning itself as a safety-focused alternative to ChatGPT. Early access was available via claude.ai and API.



Anthropic 公开上线 Claude，首款 AI 助手。Claude 把安全放在核心位置，用 Constitutional AI (CAI) 训练，追求有用、无害、诚实。这标志着 Anthropic 正式进入消费级 AI 市场，定位为 ChatGPT 的安全替代方案。

OpenAI released GPT-4, its most capable large language model at the time, featuring multimodal inputs (text and images) and dramatically improved reasoning. GPT-4 scored in the top 10% of bar exam takers and significantly outperformed GPT-3.5 on standard benchmarks. The model powered ChatGPT Plus and was made available via API. GPT-4 set a new standard for frontier AI and accelerated enterprise adoption of large language models across industries.



OpenAI 发布 GPT-4，当时最强大语言模型，支持多模态输入（文本+图片），推理能力大幅跃升。GPT-4 在律师考试中排进前 10%，各项基准远超 GPT-3.5。模型驱动 ChatGPT Plus 并开放 API。GPT-4 为前沿 AI 定下新标杆，加速了各行业采用大语言模型。

Meta released LLaMA (Large Language Model Meta AI) on February 24, 2023, a collection of foundation language models ranging from 7B to 65B parameters. Unlike proprietary models, LLaMA was released for research purposes to democratize access to large language models. Despite a research-only license, LLaMA weights leaked online within days, sparking an explosion of open-source AI development. LLaMA kickstarted the open-source LLM movement, enabling Alpaca, Vicuna, and dozens of community fine-tunes that accelerated AI research globally.



Meta 开源 LLaMA，一系列从 7B 到 65B 参数的基础语言模型。跟闭源模型不同，LLaMA 面向研究社区开放。尽管只授权研究用途，权重几天内就泄露到网上，点燃了开源 AI 大爆发。LLaMA 催生了 Alpaca、Vicuna 和数十个社区微调版本，推动全球 AI 研究提速。

Microsoft launched the new AI-powered Bing on February 7, 2023, integrating OpenAI's GPT-4 model into its search engine. The new Bing featured a conversational chat interface alongside traditional search results, allowing users to ask complex questions and get synthesized answers. Microsoft positioned Bing Chat as a "co-pilot for the web," marking the most significant challenge to Google's search dominance in decades. The launch ignited an AI search war and was later renamed Microsoft Copilot.



Microsoft 上线 AI 版 Bing，把 OpenAI 的 GPT-4 集成进搜索引擎。新 Bing 在传统搜索结果旁加了对话式聊天界面，用户可以提复杂问题并获得综合回答。Microsoft 把它定位为「网络副驾驶」，这是几十年来 Google 搜索霸主地位面临的最大挑战。后来改名为 Microsoft Copilot。

Google announced Bard, its experimental conversational AI service powered by LaMDA (Language Model for Dialogue Applications), on February 6, 2023 — just days after Microsoft launched the new AI-powered Bing. Bard was initially made available to trusted testers before opening to the public in March. The announcement signaled Google's urgency to compete in the AI chatbot space, though an early demo error briefly knocked $100B off Alphabet's market cap, highlighting the risks of rushed AI deployment.



Google 在 Microsoft 推出 AI 版 Bing 几天后，紧急宣布实验性对话 AI 服务 Bard，底层跑的是 LaMDA。Bard 先给内测用户试用，三月向公众开放。但一次早期演示出了错，Alphabet 市值一天蒸发 $100B——AI 仓促上阵的代价。

Microsoft announced a multiyear, multi-billion-dollar investment in OpenAI on January 23, 2023 — reported to total $10 billion over multiple years. The deal extended and deepened the partnership that began in 2019, giving Microsoft exclusive rights to commercialize OpenAI's technologies through Azure. Microsoft planned to integrate OpenAI models across its product suite including Office 365, Bing, and GitHub Copilot. The investment transformed Microsoft's competitive position versus Google and was the defining AI partnership deal of the era.



Microsoft 宣布对 OpenAI 追加多年期数十亿美元投资，据报总额达 $10B。这笔交易延续了 2019 年开始的合作关系，Microsoft 拿到 OpenAI 技术在 Azure 上的独家商业化权。Microsoft 计划把 OpenAI 模型整合进 Office 365、Bing 和 GitHub Copilot。这笔投资彻底改变了 Microsoft 与 Google 的竞争格局，堪称这个时代最具定义性的 AI 合作案。

OpenAI launched ChatGPT on November 30, 2022, as a free research preview built on GPT-3.5. Within five days it attracted one million users, and within two months reached 100 million — the fastest consumer app growth in history. ChatGPT demonstrated that conversational AI could be genuinely useful for writing, coding, analysis, and Q&A at scale. The release triggered a wave of AI investment and competition that reshaped the entire technology industry.



OpenAI 在 2022 年 11 月 30 日上线 ChatGPT，基于 GPT-3.5 的免费研究预览。五天内吸引一百万用户，两个月破亿——史上增长最快的消费级应用。ChatGPT 证明对话式 AI 在写作、编程、分析和问答上真的好用。这次发布引爆了一轮 AI 投资和竞争浪潮，重塑了整个科技行业。