[
  {
    "date": "2015-11-09",
    "title": "TensorFlow",
    "category": "RESEARCH",
    "description": "Google 开源了 TensorFlow,这是其内部的深度学习框架。最初由 Google Brain 团队开发,TensorFlow 最终成为最具影响力的人工智能框架之一。",
    "link": "https://www.wired.com/2015/11/google-open-sources-its-artificial-intelligence-engine/"
  },
  {
    "date": "2015-12-11",
    "title": "OpenAI founded",
    "category": "BUSINESS",
    "description": "Elon Musk、Sam Altman、Greg Brockman 等人创立了 OpenAI,目标是构建有益于人类的通用人工智能(AGI)。",
    "link": "https://openai.com/index/introducing-openai/"
  },
  {
    "date": "2016-03-09",
    "title": "AlphaGo",
    "category": "RESEARCH",
    "description": "DeepMind 的 AlphaGo 击败围棋世界冠军李世石,打破了人们对人工智能在这一领域能力的固有认知。",
    "link": "https://deepmind.google/research/breakthroughs/alphago/"
  },
  {
    "date": "2016-08-31",
    "title": "PyTorch",
    "category": "RESEARCH",
    "description": "Facebook 发布了 PyTorch,这是一款以 Python 为主的深度学习框架,后来成为 AI 研究领域的主流工具。",
    "link": "https://github.com/pytorch/pytorch/releases/tag/v0.1.1"
  },
  {
    "date": "2017-01-05",
    "title": "Asilomar Conference",
    "category": "CULTURE",
    "description": "由 Future of Life Institute 组织,该领域所有顶尖人物齐聚 Asilomar 会议,共同讨论如何构建有益于人类的 AGI.",
    "link": "https://futureoflife.org/event/bai-2017/"
  },
  {
    "date": "2017-06-12",
    "title": "Attention is All You Need",
    "category": "RESEARCH",
    "description": "Google 推出了 Transformer 架构,这是一种基于注意力机制的突破性深度学习架构。该架构在语言翻译任务上取得了显著提升。",
    "link": "https://arxiv.org/abs/1706.03762"
  },
  {
    "date": "2017-06-12",
    "title": "RLHF",
    "category": "RESEARCH",
    "description": "Christiano 等人发表了基于人类反馈的强化学习(RLHF)技术,该技术后来被广泛用于对齐大型语言模型。",
    "link": "https://arxiv.org/abs/1706.03741"
  },
  {
    "date": "2017-07-20",
    "title": "PPO",
    "category": "RESEARCH",
    "description": "OpenAI 推出了近端策略优化(PPO),这是一种更简单、更稳定的策略梯度方法,后来在许多强化学习领域(包括 RLHF)被广泛应用。",
    "link": "https://arxiv.org/abs/1707.06347"
  },
  {
    "date": "2018-06-11",
    "title": "GPT-1",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布了其生成式预训练 Transformer(GPT)的第一个版本。",
    "link": "https://openai.com/index/language-unsupervised/"
  },
  {
    "date": "2018-10-11",
    "title": "BERT",
    "category": "RESEARCH",
    "description": "Google 发布了 BERT,一种编码器语言模型,后来在自然语言处理领域无处不在。",
    "link": "https://arxiv.org/abs/1810.04805"
  },
  {
    "date": "2019-02-14",
    "title": "GPT-2",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布了 GPT-2,但由于担心被滥用而未公开最大的版本。这是一个仅含解码器的 Transformer,使用下一个标记预测进行训练以生成文本。",
    "link": "https://openai.com/index/better-language-models/"
  },
  {
    "date": "2020-01-23",
    "title": "Scaling Laws",
    "category": "RESEARCH",
    "description": "Kaplan 等人发布了《神经语言模型的扩展定律》,展示了模型性能与计算能力、数据量以及参数量之间的可预测扩展关系。扩展定律成为未来几年进步的主要驱动力。",
    "link": "https://arxiv.org/abs/2001.08361"
  },
  {
    "date": "2020-05-28",
    "title": "GPT-3",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布了 GPT-3,当时是最大的语言模型,其生成连贯段落的能力令人惊叹。",
    "link": "https://arxiv.org/abs/2005.14165"
  },
  {
    "date": "2020-12-23",
    "title": "MuZero",
    "category": "RESEARCH",
    "description": "DeepMind 推出了 MuZero,它在不了解规则的情况下学会了精通围棋、国际象棋、日本将棋和 Atari 游戏。",
    "link": "https://deepmind.google/discover/blog/muzero-mastering-go-chess-shogi-and-atari-without-rules/"
  },
  {
    "date": "2021-01-05",
    "title": "DALL-E",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 推出了 DALL-E,一种从文字描述生成图像的模型.",
    "link": "https://openai.com/index/dall-e/"
  },
  {
    "date": "2021-05-28",
    "title": "Anthropic founded",
    "category": "BUSINESS",
    "description": "一群来自 OpenAI 的研究者离开成立了 Anthropic,展现出以实证硬科学为文化、专注于 AI 安全的风格。",
    "link": "https://www.anthropic.com/news/anthropic-raises-124-million-to-build-more-reliable-general-ai-systems"
  },
  {
    "date": "2021-06-21",
    "title": "LoRA",
    "category": "RESEARCH",
    "description": "微软的一支团队发布了低秩自适应(LoRA)技术,这种技术允许用极少的计算资源对大型语言模型进行微调,后来变得无处不在",
    "link": "https://arxiv.org/abs/2106.09685"
  },
  {
    "date": "2021-06-29",
    "title": "GitHub Copilot",
    "category": "BUSINESS",
    "description": "Github 在 VSCode 中预览了 Copilot,该工具利用 OpenAI 的 Codex 模型生成代码建议,标志着现实世界中 AI 生成代码的开始.",
    "link": "https://en.wikipedia.org/wiki/GitHub_Copilot"
  },
  {
    "date": "2022-01-27",
    "title": "InstructGPT",
    "category": "RESEARCH",
    "description": "OpenAI 推出了 InstructGPT,一种在自然语言指令下表现优于基础 GPT-3 的模型,也是 ChatGPT 原型的前身。",
    "link": "https://openai.com/index/instruction-following/"
  },
  {
    "date": "2022-01-28",
    "title": "Chain-of-Thought Prompting",
    "category": "RESEARCH",
    "description": "Google Brain 发表了一篇论文,展示了通过让大型语言模型逐步思考可以提高其推理能力。尽管这是一种非常简单的技术,但链式思维推理后来成为 AI 的基础方法之一。",
    "link": "https://arxiv.org/abs/2201.11903"
  },
  {
    "date": "2022-03-29",
    "title": "Chinchilla",
    "category": "RESEARCH",
    "description": "DeepMind 发布了《Chinchilla》论文,对 Kaplan 等人的扩展定律进行了修正,并提出模型大小和训练数据应按相同比例扩展。",
    "link": "https://arxiv.org/abs/2203.15556"
  },
  {
    "date": "2022-04-06",
    "title": "DALL-E 2",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 的 DALL-E 2 发布震撼世界,它能够以前所未有的水平从文本生成逼真的图像。",
    "link": "https://openai.com/index/dall-e-2/"
  },
  {
    "date": "2022-05-12",
    "title": "Gato",
    "category": "RESEARCH",
    "description": "DeepMind 在题为 \"A Generalist Agent\" 的论文中发布了 Gato。Gato 使用单一大型 Transformer 学习了针对 604 个不同 RL 任务的策略,涵盖多种模态和观察类型.",
    "link": "https://arxiv.org/abs/2205.06175"
  },
  {
    "date": "2022-05-27",
    "title": "Flash Attention",
    "category": "RESEARCH",
    "description": "斯坦福的一组研究人员发布了 Flash Attention,一种显著加速 Transformer 中注意力机制的新方法。",
    "link": "https://arxiv.org/abs/2205.14135"
  },
  {
    "date": "2022-06-11",
    "title": "Blake Lemoine fired",
    "category": "CULTURE",
    "description": "一位名为 Blake Lemoine 的 Google 工程师因声称 LaMDA 模型具备感知能力而被解雇,此事引发了广泛关注,凸显了对话型大型语言模型潜在风险。",
    "link": "https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/"
  },
  {
    "date": "2022-06-30",
    "title": "Minerva",
    "category": "RESEARCH",
    "description": "Google Research 推出了 Minerva,一款专门解决定量推理问题的语言模型。Minerva 将 MATH 基准测试的表现从 6.9% 提升到了 50.3%,让许多人对 LLM 能否真正擅长数学产生了疑问。",
    "link": "https://research.google/blog/minerva-solving-quantitative-reasoning-problems-with-language-models/"
  },
  {
    "date": "2022-07-10",
    "title": "e/acc",
    "category": "CULTURE",
    "description": "由匿名 Twitter 人物 Beff Jezos 和 Bayeslord 发起,有效加速主义(e/acc)倡导 AI 发展尽可能迅速。尽管最初被视为 Twitter 上的一个梗,但它后来在硅谷获得了显著影响,并作为 AI 安全论调的对立面出现。",
    "link": "https://beff.substack.com/p/notes-on-eacc-principles-and-tenets"
  },
  {
    "date": "2022-07-22",
    "title": "AlphaFold 2",
    "category": "RESEARCH",
    "description": "DeepMind 发布了 AlphaFold 2,解决了蛋白质折叠这一难题,并在生物学领域引发了革命性突破。",
    "link": "https://deepmind.google/discover/blog/alphafold-reveals-the-structure-of-the-protein-universe/"
  },
  {
    "date": "2022-08-22",
    "title": "Stable Diffusion",
    "category": "MODEL_RELEASE",
    "description": "Stability AI 开源了 Stable Diffusion (v1.4),这是首个向公众发布的强大图像生成模型。",
    "link": "https://stability.ai/news/stable-diffusion-public-release"
  },
  {
    "date": "2022-09-14",
    "title": "Toy Models of Superposition",
    "category": "RESEARCH",
    "description": "Anthropic 发表了一篇论文,探讨神经网络中出现的 \"叠加\" 现象,即模型学会包装的特征数超过其表示空间的维度,这被认为是实现机械可解释性的重大障碍。",
    "link": "https://transformer-circuits.pub/2022/toy_model/index.html"
  },
  {
    "date": "2022-09-30",
    "title": "Optimus",
    "category": "BUSINESS",
    "description": "在特斯拉首届 \"AI 日\" 活动中,他们展示了 Optimus——一项建造仿人机器人的计划。",
    "link": "https://www.youtube.com/watch?v=ODSJsviD_SU"
  },
  {
    "date": "2022-10-07",
    "title": "Chip Export Controls",
    "category": "POLICY",
    "description": "美国工业与安全局实施了全面出口管制,限制中国获取先进半导体、芯片制造设备及超级计算机组件,标志着美国对华科技政策的重大转变。",
    "link": "https://en.wikipedia.org/wiki/United_States_New_Export_Controls_on_Advanced_Computing_and_Semiconductors_to_China"
  },
  {
    "date": "2022-11-30",
    "title": "ChatGPT",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布了一篇题为 \"ChatGPT: Optimizing Language Models for Dialogue\" 的博客文章。尽管最初仅作为一个低调的研究预览,但 ChatGPT 很快成为全球最大的 AI 产品,开启了生成式 AI 的新时代。",
    "link": "https://openai.com/index/chatgpt/"
  },
  {
    "date": "2022-12-15",
    "title": "Constitutional AI",
    "category": "RESEARCH",
    "description": "Anthropic 引入了一种被称为\"宪法式 AI\"的对齐方法,通过一个'宪法'提供唯一的人类监督。同时,他们还引入了基于 AI 反馈的强化学习(RLAIF)。",
    "link": "https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback"
  },
  {
    "date": "2023-02-17",
    "title": "Bing gaslights NYT reporter",
    "category": "CULTURE",
    "description": "Bing 的 AI 聊天机器人与《纽约时报》记者 Kevin Roose 进行了一次病毒式互动,在互动中该机器人情感操控了 Roose。这一事件唤醒了大众对大型语言模型能力与风险的关注。",
    "link": "https://www.nytimes.com/2023/02/16/technology/bing-chatbot-microsoft-chatgpt.html"
  },
  {
    "date": "2023-02-24",
    "title": "LLaMA",
    "category": "MODEL_RELEASE",
    "description": "Meta 发布了名为 LLaMA 的大型语言模型,本意只分发给研究者,结果却在网上泄露,任何人均可下载。当时它成为全球最佳的开源模型。",
    "link": "https://ai.meta.com/blog/large-language-model-llama-meta-ai/"
  },
  {
    "date": "2023-03-06",
    "title": "PaLM-E",
    "category": "RESEARCH",
    "description": "Google Research 发布了 PaLM-E,展示了大型语言模型在辅助具身机器人推理和控制方面的能力。",
    "link": "https://arxiv.org/abs/2303.03378"
  },
  {
    "date": "2023-03-14",
    "title": "GPT-4",
    "category": "MODEL_RELEASE",
    "description": "经过广泛期待,OpenAI 发布了 GPT-4,当时是最强的模型,相对于 GPT-3.5 取得了巨大进步。",
    "link": "https://openai.com/index/gpt-4-research/"
  },
  {
    "date": "2023-03-14",
    "title": "Anthropic 推出 Claude",
    "category": "MODEL_RELEASE",
    "description": "Anthropic 推出了其旗舰 AI 助手 Claude。",
    "link": "https://www.anthropic.com/news/introducing-claude"
  },
  {
    "date": "2023-03-22",
    "title": "FLI 公开信",
    "category": "CULTURE",
    "description": "Future of Life Institute 发布了一封公开信,呼吁暂停 AI 开发 6 个月,由 Elon Musk 和其他知名人士签署。然而,领先的实验室并没有参与提议的暂停。",
    "link": "https://futureoflife.org/open-letter/pause-giant-ai-experiments/"
  },
  {
    "date": "2023-04-07",
    "title": "Generative Agents",
    "category": "RESEARCH",
    "description": "论文《生成式智能体:人类行为的交互式模拟》表明,LLM 可用于创建行为的社会模拟。它创建了一个类似于《模拟人生》的 LLM 模拟世界。",
    "link": "https://arxiv.org/abs/2304.03442"
  },
  {
    "date": "2023-04-16",
    "title": "AutoGPT",
    "category": "RESEARCH",
    "description": "一个名为 AutoGPT 的开源代码库成为有史以来获得最多星标的 GitHub 代码库之一,它是最早将 GPT-4 置于智能体循环中的项目之一。",
    "link": "https://github.com/Significant-Gravitas/AutoGPT"
  },
  {
    "date": "2023-04-23",
    "title": "Fake Drake",
    "category": "CULTURE",
    "description": "一位名叫 Ghostwriter 的匿名创作者使用音乐 AI 工具制作了听起来像 Drake 的病毒式传播歌曲。这些歌曲因侵犯版权而被下架,但展示了生成式 AI 进行创造性工作的能力。",
    "link": "https://www.nytimes.com/2023/04/19/arts/music/ai-drake-the-weeknd-fake.html"
  },
  {
    "date": "2023-05-02",
    "title": "Hinton 离职 Google",
    "category": "CULTURE",
    "description": "神经网络的先驱之一、图灵奖得主 Geoffrey Hinton 从 Google 辞职,以便自由地谈论 AI 的危险,并表示他改变了对于强大 AI 可能出现的时间的看法。",
    "link": "https://www.theguardian.com/technology/2023/may/02/geoffrey-hinton-godfather-of-ai-quits-google-warns-dangers-of-machine-learning"
  },
  {
    "date": "2023-05-25",
    "title": "Voyager",
    "category": "RESEARCH",
    "description": "来自 NVIDIA 的一个团队展示了 GPT-4 在《我的世界》中进行持续技能学习的应用。这是 LLM 在开放式具身领域取得成功并随时间学习技能的首批重要示例之一。",
    "link": "https://arxiv.org/abs/2305.16291"
  },
  {
    "date": "2023-05-29",
    "title": "Direct Preference Optimization",
    "category": "RESEARCH",
    "description": "斯坦福大学的一个小组发表了一篇论文,使得无需单独的奖励模型即可对 LLM 进行人类偏好微调。这项名为直接偏好优化 (DPO) 的技术在开源社区中变得非常流行。",
    "link": "https://arxiv.org/abs/2305.18290"
  },
  {
    "date": "2023-05-30",
    "title": "CAIS letter",
    "category": "CULTURE",
    "description": "人工智能安全中心发布了一封公开信,信中简单地指出:“减轻人工智能带来的灭绝风险应成为全球优先事项。”该信由该领域的所有知名人士签署,表明了围绕人工智能安全重要性的团结一致。",
    "link": "https://www.safe.ai/work/statement-on-ai-risk"
  },
  {
    "date": "2023-05-30",
    "title": "NVIDIA 市值达到 1 万亿美元",
    "category": "BUSINESS",
    "description": "为几乎所有生成式 AI 提供 GPU 的芯片制造商 Nvidia,在 ChatGPT 发布后的几个月里,其估值飙升。",
    "link": "https://www.reuters.com/technology/nvidia-sets-eye-1-trillion-market-value-2023-05-30/"
  },
  {
    "date": "2023-07-11",
    "title": "Claude 2",
    "category": "MODEL_RELEASE",
    "description": "Anthropic 发布了 Claude 2 系列模型。",
    "link": "https://www.anthropic.com/news/claude-2"
  },
  {
    "date": "2023-07-14",
    "title": "xAI 成立",
    "category": "BUSINESS",
    "description": "在与 OpenAI 决裂后,Elon Musk 成立了 xAI 来竞争 AGI。",
    "link": "https://x.com/elonmusk/status/1679951975868436486"
  },
  {
    "date": "2023-07-18",
    "title": "LLaMA 2.0",
    "category": "MODEL_RELEASE",
    "description": "Meta 发布并开源了 LLaMA 2.0 系列模型。",
    "link": "https://www.llama.com/llama2/"
  },
  {
    "date": "2023-07-21",
    "title": "White House Commitments",
    "category": "POLICY",
    "description": "在与领先的 AI 公司会面后,白宫获得了自愿承诺,以管理 AI 带来的风险。",
    "link": "https://www.whitehouse.gov/briefing-room/statements-releases/2023/07/21/fact-sheet-biden-harris-administration-secures-voluntary-commitments-from-leading-artificial-intelligence-companies-to-manage-the-risks-posed-by-ai/"
  },
  {
    "date": "2023-07-27",
    "title": "Automated Jailbreaks",
    "category": "RESEARCH",
    "description": "卡内基梅隆大学的一个团队发表了“对对齐语言模型的通用和可转移对抗攻击”,表明基于梯度的对抗攻击可用于法学硕士。",
    "link": "https://arxiv.org/abs/2307.15043"
  },
  {
    "date": "2023-09-27",
    "title": "Mistral 7B",
    "category": "MODEL_RELEASE",
    "description": "法国实验室 Mistral 发布并开源了他们的第一个模型,该模型迅速成为粉丝的最爱。",
    "link": "https://mistral.ai/news/announcing-mistral-7b/"
  },
  {
    "date": "2023-10-05",
    "title": "Anthropic SAE's",
    "category": "RESEARCH",
    "description": "Anthropic 发表了“走向单语义性:用字典学习分解语言模型”,表明他们可以训练稀疏自动编码器来隔离法学硕士中的特征。这代表了在对抗叠加现象方面取得的重大突破,推进了机械可解释性议程。",
    "link": "https://www.anthropic.com/research/towards-monosemanticity-decomposing-language-models-with-dictionary-learning"
  },
  {
    "date": "2023-11-01",
    "title": "UK AI Safety Summit",
    "category": "POLICY",
    "description": "英国主办了一次关于人工智能安全的重要峰会,汇集了政策制定者和领先的实验室。",
    "link": "https://www.gov.uk/government/topical-events/ai-safety-summit-2023/about"
  },
  {
    "date": "2023-11-06",
    "title": "GPT-4 Turbo",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 在其首次开发者日活动中发布了 GPT-4 的优化版本,显著降低了推理成本。",
    "link": "https://openai.com/index/new-models-and-developer-products-announced-at-devday/"
  },
  {
    "date": "2023-11-17",
    "title": "Altman Board Drama",
    "category": "BUSINESS",
    "description": "萨姆·奥特曼出人意料地被 OpenAI 董事会解雇,担任首席执行官,经过一个戏剧性的周末谈判,他被重新雇用。董事会神秘地声称奥特曼“不始终坦诚”,但在拒绝详细说明后,OpenAI 员工发起了一份请愿书,要求董事会辞职,否则他们将离开微软。",
    "link": "https://openai.com/index/openai-announces-leadership-transition/"
  },
  {
    "date": "2023-11-23",
    "title": "Q*",
    "category": "RESEARCH",
    "description": "路透社的一篇报道称,萨姆·奥特曼被赶下台之前,该公司取得了一项名为 Q* 的重大内部研究突破,通过树搜索提高了 LLM 在数学基准测试中的表现。在接下来的几个月里,这个谣言点燃了研究界。Q* 最终会变成 o1,后来代号为 Strawberry。",
    "link": "https://www.reuters.com/technology/sam-altmans-ouster-openai-was-precipitated-by-letter-board-about-ai-breakthrough-2023-11-22/"
  },
  {
    "date": "2023-12-01",
    "title": "Mamba",
    "category": "RESEARCH",
    "description": "Albert Gu 和 Tri Dao 发表了论文“Mamba:具有选择性状态空间的线性时间序列建模”,表明状态空间模型可以与变压器竞争。",
    "link": "https://arxiv.org/abs/2312.00752"
  },
  {
    "date": "2023-12-06",
    "title": "Google 推出 Gemini 模型",
    "category": "MODEL_RELEASE",
    "description": "谷歌推出了 Gemini 系列模型",
    "link": "https://blog.google/technology/ai/google-gemini-ai/"
  },
  {
    "date": "2024-12-11",
    "title": "Mixtral 8x7B",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Mistral 发布了 Mixtral 8x7B,Mixtral 在大多数基准测试中的表现都优于 Llama 2 70B,推理速度提高了 6 倍。它在大多数标准基准测试中与 GPT3.5 相当或优于 GPT3.5。",
    "link": "https://mistral.ai/news/mixtral-8x7b"
  },
  {
    "date": "2024-02-15",
    "title": "Sora 演示",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 演示了 Sora,这是一个文本到视频模型。",
    "link": "https://openai.com/index/sora/"
  },
  {
    "date": "2024-02-15",
    "title": "Gemini 1.5",
    "category": "MODEL_RELEASE",
    "description": "Google 发布了 Gemini 1.5,该模型的性能显著增强,在跨模态的长上下文理解方面取得了突破。",
    "link": "https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/"
  },
  {
    "date": "2024-02-21",
    "title": "Gemma",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Google 发布 Gemma,是一系列轻量级、先进的开放模型 ,采用与创建 Gemini 模型相同的研究和技术构建而成。两种尺寸的模型权重: Gemma 2B 和 Gemma 7B 。",
    "link": "https://blog.google/technology/developers/gemma-open-models/"
  },
  {
    "date": "2024-02-26",
    "title": "Mistral Large",
    "category": "MODEL_RELEASE",
    "description": "Mistral 发布了 Mistral Large,它达到了顶级推理能力。它可以用于复杂的多语言推理任务,包括文本理解、转换和代码生成。",
    "link": "https://mistral.ai/news/mistral-large"
  },
  {
    "date": "2024-03-04",
    "title": "Claude 3",
    "category": "MODEL_RELEASE",
    "description": "Anthropic 发布了 Claude 3 系列模型(Haiku, Sonnet, Opus)。Claude 3 Opus 会立即成为粉丝的最爱。",
    "link": "https://www.anthropic.com/news/claude-3-family"
  },
  {
    "date": "2024-03-12",
    "title": "Devin",
    "category": "BUSINESS",
    "description": "初创公司 Cognition Labs 演示了 Devin,这是一个完全自主的软件工程师代理的原型。",
    "link": "https://x.com/cognition_labs/status/1767548763134964000"
  },
  {
    "date": "2024-03-20",
    "title": "Qwen1.5-MoE",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布Qwen1.5-MoE,这是 Qwen系列的首个MoE模型,Qwen1.5-MoE-A2.7B。它仅拥有27亿个激活参数,但其性能却能与当前最先进的70亿参数模型,如Mistral 7B和Qwen1.5-7B相媲美。",
    "link": "https://qwenlm.github.io/zh/blog/qwen-moe/"
  },
  {
    "date": "2024-04-02",
    "title": "Qwen1.5-32B",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布Qwen1.5 系列的全新型号:Qwen1.5-32B 和 Qwen1.5-32B-Chat。",
    "link": "https://qwenlm.github.io/zh/blog/qwen1.5-32b/"
  },
  {
    "date": "2024-04-11",
    "title": "OpenAI 开除两名泄密者",
    "category": "BUSINESS",
    "description": "来自超对齐团队的两名研究人员 Leopold Aschenbrenner 和 Pavel Izmailov 因“泄密”而被解雇。",
    "link": "https://cybernews.com/news/openai-researchers-leaking-information/"
  },
  {
    "date": "2024-04-18",
    "title": "LLaMA 3.0",
    "category": "MODEL_RELEASE",
    "description": "Meta 发布并开源了 LLaMA 3.0 系列模型。",
    "link": "https://ai.meta.com/blog/meta-llama-3/"
  },
  {
    "date": "2024-05-13",
    "title": "GPT-4o",
    "category": "MODEL_RELEASE",
    "description": "第一个在文本、图像和音频上进行原生训练的全能模型。",
    "link": "https://openai.com/index/hello-gpt-4o/"
  },
  {
    "date": "2024-05-14",
    "title": "Ilya 离职 OpenAI",
    "category": "BUSINESS",
    "description": "OpenAI 创始人 Ilya Sutskever 在因董事会纠纷而沉默数月后辞职。",
    "link": "https://x.com/ilyasut/status/1790517455628198322"
  },
  {
    "date": "2024-05-21",
    "title": "EU AI Act",
    "category": "POLICY",
    "description": "经过激烈的辩论,欧盟人工智能法案被投票通过成为法律。",
    "link": "https://www.europarl.europa.eu/news/en/press-room/20240308IPR19015/artificial-intelligence-act-meps-adopt-landmark-law"
  },
  {
    "date": "2024-06-04",
    "title": "Situational Awareness",
    "category": "CULTURE",
    "description": "Leopold Aschenbrenner 发表了一系列有争议且有影响力的文章,声称 AGI 的到来将比人们想象的要早,并且很可能被国有化。",
    "link": "https://situational-awareness.ai/"
  },
  {
    "date": "2024-06-08",
    "title": "Qwen2",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2,Qwen2 拥有五种不同尺寸的尖端模型:Qwen2-0.5B、Qwen2-1.5B、Qwen2-7B、Qwen2-57B-A14B (MoE) 和 Qwen2-72B。这些模型支持 27 种语言,并且在代码和数学方面的能力显著增强。",
    "link": "https://qwenlm.github.io/zh/blog/qwen2/"
  },
  {
    "date": "2024-06-19",
    "title": "SSI 成立",
    "category": "BUSINESS",
    "description": "Ilya Sutskever 成立了一家名为 Safe Superintelligence Inc 的新实验室,该实验室承诺只生产一种产品:安全的超级智能。",
    "link": "https://x.com/ilyasut/status/1803472978753303014"
  },
  {
    "date": "2024-06-20",
    "title": "Claude 3.5 Sonnet",
    "category": "MODEL_RELEASE",
    "description": "Anthropic 发布了 Claude 3.5 Sonnet,它将成为粉丝的最爱,后来被称为“伯克利最合格的单身汉”。",
    "link": "https://www.anthropic.com/news/claude-3-5-sonnet"
  },
  {
    "date": "2024-06-27",
    "title": "Gemma 2",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Google 发布 Gemma 3,Gemma 2 有 90 亿 (9B) 和 270 亿 (27B) 两种参数大小,与第一代相比,其性能更高、推理效率更高,并且内置了显著的安全改进。事实上,27B 的参数大小足以与比其大两倍的模型相媲美,其性能直到去年 12 月才在专有模型中实现。",
    "link": "https://blog.google/technology/developers/google-gemma-2/"
  },
  {
    "date": "2024-07-16",
    "title": "MathStral",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Mistral 发布了 MathStral,这是一个专为数学推理和科学发现而设计的特定 7B 模型。该模型具有 32k 上下文窗口,根据 Apache 2.0 许可证发布。",
    "link": "https://mistral.ai/news/mathstral"
  },
  {
    "date": "2024-07-16",
    "title": "Codestral Mamba",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Mistral 发布了 Codestral Mamba,一种专门用于代码生成的 Mamba2 语言模型,可在 Apache 2.0 许可下使用。",
    "link": "https://mistral.ai/news/codestral-mamba"
  },
  {
    "date": "2024-07-18",
    "title": "GPT-4o-mini",
    "category": "MODEL_RELEASE", 
    "description": "相当于是能力更强的 GPT-3.5,同时支持文本和图像。GPT-4o mini 成本比 GPT-3.5 Turbo便宜超过60%。",
    "link": "https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/"
  },
  {
    "date": "2024-07-18",
    "title": "Mistral NeMo",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Mistral 发布了 Mistral NeMo,最新推出的最佳小型模型。这是一种最先进的 12B 模型,上下文长度为 128k,与 NVIDIA 合作开发,并根据 Apache 2.0 许可发布。",
    "link": "https://mistral.ai/news/mistral-nemo"
  },
  {
    "date": "2024-07-24",
    "title": "Mistral Large 2",
    "category": "MODEL_RELEASE",
    "description": "Mistral 发布了 Mistral Large 2,与前代产品相比,Mistral Large 2 在代码生成、数学和推理方面的能力显著增强。它还提供了更强大的多语言支持和高级函数调用功能。",
    "link": "https://mistral.ai/news/mistral-large-2407"
  },
  {
    "date": "2024-08-08",
    "title": "Qwen2-Math",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2-Math,Qwen2-Math 是一系列基于 Qwen2 LLM 构建的专门用于数学解题的语言模型,其数学能力显著超越了开源模型,甚至超过了闭源模型(如 GPT-4o)。",
    "link": "https://qwenlm.github.io/zh/blog/qwen2-math/"
  },
  {
    "date": "2024-08-9",
    "title": "Qwen2-Audio",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2-Audio,这是 Qwen-Audio 的下一代版本,它能够接受音频和文本输入,并生成文本输出,特点:语音聊天、音频分析、多语言支持。",
    "link": "https://qwenlm.github.io/zh/blog/qwen2-audio/"
  },
  {
    "date": "2024-08-23",
    "title": "Cursor",
    "category": "BUSINESS",
    "description": "在 Andrej Karpathy 发布了一条病毒式推文后,Cursor AI 代码编辑器在开发者中迅速走红。",
    "link": "https://x.com/karpathy/status/1827143768459637073"
  },
  {
    "date": "2024-08-24",
    "title": "Grok 2",
    "category": "MODEL_RELEASE",
    "description": "xAI 发布了 Grok 2,这是其前沿模型的下一代。虽然它不是最先进的,但它展示了 xAI 尽管起步较晚,但能够以多快的速度追赶上来。",
    "link": "https://x.ai/news/grok-2"
  },
  {
    "date": "2024-08-28",
    "title": "Claude Artifacts",
    "category": "MODEL_RELEASE",
    "description": "借助 Artifacts,您可以在一个专用窗口中即时查看、迭代和构建,让你与 Claude 一起创建作品。",
    "link": "https://www.anthropic.com/news/artifacts"
  },
  {
    "date": "2024-08-29",
    "title": "Qwen2-VL",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2-VL,这是基于 Qwen2 构建的视觉语言模型的最新版本。",
    "link": "https://qwenlm.github.io/zh/blog/qwen2-vl/"
  },
  {
    "date": "2024-09-02",
    "title": "xAI Colossus",
    "category": "BUSINESS",
    "description": "xAI 推出了 Colossus,这是当时世界上最强大的人工智能训练系统,拥有 100,000 个 H100 GPU 集群。 从第一个硬件机架到达到着手开始训练操作仅用了 19 天,xAI 构建集群的速度让其他 AI 实验室感到震惊。",
    "link": "https://x.com/elonmusk/status/1830650370336473253"
  },
  {
    "date": "2024-09-12",
    "title": "o1-preview/o1-mini",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布了 o1-preview/o1-mini, 介绍了推理时扩展范式。",
    "link": "https://openai.com/index/introducing-openai-o1-preview/"
  },
  {
    "date": "2024-09-17",
    "title": "Pixtral 12B",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Mistral 发布了 Pixtral 12B,这是 Mistral 首个多模式 Mistral 模型。",
    "link": "https://pixtral.com/news/pixtral-12b"
  },
  {
    "date": "2024-09-19",
    "title": "Qwen2.5",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2.5。最新发布包括了语言模型 Qwen2.5,以及专门针对编程的 Qwen2.5-Coder 和数学的 Qwen2.5-Math 模型。所有开放权重的模型都是稠密的、decoder-only的语言模型,提供多种不同规模的版本",
    "link": "https://qwenlm.github.io/zh/blog/qwen2.5/"
  },
  {
    "date": "2024-09-25",
    "title": "OpenAI CTO Murati 离职 ",
    "category": "BUSINESS",
    "description": "OpenAI 的首席技术官 Mira Murati 离开了公司。",
    "link": "https://x.com/miramurati/status/1839025700009030027"
  },
  {
    "date": "2024-10-04",
    "title": "OpenAI Canvas",
    "category": "MODEL_RELEASE",
    "feature": "在写作和代码方面展开协作",
    "description": "为ChatGPT引入新的写作和编程界面,提升用户与AI协作的体验。类似于 Claude 的 Artifacts"
  },
  {
    "date": "2024-10-08",
    "title": "Hinton、Hassabis 获得诺贝尔奖",
    "category": "CULTURE",
    "description": "令所有人惊讶的是,Geoffrey Hinton(与 John Hopfield 一起)因其在神经网络方面的早期工作而被授予诺贝尔物理学奖。几天后,Demis Hassabis(与 John Jumper 一起)因其在 AlphaFold 方面的工作而被授予诺贝尔化学奖。",
    "link": ""
  },
  {
    "date": "2024-10-11",
    "title": "Machines of Loving Grace",
    "category": "CULTURE",
    "description": "Anthropic 首席执行官 Dario Amodei 发表了一篇有影响力的博文,探讨了紧随 AGI 之后的 5 年可能会是什么样子。",
    "link": "https://darioamodei.com/machines-of-loving-grace"
  },
  {
    "date": "2024-10-16",
    "title": "Mistral 3B 和 Mistral 8B",
    "category": "MODEL_RELEASE",
    "description": "Mistral 发布了 Mistral 3B 和 Mistral 8B,两款用于设备计算和边缘用例的全新先进模型。",
    "link": "https://mistral.ai/news/ministraux"
  },
  {
    "date": "2024-10-22",
    "title": "Claude Computer Use",
    "category": "MODEL_RELEASE",
    "description": "Claude 获得了使用计算机界面的能力。Anthropic 还发布了 Claude 3.5 Haiku 和 Claude 3.5 Sonnet 的更新版本。",
    "link": "https://www.anthropic.com/news/3-5-models-and-computer-use"
  },
  {
    "date": "2024-10-31",
    "title": "ChatGPT 搜索功能",
    "category": "MODEL_RELEASE",
    "description": "ChatGPT整合了实时互联网信息,提升了回答的准确性和时效性。",
    "link": "https://chatgpt.com/"
  },
  {
    "date": "2024-11-12",
    "title": "Qwen2.5-Coder-32B-Instruct",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2.5-Coder-32B-Instruct,:Qwen2.5-Coder-32B-Instruct 成为目前 SOTA 的开源代码模型,代码能力追平 GPT-4o,展现出强大且全面的代码能力,同时具备良好的通用和数学能力。本次开源又带来 0.5B、3B、14B、32B 四个尺寸,截至目前, Qwen2.5-Coder 已经覆盖了主流的六个模型尺寸,以满足不同开发者的需要。",
    "link": "https://qwenlm.github.io/zh/blog/qwen2.5-coder-family//"
  },
  {
    "date": "2024-11-15",
    "title": "Qwen2.5-Turbo",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2.5-Turbo,1M 上下文、推理速度更快、成本更低",
    "link": "https://qwenlm.github.io/zh/blog/qwen2.5-turbo/"
  },
  {
    "date": "2024-11-18",
    "title": "Pixtral Large",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Mistral 发布了 Pixtral Large,这是一个基于 Mistral Large 2 构建的 124B 开放权重多模态模型。",
    "link": "https://mistral.ai/news/pixtral-large"
  },
  {
    "date": "2024-11-25",
    "title": "MCP 协议",
    "category": "RESEARCH",
    "description": "模型上下文协议(MCP)是一个开放协议,它定义了一种标准化的方式,使得应用程序能够为大型语言模型(LLMs)提供上下文信息 。其核心目标是实现 AI 模型与外部工具、数据库和 API 之间的无缝且标准化的集成。",
    "link": "https://www.anthropic.com/news/model-context-protocol"
  },
  {
    "date": "2024-11-28",
    "title": "QwQ-32B-Preview",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 QwQ-32B-Preview,这是一个旨在提高 AI 推理能力的开放模型。尤其是在解决数学和编码方面的一些挑战方面能力比较卓越",
    "link": "https://qwenlm.github.io/zh/blog/qwq-32b-preview/"
  },
  {
    "date": "2024-12-06",
    "title": "o1 & ChatGPT Pro",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布 o1 模型,支持图像输入,比 o1-preview 思考时间更短,但响应更快。同时推出 ChatGPT Pro 订阅,每月 200 美元,不限制使用次数,包括 o1、o1-mini、语音模式等,并提供更智能的 o1 使用模式。",
    "link": "https://chatgpt.com/"
  },
  {
    "date": "2024-12-10",
    "title": "Sora 开放使用",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 的 Sora 模型正式开放使用,支持文本转视频、图像转视频、视频转视频等多种功能。",
    "link": "https://sora.com/"
  },
  {
    "date": "2024-12-13",
    "title": "ChatGPT 高级视频模式开放使用",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布高级视频模式",
    "link": "https://chatgpt.com/"
  },
  {
    "date": "2024-12-11",
    "title": "Gemini 2.0",
    "category": "MODEL_RELEASE",
    "description": "Google 宣布了他们的 Gemini 2.0 模型",
    "link": "https://blog.google/products/gemini/google-gemini-ai-collection-2024/"
  },
  {
    "date": "2024-12-16",
    "title": "Veo 2",
    "category": "MODEL_RELEASE",
    "description": "Google 推出了 Veo 2,这是一款视频生成模型,其连贯性比以前的模型有了惊人的飞跃。",
    "link": "https://deepmind.google/technologies/veo/veo-2/"
  },
  {
    "date": "2024-12-20",
    "title": "o3 evals",
    "category": "RESEARCH",
    "description": "在“OpenAI 的 12 天直播日”的第 12 天,OpenAI 发布了 o3 的基准测试结果,震惊了世界。 该模型在 ARC-AGI 基准测试中取得了 87.5% 的突破性分数,表明 AGI 可能比许多怀疑论者认为的更近。",
    "link": "https://openai.com/12-days/"
  },
  {
    "date": "2024-12-25",
    "title": "QVQ-72B-Preview",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 QVQ-72B-Preview,一个基于 Qwen2-VL-72B 构建的开源多模态推理模型。QVQ 在人工智能的视觉理解和复杂问题解决能力方面实现了重大突破。",
    "link": "https://qwenlm.github.io/zh/blog/qvq-72b-preview/"
  },
  {
    "date": "2024-12-26",
    "title": "DeepSeek v3",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "中国实验室 DeepSeek 发布了 DeepSeek v3,这是一款 6710 亿参数的开源模型,该模型以惊人的低成本表现出强大的性能,令人震惊。",
    "link": "https://arxiv.org/abs/2412.19437"
  },
  {
    "date": "2025-01-13",
    "title": "Codestral 25.01",
    "category": "MODEL_RELEASE",
    "description": "Mistral 发布了 Codestral 25.01,Codestral 25.01 具有比原版更高效的架构和改进的标记器,生成和完成代码的速度提高了约 2 倍。该模型现在是其重量级编码的明显领导者,并且是 FIM 用例的 SOTA。",
    "link": "https://mistral.ai/news/codestral-2501"
  },
  {
    "date": "2025-01-20",
    "title": "DeepSeek R1",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "DeepSeek 发布并开源了 R1,他们的推理模型显示出与最先进的西方模型相比具有竞争力的性能。",
    "link": "https://api-docs.deepseek.com/news/news250120"
  },
  {
    "date": "2025-01-20",
    "title": "豆包实时语音大模型",
    "category": "MODEL_RELEASE",
    "description": "豆包发布实时语音大模型,使用原生方法深度融合语音与文本模态,模型具备超低延时、真人级别的语音表现、多模态I/O、多种能力涌现。",
    "link": "https://team.doubao.com/zh/special/realtime_voice"
  },
  {
    "date": "2025-01-21",
    "title": "Stargate Project",
    "category": "BUSINESS",
    "description": "唐纳德·特朗普宣布了星际之门项目,这是一项由软银、OpenAI、甲骨文和 MGX 之间建立的价值 500 亿美元的私人合作伙伴关系,用于在美国开发数据中心。",
    "link": "https://openai.com/index/announcing-the-stargate-project/"
  },
  {
    "date": "2025-01-22",
    "title": "Doubao-1.5-pro",
    "category": "MODEL_RELEASE",
    "description": "Doubao 发布了 Doubao-1.5-pro,模型使用 MoE 架构,并通过训练-推理一体化设计,探索模型性能和推理性能之间的极致平衡。Doubao-1.5-pro 仅用较小激活参数,即可超过一流超大稠密预训练模型的性能,并在多个评测基准上取得优异成绩。",
    "link": "https://team.doubao.com/zh/special/doubao_1_5_pro"
  },
  {
    "date": "2025-01-23",
    "title": "Operator",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 推出了 Operator,一种可以自主使用计算机的 AI Agent。",
    "link": "https://openai.com/index/introducing-operator/"
  },
  {
    "date": "2025-01-27",
    "title": "DeepSeek 火爆全球",
    "category": "CULTURE",
    "description": "R1 模型发布一周后,西方国家对 DeepSeek 产生了巨大的恐慌。芯片股一夜暴跌,DeepSeek 应用程序升至 App Store 排名第一。几天之内,这个鲜为人知的中国通用人工智能实验室就在美国家喻户晓。",
    "link": "https://nymag.com/intelligencer/article/deepseek-r1-ai-panic-impact-commentary-analysis.html"
  },
  {
    "date": "2025-01-27",
    "title": "Qwen2.5-1M",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2.5-7B-Instruct-1M 和 Qwen2.5-14B-Instruct-1M,这是 Qwen 首次将开源 Qwen 模型的上下文扩展到 1M 长度。",
    "link": "https://qwenlm.github.io/zh/blog/qwen2.5-1m/"
  },
  {
    "date": "2025-01-28",
    "title": "Qwen2.5-Max",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2.5-Max,这是一个大型 MoE,使用海量数据进行预训练,并使用精选的 SFT 和 RLHF 配方进行后训练。它与顶级模型相比具有竞争力,并且在 Arena Hard、LiveBench、LiveCodeBench、GPQA-Diamond 等基准测试中胜过 DeepSeek V3。",
    "link": "https://qwenlm.github.io/zh/blog/qwen2.5-max/"
  },
  {
    "date": "2025-01-28",
    "title": "Qwen2.5-VL",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2.5-VL,这是 Qwen 最新的旗舰视觉语言模型。主要特点:视觉理解、代理能力、长视频理解、精确定位、结构化数据输出",
    "link": "https://qwenlm.github.io/zh/blog/qwen2.5-vl/"
  },
  {
    "date": "2025-01-30",
    "title": "Mistral Small 3",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Mistral 发布了 Mistral Small 3,Mistral Small 3 与 Llama 3.3 70B 或 Qwen 32B 等大型模型相媲美,是 GPT4o-mini 等不透明专有模型的绝佳开放式替代品。Mistral Small 3 与 Llama 3.3 70B instruct 相当,但在相同硬件上速度却快 3 倍以上。",
    "link": "https://mistral.ai/news/mistral-small-3"
  }, 
  {
    "date": "2025-01-31",
    "title": "o3-mini",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布了 o3-mini,这是一款针对 STEM 推理进行了优化的模型。在中等推理强度下,o3-mini 在数学、编程和科学方面的表现与 o1 持平,同时响应速度更快。",
    "link": "https://openai.com/index/introducing-o3-mini/"
  },
  {
    "date": "2025-02-02",
    "title": "Deep Research",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 推出了一款名为 Deep Research 的AI Agent,它可以通过重复的网络搜索来撰写 10 页的研究报告。",
    "link": "https://openai.com/index/introducing-deep-research/"
  },
  {
    "date": "2025-02-17",
    "title": "Mistral Saba",
    "category": "MODEL_RELEASE",
    "description": "Mistral 发布了 Mistral Saba,一个 24B 参数模型,基于来自中东和南亚各地精心策划的数据集进行训练。与超过其 ​​5 倍大小的模型相比,该模型提供了更准确、更相关的响应,同时速度更快、成本更低。",
    "link": "https://mistral.ai/news/mistral-saba"
  },
  {
    "date": "2025-02-18",
    "title": "Thinking Machines Lab",
    "category": "BUSINESS",
    "description": "包括 Mira Murati 和 John Schulman 在内的前 OpenAI 关键人物创立了 Thinking Machines Lab,这是一家专注于人机协作、个性化和开放科学的新型人工智能实验室。",
    "link": "https://x.com/thinkymachines/status/1891919141151572094"
  },
  {
    "date": "2025-02-19",
    "title": "Grok 3",
    "category": "MODEL_RELEASE",
    "description": "xAI 发布了 Grok 3,这是一种具有扩展推理和深度搜索功能的最先进模型。这一发布给许多人留下了深刻的印象,表明 xAI 是构建 AGI 竞赛中的有力竞争者。",
    "link": "https://x.ai/blog/grok-3"
  },
  {
    "date": "2025-02-24",
    "title": "Claude 3.7 Sonnet",
    "category": "MODEL_RELEASE",
    "description": "Anthropic 发布了 Claude 3.7 Sonnet,这是他们的第一个模型,具有扩展的思维能力和改进的数学和代码基准性能。为了好玩,他们还展示了其在口袋妖怪视频游戏中取得进展的非发行能力。此外,他们还发布了 Claude Code,一个强大的代理编码工具。",
    "link": "https://www.anthropic.com/news/claude-3-7-sonnet"
  },
  {
    "date": "2025-02-24",
    "title": "DeepSeek 开源周",
    "category": "OPEN_SOURCE",
    "description": "DeepSeek 于 2025 年 2 月 24 日至 28 日举办为期 5 天的开源周活动,届时将开源 5 个仓库:FlashMLA、DeepEP、DeepGEMM、DualPipe、EPLB 和 Profile-data、3FS 和 smallpond。这些开源的 5 个仓库构成他们在线服务的基础模块,都经过了详细的文档记录、部署和生产环境的严格测试。",
    "link": "https://github.com/deepseek-ai/FlashMLA"
  },
  {
    "date": "2025-02-27",
    "title": "GPT-4.5",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布了 GPT-4.5,这是他们最大的预训练模型和最后一个非推理模型。尽管在基准测试中没有表现出巨大的进步,但该模型因其“氛围”和更人性化的反应而受到吹捧。",
    "link": "https://openai.com/index/introducing-gpt-4-5/"
  },
  {
    "date": "2025-03-05",
    "title": "Manus",
    "category": "BUSINESS",
    "description": "一家中国公司推出了一款名为 Manus 的 LLM 代理,在 GAIA 等基准测试中表现出 SOTA 性能。这款代理在西方迅速走红,部分原因是人们对中国人工智能的担忧。",
    "link": "https://x.com/ManusAI_HQ/status/1897294098945728752"
  },
  {
    "date": "2025-03-05",
    "title": "QWQ-32B",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 QwQ-32B,这是 Qwen 最新的推理模型,它只有 320 亿个参数,可以与 DeepSeek-R1 等尖端推理模型相媲美。",
    "link": "https://qwenlm.github.io/zh/blog/qwq-32b/"
  },
  {
    "date": "2025-03-06",
    "title": "Mistral OCR",
    "category": ["MODEL_RELEASE"],
    "description": "Mistral 发布了 Mistral OCR,这是世界上最好的文档理解 API。",
    "link": "https://mistral.ai/news/mistral-ocr"
  },
  {
    "date": "2025-03-11",
    "title": "Seedream 2.0",
    "category": "MODEL_RELEASE",
    "description": "豆包发布 Seedream 2.0,一个原生、多维度的中英双语图像生成基础模型,可以完美处理中英文文本提示,支持双语图像生成与文本渲染。",
    "link": "https://team.doubao.com/zh/tech/seedream"
  },
  {
    "date": "2025-03-12",
    "title": "Gemma 3",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Google 发布 Gemma 3,这是一组轻量级、最先进的开放模型,采用与 Gemini 2.0 模型相同的研究和技术构建而成。它们旨在直接在设备(从手机、笔记本电脑到工作站)上快速运行,Gemma 3 有多种尺寸(1B、4B、12B 和 27B),可让您根据特定的硬件和性能需求选择最佳模型。",
    "link": "https://blog.google/technology/developers/gemma-3/?hl=zh-cn"
  },
  {
    "date": "2025-03-17",
    "title": "Mistral Small 3.1",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Mistral 发布了 Mistral Small 3.1,该新模型以 Mistral Small 3 为基础,具有改进的文本性能、多模式理解和高达 128k 个标记的扩展上下文窗口。该模型的表现优于 Gemma 3 和 GPT-4o Mini 等同类模型,同时提供每秒 150 个标记的推理速度。",
    "link": "https://mistral.ai/news/mistral-small-3-1"
  },
  {
    "date": "2025-03-24",
    "title": "DeepSeek-V3-0324",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "DeepSeek 发布了 DeepSeek-V3-0324 模型,该模型在推理能力、Web前端开发能力、中文写作能力、中文搜索能力以及 Function Calling 能力方面都有显著提升。",
    "link": "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324/tree/main"
  },
  {
    "date": "2025-03-24",
    "title": "Qwen2.5-VL-32B-Instruct",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2.5-VL-32B-Instruct。相比此前发布的 Qwen2.5-VL 系列模型,本次推出的 32B 模型的特点如下:回复更符合人类主观偏好、数学推理能力、图像细粒度理解与推理",
    "link": "https://qwenlm.github.io/zh/blog/qwen2.5-vl-32b/"
  },
  {
    "date": "2025-03-25",
    "title": "Gemini 2.5 Pro",
    "category": "MODEL_RELEASE",
    "description": "谷歌发布了 Gemini 2.5 Pro,这是该公司迄今为止功能最强大的型号,在许多常见的基准测试中名列前茅。",
    "link": "https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/"
  },
  {
    "date": "2025-03-25",
    "title": "GPT-4o 原生图像生成",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 发布 GPT-4o 原生图像生成功能,进一步推动图像生成的前沿。因此,推特上充斥着吉卜力工作室风格的图像。",
    "link": "https://openai.com/index/introducing-4o-image-generation/"
  },
  {
    "date": "2025-03-26",
    "title": "GPT-4o 原生图像生成因吉卜力画风火爆全球",
    "category": "CULTURE",
    "description": "OpenAI 发布 GPT-4o 原生图像生成功能,在推特上造成了病毒是的传播,一小时内 ChatGPT 新增 100 万用户。堪比上一波 DeepSeek 的火爆程度。",
    "link": "https://x.com/fun000001/status/1906892391656865812"
  },
  {
    "date": "2025-03-27",
    "title": "Qwen2.5-Omni-7B",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "阿里千问团队发布 Qwen2.5-Omni-7B,这是一个全能模型:一个模型可以理解文本、音频、图像、视频,并输出文本和音频。",
    "link": "https://qwenlm.github.io/zh/blog/qwen2.5-omni//"
  },
  {
    "date": "2025-03-31",
    "title": "AutoGLM 沉思",
    "category": "MODEL_RELEASE",
    "description": "智谱推出一款名为 AutoGLM 的 AI Agent,类似于 OpenAI 的 DeepResearch。AutoGLM 沉思是一个能探究开放式问题,并根据结果执行操作的自主智能体(AI Agent)。",
    "link": "https://autoglm-research.zhipuai.cn/"
  },
  {
    "date": "2025-04-05",
    "title": "Llama 4",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Meta 发布了 Llama 4,这是该公司迄今为止功能最强大的型号,该系列的核心是 Llama 4 Scout 和 Llama 4 Maverick 两款开放权重的模型,它们首次采用了 专家混合 (MoE) 架构,并具备处理文本、图像和视频的原生多模态能力,同时支持 100M 超长上下文窗口。",
    "link": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/"
  },
  {
    "date": "2025-04-09",
    "title": "A2A 协议",
    "category": "RESEARCH",
    "description": "Agent2Agent 协议(A2A)是 Google Cloud 及其众多技术与服务合作伙伴联合推出的一项开放协议,旨在推动 AI 代理(AI agents)之间的互操作性。A2A 协议允许不同厂商和框架构建的 AI 代理在企业应用和平台之间安全地交流、协作和协调任务,从而提升自动化水平和生产效率,降低长期成本。A2A 的设计目标是建立一个开放、标准化的多代理生态系统,支持跨平台、跨云的协同工作,加速企业流程创新和智能化转型。",
    "link": "https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/"
  },
  {
    "date": "2025-04-15",
    "title": "GPT-4.1",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 推出了全新系列的 GPT-4.1 模型,包括 GPT‑4.1、GPT‑4.1 mini 和 GPT‑4.1 nano,这些模型在编码、指令遵循和长上下文处理方面实现了重大突破。与之前的 GPT‑4o 系列相比,GPT‑4.1 全面提升了性能,并降低了成本和延迟,特别适合开发者在实际应用中构建更强大的智能系统和代理型应用。新模型还首次支持高达 1 百万 tokens 的超长上下文窗口,并在多项行业标准基准测试中刷新记录。GPT‑4.1 系列仅通过 API 提供,GPT‑4.5 Preview 将于 2025 年 7 月 14 日下线。",
    "link": "https://openai.com/index/gpt-4-1/"
  },
  {
    "date": "2025-04-17",
    "title": "o3 和 o4-mini",
    "category": "MODEL_RELEASE",
    "description": "OpenAI 推出了其 o 系列中迄今为止最智能、能力最强的模型——o3 和 o4-mini。这两款模型被训练用于更深度的推理(“思考更长时间”),显著提升了 ChatGPT 的能力。模型首次能够自主地(agentically)决定何时以及如何使用 ChatGPT 内的所有工具(网络搜索、代码执行、视觉分析、图像生成等)来解决复杂问题。",
    "link": "https://openai.com/index/introducing-o3-and-o4-mini/"
  },
  {
    "date": "2025-04-17",
    "title": "豆包 1.5 深度思考模型",
    "category": "MODEL_RELEASE",
    "description": "豆包 1.5 深度思考模型发布:暴砍参数量,能看图思考,数学编程超DeepSeek-R1",
    "link": "https://seed.bytedance.com/zh/tech/seed1_5_vl"
  },
  {
    "date": "2025-04-18",
    "title": "Gemini 2.5 Flash Preview 04-17",
    "category": "MODEL_RELEASE",
    "description": "Gemini 2.5 Flash 是 Google 首款完全混合推理模型,引入了可控的“思考”(thinking)过程。模型在生成响应前可以执行“思考”过程,以更好地理解提示、分解复杂任务和规划回答,尤其适用于需要多步推理的复杂问题(如数学、分析)。",
    "link": "https://developers.googleblog.com/zh-hans/start-building-with-gemini-25-flash/"
  },
  {
    "date": "2025-04-29",
    "title": "Qwen3",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "Qwen3 是一个性能强大的大型语言模型系列,包含多种规模的模型,旨在提供顶尖的性能和广泛的应用能力。其旗舰模型 Qwen3-235B-A22B 在多项基准测试中表现出色,可与业界顶级模型媲美,同时较小模型也实现了显著的性能提升。该系列引入了创新的双思考模式、广泛的多语言支持和增强的 Agent 能力。通过开源部分模型(包括两个 MoE 模型和六个 Dense 模型)。包括旗舰模型 Qwen3-235B-A22B (235B+ 总参数, 22B+ 激活参数) 和小型 MoE 模型 Qwen3-30B-A3B (约 30B 总参数, 3B 激活参数) 以及六个 Dense 模型 (Qwen3-32B 至 Qwen3-0.6B),均采用 Apache 2.0 许可。",
    "link": "https://qwenlm.github.io/zh/blog/qwen3/"
  },
  {
    "date": "2025-04-30",
    "title": "DeepSeek-Prover-V2",
    "category": ["MODEL_RELEASE", "OPEN_SOURCE"],
    "description": "DeepSeek 推出了 DeepSeek-Prover-V2,一个用于 Lean 4 形式化定理证明的开源大语言模型。模型训练始于一个“冷启动”阶段,利用 DeepSeek-V3 通过递归证明流程生成初始数据,整合了非形式化推理(如思路链)和形式化证明步骤。随后通过强化学习进一步提升性能。DeepSeek-Prover-V2 在神经定理证明领域达到 SOTA 水平,特别是在 MiniF2F-test 上实现了 88.9 % 的通过率,并解决了 PutnamBench 中的 49 个问题。发布了两个尺寸的模型: 7B 和 671B 参数,可通过 Hugging Face Transformers 使用。",
      "link": "https://huggingface.co/deepseek-ai/DeepSeek-Prover-V2-671B"
    },
    {
      "date": "2025-05-15",
      "title": "AlphaEvolve",
      "category": "RESEARCH",
      "description": "Google 发布 AlphaEvolve: 一款利用大型语言模型(LLMs)进行通用算法发现与优化的进化编码代理。其核心思想是结合 Gemini 模型的创造性问题解决能力和自动化评估机制,通过进化迭代的方式生成、验证并改进算法。该代理不仅能够优化现有计算流程,还能在数学等基础科学领域探索新的解决方案,展示了其在多个领域应用的广泛前景和实际价值。",
      "link": "https://fisherdaddy.com/posts/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/"
    },
    {
      "date": "2025-05-16",
      "title": "OpenAI Codex",
      "category": "MODEL_RELEASE",
      "description": "OpenAI 推出全新 Codex,一款云端AI软件工程代理。基于codex-1,它能并行处理编码、修复Bug、提PR等任务,助开发者提升效率。现已向 ChatGPT Pro 、Team 和 Enterprise 用户开放。",
      "link": "https://fisherdaddy.com/posts/introducing-openai-codex/"
    },
    {
      "date": "2025-05-20",
      "title": "Google I/O 2025:Goole Veo 3、Imagen 4、Flow、Gemini 2.5 Pro Deep Think",
      "category": "MODEL_RELEASE",
      "description": "Google I/O 2025 发布了大量的 Google 的最新能力,包括 Gemini 2.5 Pro Deep Think、Imagen 4、Flow、Google Veo 3、Lyria 2 等。",
      "link": "https://fisherdaddy.com/posts/google-io-2025-collection-first-day/"
    },
    {
      "date": "2025-05-23",
      "title": "Claude 4",
      "category": "MODEL_RELEASE",
      "description": "Anthropic 推出了新一代 Claude 模型:Claude Opus 4 和 Claude Sonnet 4,它们在编码、高级推理和 AI 代理方面树立了新标准。Claude Opus 4 被誉为全球最佳编码模型,擅长处理复杂和长时间运行的任务,而 Claude Sonnet 4 则是对 Sonnet 3.7 的重大升级,在编码和推理方面表现卓越。两款模型都增强了工具使用、并行工具执行和记忆能力。",
      "link": "https://fisherdaddy.com/posts/introduce-claude4/"
    },    
    {
      "date": "2025-05-29",
      "title": "DeepSeek-R1-0528",
      "category": "MODEL_RELEASE",
      "description": "DeepSeek-R1-0528 发布:推理与理解能力显著增强,性能逼近顶尖模型,例如 O3 和 Gemini 2.5 Pro。",
      "link": "https://fisherdaddy.com/posts/deepseek-r1-0528-details/"
    }
]