{
 "meta": {
  "title": "Alpha 图谱 · DeepMind Alpha Family Tracker",
  "lastUpdated": "2026-11-13",
  "maintainer": "Slovian Inc",
  "disclaimer": "独立整理的研究索引，与 Google DeepMind 无隶属关系。所有条目均附原始来源，如有出入以来源为准。"
 },
 "disciplines": [
  {
   "key": "games",
   "zh": "博弈与强化学习",
   "en": "Games & RL",
   "color": "#7c8cff"
  },
  {
   "key": "biology",
   "zh": "结构生物学与蛋白质",
   "en": "Structural Biology",
   "color": "#2fbf8f"
  },
  {
   "key": "genomics",
   "zh": "基因组学",
   "en": "Genomics",
   "color": "#e3a03a"
  },
  {
   "key": "math",
   "zh": "数学与形式证明",
   "en": "Mathematics",
   "color": "#e0617a"
  },
  {
   "key": "cs",
   "zh": "算法与计算机科学",
   "en": "Algorithms & CS",
   "color": "#4fb3e8"
  },
  {
   "key": "chips",
   "zh": "芯片设计",
   "en": "Chip Design",
   "color": "#a77cf0"
  },
  {
   "key": "quantum",
   "zh": "量子计算",
   "en": "Quantum Computing",
   "color": "#22c2c9"
  },
  {
   "key": "earth",
   "zh": "地球科学",
   "en": "Earth Science",
   "color": "#8bbf3a"
  },
  {
   "key": "security",
   "zh": "软件安全与形式化验证",
   "en": "Software Verification",
   "color": "#f07c4f"
  }
 ],
 "projects": [
  {
   "id": "alphago",
   "name": "AlphaGo",
   "date": "2016-01-27",
   "discipline": "games",
   "tags": [
    "games"
   ],
   "status": "superseded",
   "tagline": "首个击败人类职业围棋顶尖棋手的 AI",
   "problem": "围棋状态空间巨大，长期被视为 AI 难以攻克的“圣杯”。",
   "result": "以策略网络 + 价值网络结合蒙特卡洛树搜索；2016 年 3 月在首尔以 4:1 战胜李世石，第 2 局“第 37 手”成为经典。",
   "metric": "4:1 战胜李世石（2016.03）",
   "applications": [
    "推动深度强化学习研究浪潮",
    "思想被 AlphaZero、AlphaProof 等后续系统继承"
   ],
   "openSource": "未开源（论文公开）",
   "parents": [],
   "children": [
    "alphago-zero"
   ],
   "links": [
    {
     "label": "Nature 论文 (2016)",
     "url": "https://www.nature.com/articles/nature16961",
     "type": "paper"
    },
    {
     "label": "DeepMind 官方页面",
     "url": "https://deepmind.google/research/alphago/",
     "type": "official"
    },
    {
     "label": "AlphaGo at 10（2026 回顾）",
     "url": "https://deepmind.google/blog/10-years-of-alphago/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphago-zero",
   "name": "AlphaGo Zero",
   "date": "2017-10-19",
   "discipline": "games",
   "tags": [
    "games"
   ],
   "status": "superseded",
   "tagline": "不用人类棋谱，从零自我对弈学会围棋",
   "problem": "能否完全摆脱人类数据，仅凭规则与自我对弈达到超人水平？",
   "result": "单一网络 + 自我对弈强化学习，以 100:0 击败击败李世石的 AlphaGo 版本。",
   "metric": "对 AlphaGo Lee 版本 100:0",
   "applications": [
    "“从零学习”范式的里程碑"
   ],
   "openSource": "未开源（论文公开）",
   "parents": [
    "alphago"
   ],
   "children": [
    "alphazero"
   ],
   "links": [
    {
     "label": "Nature 论文 (2017)",
     "url": "https://www.nature.com/articles/nature24270",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphago-zero-starting-from-scratch/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphazero",
   "name": "AlphaZero",
   "date": "2018-12-07",
   "discipline": "games",
   "tags": [
    "games"
   ],
   "status": "superseded",
   "tagline": "一个算法通吃国际象棋、将棋与围棋",
   "problem": "把 AlphaGo Zero 的方法推广到多种棋类，迈向更通用的学习系统。",
   "result": "同一算法仅凭规则自学，在三种棋类中均击败当时的世界冠军级程序（如 Stockfish、Elmo）。",
   "metric": "三种棋类均达到超人水平",
   "applications": [
    "改变了国际象棋界对开局与弃子的理解",
    "其搜索+强化学习框架被 AlphaDev、AlphaTensor、AlphaProof 借用"
   ],
   "openSource": "未开源（论文公开）",
   "parents": [
    "alphago-zero"
   ],
   "children": [
    "muzero",
    "alphatensor",
    "alphadev",
    "alphaproof"
   ],
   "links": [
    {
     "label": "Science 论文 (2018)",
     "url": "https://www.science.org/doi/10.1126/science.aar6404",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphazero-shedding-new-light-on-chess-shogi-and-go/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "muzero",
   "name": "MuZero",
   "date": "2020-12-23",
   "discipline": "games",
   "tags": [
    "games",
    "cs"
   ],
   "status": "research",
   "related": true,
   "tagline": "（相关项目）不知道规则也能规划",
   "problem": "现实问题往往没有明确规则，系统需要自己学习环境模型。",
   "result": "学习环境的内部模型并据此规划，在围棋、国际象棋、将棋上达到 AlphaZero 水平，同时在 Atari 游戏上刷新纪录。",
   "metric": "无规则条件下匹配 AlphaZero",
   "applications": [
    "基于学习模型的规划（model-based RL）研究"
   ],
   "openSource": "未开源（论文公开）",
   "note": "MuZero 不以 Alpha 命名，但直接继承 AlphaZero，作为谱系参考收录。",
   "parents": [
    "alphazero"
   ],
   "children": [],
   "links": [
    {
     "label": "Nature 论文 (2020)",
     "url": "https://www.nature.com/articles/s41586-020-03051-4",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/muzero-mastering-go-chess-shogi-and-atari-without-rules/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphastar",
   "name": "AlphaStar",
   "date": "2019-10-30",
   "discipline": "games",
   "tags": [
    "games"
   ],
   "status": "superseded",
   "tagline": "《星际争霸 II》宗师级 AI",
   "problem": "不完全信息、实时、长时程、巨大动作空间的即时战略游戏。",
   "result": "多智能体强化学习 + 联赛训练，三个种族均达到 Battle.net 宗师（Grandmaster）段位，高于 99.8% 的排位玩家。",
   "metric": "宗师段位，超过 99.8% 活跃玩家",
   "applications": [
    "多智能体训练（league training）方法论"
   ],
   "openSource": "未开源（论文公开；部分对局回放公开）",
   "parents": [],
   "children": [],
   "links": [
    {
     "label": "Nature 论文 (2019)",
     "url": "https://www.nature.com/articles/s41586-019-1724-z",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphastar-grandmaster-level-in-starcraft-ii-using-multi-agent-reinforcement-learning/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphafold-1",
   "name": "AlphaFold (v1)",
   "date": "2018-12-02",
   "discipline": "biology",
   "tags": [
    "biology"
   ],
   "status": "superseded",
   "tagline": "CASP13 冠军，首次让深度学习领跑蛋白质结构预测",
   "problem": "仅从氨基酸序列预测蛋白质 3D 结构——50 年的“蛋白质折叠问题”。",
   "result": "用深度网络预测残基间距离分布，再优化得到结构，在 CASP13（2018）中排名第一；论文 2020 年 1 月发表于 Nature。",
   "metric": "CASP13 第一名",
   "applications": [
    "奠定 AlphaFold 2 的基础"
   ],
   "openSource": "论文公开",
   "note": "日期取 CASP13 结果公布时间（2018 年 12 月），论文发表于 2020-01。",
   "parents": [],
   "children": [
    "alphafold-2"
   ],
   "links": [
    {
     "label": "Nature 论文 (2020)",
     "url": "https://www.nature.com/articles/s41586-019-1923-7",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphafold-using-ai-for-scientific-discovery-2020/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphafold-2",
   "name": "AlphaFold 2",
   "date": "2020-11-30",
   "discipline": "biology",
   "tags": [
    "biology"
   ],
   "status": "active",
   "tagline": "基本解决单链蛋白质结构预测，2024 年诺贝尔化学奖成果",
   "problem": "达到接近实验精度的蛋白质结构预测。",
   "result": "在 CASP14（2020）中达到原子级精度，被组织方称为解决了 50 年的重大挑战；2021 年 7 月在 Nature 发表并开源。Demis Hassabis 与 John Jumper 因此分享 2024 年诺贝尔化学奖。",
   "metric": "CASP14 原子级精度",
   "applications": [
    "药物发现与靶点研究",
    "酶设计、疫苗研究",
    "全球数百万研究者使用"
   ],
   "openSource": "代码与权重开源（GitHub）",
   "parents": [
    "alphafold-1"
   ],
   "children": [
    "alphafold-db",
    "alphafold-multimer",
    "alphafold-3",
    "alphamissense"
   ],
   "links": [
    {
     "label": "Nature 论文 (2021)",
     "url": "https://www.nature.com/articles/s41586-021-03819-2",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客（CASP14）",
     "url": "https://deepmind.google/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology/",
     "type": "blog"
    },
    {
     "label": "GitHub: alphafold",
     "url": "https://github.com/google-deepmind/alphafold",
     "type": "code"
    }
   ]
  },
  {
   "id": "alphafold-db",
   "name": "AlphaFold Protein Structure Database",
   "date": "2021-07-22",
   "discipline": "biology",
   "tags": [
    "biology"
   ],
   "status": "active",
   "tagline": "与 EMBL-EBI 合作，免费公开 2 亿+ 蛋白质结构",
   "problem": "让全世界研究者免费获得已知蛋白质的预测结构。",
   "result": "2021 年上线；2022 年 7 月扩展至 2 亿多个结构，覆盖几乎所有已编目蛋白质（扩大 200 多倍）。",
   "metric": "2 亿+ 预测结构",
   "applications": [
    "基础生物学研究",
    "被忽视疾病研究",
    "可通过 Google Cloud 公共数据集批量下载"
   ],
   "openSource": "免费开放（CC-BY 4.0）",
   "parents": [
    "alphafold-2"
   ],
   "children": [],
   "links": [
    {
     "label": "AlphaFold DB（EMBL-EBI）",
     "url": "https://alphafold.ebi.ac.uk/",
     "type": "official"
    },
    {
     "label": "DeepMind 博客（2 亿结构）",
     "url": "https://deepmind.google/blog/alphafold-reveals-the-structure-of-the-protein-universe/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphafold-multimer",
   "name": "AlphaFold-Multimer",
   "date": "2021-10-04",
   "discipline": "biology",
   "tags": [
    "biology"
   ],
   "status": "superseded",
   "tagline": "把 AlphaFold 扩展到蛋白质复合物",
   "problem": "预测多条蛋白质链如何组装成复合物。",
   "result": "针对多链输入训练的 AlphaFold 版本，显著提升复合物界面预测精度。",
   "metric": "多链复合物预测",
   "applications": [
    "蛋白质相互作用研究"
   ],
   "openSource": "集成在 AlphaFold 2 开源代码中",
   "parents": [
    "alphafold-2"
   ],
   "children": [
    "alphafold-3"
   ],
   "links": [
    {
     "label": "bioRxiv 预印本",
     "url": "https://www.biorxiv.org/content/10.1101/2021.10.04.463034",
     "type": "paper"
    }
   ]
  },
  {
   "id": "alphafold-3",
   "name": "AlphaFold 3",
   "date": "2024-05-08",
   "discipline": "biology",
   "tags": [
    "biology"
   ],
   "status": "active",
   "tagline": "预测蛋白质、DNA、RNA、配体等所有生命分子的相互作用",
   "problem": "从单一蛋白扩展到生物分子复合体（核酸、小分子、离子、修饰）。",
   "result": "与 Isomorphic Labs 联合发布，采用扩散式结构生成；对蛋白-配体等相互作用的预测较既有方法显著提升。2024 年 11 月向学术界开放代码与权重；2026 年代码许可改为 Apache 2.0（权重条款不变），v3.0.5 于 2026-10-30 发布。",
   "metric": "全分子类型复合物结构预测",
   "applications": [
    "药物设计（Isomorphic Labs）",
    "AlphaProtein Novo 等设计管线的评估器"
   ],
   "openSource": "推理代码开源（Apache 2.0）；权重需申请，限非商业",
   "parents": [
    "alphafold-2",
    "alphafold-multimer"
   ],
   "children": [
    "alphafold-server",
    "alphaproteo",
    "alphaprotein-novo"
   ],
   "links": [
    {
     "label": "Nature 论文 (2024)",
     "url": "https://www.nature.com/articles/s41586-024-07487-w",
     "type": "paper"
    },
    {
     "label": "Google 官方博客",
     "url": "https://blog.google/innovation-and-ai/products/google-deepmind-isomorphic-alphafold-3-ai-model/",
     "type": "blog"
    },
    {
     "label": "GitHub: alphafold3（含 2026 版本记录）",
     "url": "https://github.com/google-deepmind/alphafold3",
     "type": "code"
    },
    {
     "label": "GitHub Releases",
     "url": "https://github.com/google-deepmind/alphafold3/releases",
     "type": "code"
    }
   ]
  },
  {
   "id": "alphafold-server",
   "name": "AlphaFold Server",
   "date": "2024-05-08",
   "discipline": "biology",
   "tags": [
    "biology"
   ],
   "status": "active",
   "tagline": "免费网页版 AlphaFold 3",
   "problem": "让没有算力和编程背景的科研人员也能使用 AlphaFold 3。",
   "result": "随 AlphaFold 3 一同推出的免费在线平台，供非商业研究使用。",
   "metric": "免费网页访问",
   "applications": [
    "教学与非商业科研"
   ],
   "openSource": "免费在线服务（非商业）",
   "parents": [
    "alphafold-3"
   ],
   "children": [],
   "links": [
    {
     "label": "AlphaFold Server",
     "url": "https://alphafoldserver.com/",
     "type": "official"
    },
    {
     "label": "发布博客",
     "url": "https://blog.google/innovation-and-ai/products/google-deepmind-isomorphic-alphafold-3-ai-model/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphamissense",
   "name": "AlphaMissense",
   "date": "2023-09-19",
   "discipline": "genomics",
   "tags": [
    "genomics",
    "biology"
   ],
   "status": "active",
   "tagline": "为 7100 万种错义突变给出致病性预测",
   "problem": "绝大多数错义突变（改变单个氨基酸）的致病性未知。",
   "result": "基于 AlphaFold 改造，将 7100 万种可能错义突变中的 89% 分类为“可能致病”或“可能良性”；人类专家确认的仅 0.1%。",
   "metric": "89% / 7100 万变异完成分类",
   "applications": [
    "罕见病诊断研究",
    "与 AlphaGenome 组合成 AVI 评分（AlphaGenome Atlas）"
   ],
   "openSource": "预测目录与代码公开（非商业）",
   "parents": [
    "alphafold-2"
   ],
   "children": [
    "alphagenome-atlas"
   ],
   "links": [
    {
     "label": "Science 论文 (2023)",
     "url": "https://www.science.org/doi/10.1126/science.adg7492",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/a-catalogue-of-genetic-mutations-to-help-pinpoint-the-cause-of-diseases/",
     "type": "blog"
    },
    {
     "label": "GitHub: alphamissense",
     "url": "https://github.com/google-deepmind/alphamissense",
     "type": "code"
    }
   ]
  },
  {
   "id": "alphaproteo",
   "name": "AlphaProteo",
   "date": "2024-09-05",
   "discipline": "biology",
   "tags": [
    "biology"
   ],
   "status": "active",
   "tagline": "从头设计高亲和力蛋白结合体",
   "problem": "为目标蛋白设计能牢固结合的新蛋白，用于药物与生物传感器研究。",
   "result": "首次为 VEGF-A 设计出成功的结合体；在 7 个测试靶点上实验成功率更高，结合亲和力比现有最佳方法高 3–300 倍；对 TNFα 未能成功（官方明确列出的局限）。",
   "metric": "亲和力提升 3–300 倍（7 个靶点）",
   "applications": [
    "药物发现",
    "生物传感器",
    "疾病机制研究"
   ],
   "openSource": "未开源（技术报告公开）",
   "parents": [
    "alphafold-3"
   ],
   "children": [
    "alphaprotein-novo"
   ],
   "links": [
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphaproteo-generates-novel-proteins-for-biology-and-health-research/",
     "type": "blog"
    },
    {
     "label": "arXiv 技术报告",
     "url": "https://arxiv.org/abs/2409.08022",
     "type": "paper"
    }
   ]
  },
  {
   "id": "alphaprotein-novo",
   "name": "AlphaProtein Novo",
   "date": "2026-10-01",
   "discipline": "biology",
   "tags": [
    "biology"
   ],
   "status": "active",
   "tagline": "从头设计“自然界没有”的酶",
   "problem": "从零设计具有可观催化活性的酶，用于非天然底物与新化学反应。",
   "result": "扩散模型根据催化基序和配体环境同时生成蛋白结构与序列，结合 LigandMPNN 与 AlphaFold 3 评估组成端到端管线。媒体报道称已在湿实验中验证（如合成哌啶的氮宾转移酶、降解增塑剂 DEHP 的酶）。",
   "metric": "开源端到端酶设计管线",
   "applications": [
    "生物催化与制药合成",
    "环境污染物降解"
   ],
   "openSource": "代码开源（GitHub）；权重需按条款下载",
   "note": "日期取 bioRxiv 预印本（2026-10-01）；研究尚未经同行评审，湿实验细节来自二手报道。",
   "uncertain": true,
   "parents": [
    "alphaproteo",
    "alphafold-3"
   ],
   "children": [],
   "links": [
    {
     "label": "GitHub: alphaprotein-novo",
     "url": "https://github.com/google-deepmind/alphaprotein-novo",
     "type": "code"
    },
    {
     "label": "bioRxiv 预印本",
     "url": "https://www.biorxiv.org/content/10.64898/2026.10.01.756017v1",
     "type": "paper"
    },
    {
     "label": "IFAI 报道",
     "url": "https://ifai.io/en/news/deepmind-alpha-protein-novo",
     "type": "news"
    }
   ]
  },
  {
   "id": "alphatensor",
   "name": "AlphaTensor",
   "date": "2022-10-05",
   "discipline": "cs",
   "tags": [
    "cs",
    "math"
   ],
   "status": "superseded",
   "tagline": "AI 发现更快的矩阵乘法算法",
   "problem": "自 1969 年 Strassen 以来，小尺寸矩阵乘法所需的最少乘法次数仍是开放问题。",
   "result": "把算法发现变成单人“张量游戏”，用 AlphaZero 式强化学习求解，在多种尺寸上找到优于已知方法的算法（例如在模 2 下 4×4 矩阵用 47 次乘法）。",
   "metric": "首次改进 Strassen 4×4（模 2）结果",
   "applications": [
    "面向特定硬件的矩阵乘法优化"
   ],
   "openSource": "发现的算法公开（GitHub）",
   "parents": [
    "alphazero"
   ],
   "children": [
    "alphatensor-quantum",
    "alphaevolve"
   ],
   "links": [
    {
     "label": "Nature 论文 (2022)",
     "url": "https://www.nature.com/articles/s41586-022-05172-4",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/discovering-novel-algorithms-with-alphatensor/",
     "type": "blog"
    },
    {
     "label": "GitHub: alphatensor",
     "url": "https://github.com/google-deepmind/alphatensor",
     "type": "code"
    }
   ]
  },
  {
   "id": "alphatensor-quantum",
   "name": "AlphaTensor-Quantum",
   "date": "2024-02-22",
   "discipline": "quantum",
   "tags": [
    "quantum",
    "cs"
   ],
   "status": "research",
   "tagline": "用 AlphaTensor 优化量子电路",
   "problem": "减少容错量子计算中代价高昂的 T 门数量。",
   "result": "把 T 门数量优化建模为张量分解，并利用 AlphaTensor 方法在多个基准电路上降低 T 门数。",
   "metric": "T-count 优化",
   "applications": [
    "量子算法编译"
   ],
   "openSource": "预印本公开",
   "parents": [
    "alphatensor"
   ],
   "children": [],
   "links": [
    {
     "label": "arXiv 论文",
     "url": "https://arxiv.org/abs/2402.14396",
     "type": "paper"
    }
   ]
  },
  {
   "id": "alphadev",
   "name": "AlphaDev",
   "date": "2023-06-07",
   "discipline": "cs",
   "tags": [
    "cs"
   ],
   "status": "superseded",
   "tagline": "在汇编层面发现更快的排序算法",
   "problem": "排序、哈希等基础例程已被人类优化数十年，还能更快吗？",
   "result": "把汇编指令编写变成游戏，用强化学习发现新的排序例程；合入 LLVM libc++ 后，短序列排序最快提升约 70%，超大序列约 1.7%。",
   "metric": "短序列排序最快 +70%（LLVM libc++）",
   "applications": [
    "每天被调用数万亿次的基础库"
   ],
   "openSource": "排序代码已合入 LLVM libc++",
   "parents": [
    "alphazero"
   ],
   "children": [
    "alphaevolve"
   ],
   "links": [
    {
     "label": "Nature 论文 (2023)",
     "url": "https://www.nature.com/articles/s41586-023-06004-9",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphadev-discovers-faster-sorting-algorithms/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphacode",
   "name": "AlphaCode",
   "date": "2022-02-02",
   "discipline": "cs",
   "tags": [
    "cs"
   ],
   "status": "superseded",
   "tagline": "达到竞赛程序员中游水平的代码生成系统",
   "problem": "解决需要理解自然语言题意、设计算法的编程竞赛题。",
   "result": "大规模采样 + 过滤聚类，在 Codeforces 10 场新比赛中估计排名约前 54%；论文 2022 年 12 月发表于 Science。",
   "metric": "Codeforces 约前 54%",
   "applications": [
    "代码生成研究"
   ],
   "openSource": "竞赛数据集 CodeContests 公开",
   "parents": [],
   "children": [
    "alphacode-2"
   ],
   "links": [
    {
     "label": "Science 论文 (2022)",
     "url": "https://www.science.org/doi/10.1126/science.abq1158",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/competitive-programming-with-alphacode/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphacode-2",
   "name": "AlphaCode 2",
   "date": "2023-12-06",
   "discipline": "cs",
   "tags": [
    "cs"
   ],
   "status": "superseded",
   "tagline": "基于 Gemini 的竞赛编程系统",
   "problem": "借助更强的基础模型提升竞赛编程能力。",
   "result": "随 Gemini 发布，技术报告称其在评测比赛中估计表现优于约 85% 的参赛者。",
   "metric": "约 85 百分位（技术报告）",
   "applications": [
    "思路融入 Gemini 的代码与推理能力"
   ],
   "openSource": "未开源（技术报告公开）",
   "parents": [
    "alphacode"
   ],
   "children": [],
   "links": [
    {
     "label": "AlphaCode 2 技术报告 (PDF)",
     "url": "https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf",
     "type": "paper"
    }
   ]
  },
  {
   "id": "alphageometry",
   "name": "AlphaGeometry",
   "date": "2024-01-17",
   "discipline": "math",
   "tags": [
    "math"
   ],
   "status": "superseded",
   "tagline": "接近 IMO 金牌选手水平的几何证明系统",
   "problem": "奥数几何证明需要构造辅助线，且人类示范数据稀缺。",
   "result": "神经语言模型 + 符号推理引擎，用 1 亿条合成数据训练；在 30 道 IMO 几何题（IMO-AG-30）中限时解出 25 道，此前最佳方法为 10 道，人类金牌选手平均 25.9 道。",
   "metric": "IMO-AG-30：25/30",
   "applications": [
    "神经-符号结合的数学推理范式"
   ],
   "openSource": "代码与权重开源（GitHub）",
   "parents": [],
   "children": [
    "alphageometry-2"
   ],
   "links": [
    {
     "label": "Nature 论文 (2024)",
     "url": "https://www.nature.com/articles/s41586-023-06747-5",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/",
     "type": "blog"
    },
    {
     "label": "GitHub: alphageometry",
     "url": "https://github.com/google-deepmind/alphageometry",
     "type": "code"
    }
   ]
  },
  {
   "id": "alphageometry-2",
   "name": "AlphaGeometry 2",
   "date": "2024-07-25",
   "discipline": "math",
   "tags": [
    "math"
   ],
   "status": "active",
   "tagline": "几何题达到 IMO 金牌选手水平",
   "problem": "覆盖更广的几何语言并提升搜索与证明能力。",
   "result": "基于 Gemini 的语言模型与更强符号引擎；在 IMO 2024 中解出几何题。2025 年 2 月论文称其在 2000–2024 年 IMO 几何题上达到金牌选手水平。",
   "metric": "IMO 几何金牌水平（论文）",
   "applications": [
    "与 AlphaProof 组合参加 IMO 2024"
   ],
   "openSource": "未开源（论文公开）",
   "parents": [
    "alphageometry"
   ],
   "children": [],
   "links": [
    {
     "label": "arXiv 论文 (2025)",
     "url": "https://arxiv.org/abs/2502.03544",
     "type": "paper"
    },
    {
     "label": "IMO 2024 银牌水平博客",
     "url": "https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphaproof",
   "name": "AlphaProof",
   "date": "2024-07-25",
   "discipline": "math",
   "tags": [
    "math"
   ],
   "status": "active",
   "tagline": "用强化学习在 Lean 中寻找形式化证明，IMO 2024 银牌水平",
   "problem": "让 AI 产出可被机器严格验证的数学证明。",
   "result": "AlphaZero 式智能体在数百万自动形式化的题目上训练，并提出“测试时强化学习”；IMO 2024 中与 AlphaGeometry 2 合计 28/42 分（银牌上沿），解出含最难的第 6 题在内的 3 道非几何题。方法于 2025-11-12 发表于 Nature。",
   "metric": "IMO 2024：28/42（银牌水平）",
   "applications": [
    "形式化数学",
    "成为 AlphaProof Nexus 的专用证明工具"
   ],
   "openSource": "未开源（Nature 论文公开）",
   "note": "计算耗时为数天，远超人类考试时限（论文已说明）。",
   "parents": [
    "alphazero"
   ],
   "children": [
    "alphaproof-nexus",
    "alphaproof-code"
   ],
   "links": [
    {
     "label": "Nature 论文 (2025)",
     "url": "https://www.nature.com/articles/s41586-025-09833-y",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客（IMO 2024）",
     "url": "https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphaproof-nexus",
   "name": "AlphaProof Nexus",
   "date": "2026-10-08",
   "discipline": "math",
   "tags": [
    "math"
   ],
   "status": "active",
   "tagline": "自主攻克数十个开放数学问题，论文发表于 Science",
   "problem": "形式化证明能否从竞赛题走向真正的研究级开放问题？",
   "result": "多智能体在 Lean 编译器反馈下搜索证明，并可调用 AlphaProof；在 353 个尝试的 Erdős 问题中解决 9 个（其中 2 个悬而未决 50 多年），解决 492 个 OEIS 开放猜想中的 44 个，以及代数几何、优化、量子光学、图论中的若干问题。",
   "metric": "Erdős 问题 9/353；OEIS 猜想 44/492",
   "applications": [
    "数学家的 AI 合作者",
    "所有证明经 Lean 机器验证"
   ],
   "openSource": "Lean 证明与自然语言证明开源（GitHub）",
   "parents": [
    "alphaproof",
    "alphaevolve"
   ],
   "children": [],
   "links": [
    {
     "label": "EurekAlert!（Science 新闻稿）",
     "url": "https://www.eurekalert.org/news-releases/1146436",
     "type": "news"
    },
    {
     "label": "arXiv 预印本",
     "url": "https://arxiv.org/abs/2605.22763",
     "type": "paper"
    },
    {
     "label": "GitHub: 结果仓库",
     "url": "https://github.com/google-deepmind/alphaproof-nexus-results",
     "type": "code"
    }
   ]
  },
  {
   "id": "alphaproof-code",
   "name": "AlphaProof Code",
   "date": "2026-11-10",
   "discipline": "security",
   "tags": [
    "security",
    "math",
    "cs"
   ],
   "status": "active",
   "tagline": "AI 智能体形式化验证 Signal 协议的实际代码",
   "problem": "形式化验证是最高标准的安全保证，但极其耗人力，大多只停留在协议设计层面。",
   "result": "多 LLM 智能体框架：人类专家与 AI 确定规格，智能体用 Lean 写出 600 多万行证明，证明 libsignal 一对一聊天的关键安全性质（前向保密、认证、入侵后恢复），耗时约 4 周；重新引入的 2 个历史漏洞均被识别。",
   "metric": "约 600 万+ 行 Lean 证明，4 周完成",
   "applications": [
    "密码协议与关键软件安全",
    "验证产物开源供社区复核"
   ],
   "openSource": "形式化规格与证明开源",
   "note": "范围限定：一对一聊天、研究者自建的库版本，未覆盖群聊与密钥生成（据 TNW 对论文的解读）。",
   "parents": [
    "alphaproof"
   ],
   "children": [],
   "links": [
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/introducing-alphaproof-code/",
     "type": "blog"
    },
    {
     "label": "The Next Web 解读",
     "url": "https://thenextweb.com/news/google-deepmind-alphaproof-code-signal-encryption-formal-verification",
     "type": "news"
    }
   ]
  },
  {
   "id": "alphachip",
   "name": "AlphaChip",
   "date": "2021-06-09",
   "discipline": "chips",
   "tags": [
    "chips",
    "cs"
   ],
   "status": "active",
   "tagline": "用强化学习做芯片布局，用于历代 TPU",
   "problem": "芯片宏单元布局（floorplanning）耗时数周到数月。",
   "result": "2020 年发布预印本、2021 年发表于 Nature 并开源；2024 年正式命名 AlphaChip，官方称自 2020 年起每一代 TPU（含第 6 代 Trillium）都使用其生成的布局。",
   "metric": "用于每一代 Google TPU（官方）",
   "applications": [
    "Google TPU 设计",
    "官方称亦被外部芯片公司采用"
   ],
   "openSource": "开源（circuit_training）",
   "note": "该方法曾引发外部复现争议，Nature 于 2024 年为原论文发布附录（addendum）。",
   "uncertain": true,
   "parents": [],
   "children": [],
   "links": [
    {
     "label": "Nature 论文 (2021)",
     "url": "https://www.nature.com/articles/s41586-021-03544-w",
     "type": "paper"
    },
    {
     "label": "Nature 附录 (2024)",
     "url": "https://www.nature.com/articles/s41586-024-08032-5",
     "type": "paper"
    },
    {
     "label": "DeepMind 博客（2024 命名）",
     "url": "https://deepmind.google/blog/how-alphachip-transformed-computer-chip-design/",
     "type": "blog"
    },
    {
     "label": "GitHub: circuit_training",
     "url": "https://github.com/google-research/circuit_training",
     "type": "code"
    }
   ]
  },
  {
   "id": "alphaqubit",
   "name": "AlphaQubit",
   "date": "2024-11-20",
   "discipline": "quantum",
   "tags": [
    "quantum"
   ],
   "status": "superseded",
   "tagline": "神经网络量子纠错解码器",
   "problem": "量子比特极易出错，需要准确识别错误的解码器。",
   "result": "与 Google Quantum AI 合作，在 Sycamore 真实数据上训练；最大实验中比张量网络方法少 6% 错误、比相关匹配方法少 30% 错误。",
   "metric": "较张量网络方法错误减少 6%",
   "applications": [
    "容错量子计算"
   ],
   "openSource": "未开源（Nature 论文公开）",
   "parents": [],
   "children": [
    "alphaqubit-2"
   ],
   "links": [
    {
     "label": "Nature 论文 (2024)",
     "url": "https://www.nature.com/articles/s41586-024-08148-8",
     "type": "paper"
    },
    {
     "label": "Google 官方博客",
     "url": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/alphaqubit-quantum-error-correction/",
     "type": "blog"
    }
   ]
  },
  {
   "id": "alphaqubit-2",
   "name": "AlphaQubit 2",
   "date": "2025-12-08",
   "discipline": "quantum",
   "tags": [
    "quantum"
   ],
   "status": "research",
   "tagline": "可扩展、实时的神经量子纠错解码器",
   "problem": "解码器需同时做到准确、快速、可扩展，此前没有神经解码器满足全部要求。",
   "result": "在表面码与色码上接近最优逻辑错误率；在商用加速器上实现每周期低于 1 微秒的实时解码（表面码距离 11、色码距离 9 为首次实时解码）。",
   "metric": "< 1 μs/周期 实时解码",
   "applications": [
    "超导量子计算机的实时纠错"
   ],
   "openSource": "预印本公开",
   "note": "目前为 arXiv 预印本（2025 年 12 月），尚未见同行评审版本或官方博客；日期为约略值。",
   "uncertain": true,
   "parents": [
    "alphaqubit"
   ],
   "children": [],
   "links": [
    {
     "label": "arXiv 论文",
     "url": "https://arxiv.org/abs/2512.07737",
     "type": "paper"
    }
   ]
  },
  {
   "id": "alphaevolve",
   "name": "AlphaEvolve",
   "date": "2025-05-14",
   "discipline": "cs",
   "tags": [
    "cs",
    "math",
    "chips",
    "quantum",
    "genomics"
   ],
   "status": "active",
   "tagline": "Gemini 驱动的进化式编码智能体，发现与优化算法",
   "problem": "让大模型的创造力与自动评估器结合，系统性发现可验证的新算法。",
   "result": "Gemini 生成代码 + 自动评估 + 进化框架；其数据中心调度方案持续回收 Google 全球约 0.7% 算力，并改进矩阵乘法与多个开放数学问题。一年后（2026-05）已用于 DNA 测序纠错（DeepConsensus 变异检测错误降 30%）、Willow 量子处理器电路（误差降 10 倍）等；2026-07-09 在 Google Cloud 全面可用。",
   "metric": "回收约 0.7% 全球算力",
   "applications": [
    "数据中心调度、芯片设计、AI 训练",
    "基因组学、量子物理、电网",
    "Google Cloud 企业客户"
   ],
   "openSource": "商业服务（Google Cloud / Gemini Enterprise）",
   "parents": [
    "alphatensor",
    "alphadev"
   ],
   "children": [
    "alphaproof-nexus"
   ],
   "links": [
    {
     "label": "DeepMind 博客 (2025)",
     "url": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/",
     "type": "blog"
    },
    {
     "label": "arXiv 白皮书",
     "url": "https://arxiv.org/abs/2506.13131",
     "type": "paper"
    },
    {
     "label": "一周年影响（2026-05）",
     "url": "https://deepmind.google/blog/alphaevolve-impact/",
     "type": "blog"
    },
    {
     "label": "Google Cloud 全面可用（2026-07）",
     "url": "https://cloud.google.com/blog/products/ai-machine-learning/alphaevolve-is-available-for-everyone",
     "type": "official"
    }
   ]
  },
  {
   "id": "alphagenome",
   "name": "AlphaGenome",
   "date": "2025-06-25",
   "discipline": "genomics",
   "tags": [
    "genomics"
   ],
   "status": "active",
   "tagline": "读懂百万碱基 DNA 的调控“暗物质”",
   "problem": "98% 的非编码基因组如何调控基因，变异如何致病？",
   "result": "输入最长 100 万碱基对的 DNA 序列，预测数千种调控相关分子性质，并通过比较突变前后预测为变异打分；2026 年 1 月发表于 Nature。",
   "metric": "1 Mb 输入、碱基级分辨率",
   "applications": [
    "罕见病与复杂性状研究",
    "通过 API 供非商业研究使用"
   ],
   "openSource": "API 免费用于非商业研究；客户端开源",
   "parents": [],
   "children": [
    "alphagenome-atlas",
    "alphagenome-2"
   ],
   "links": [
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/",
     "type": "blog"
    },
    {
     "label": "论文 PDF",
     "url": "https://storage.googleapis.com/deepmind-media/papers/alphagenome.pdf",
     "type": "paper"
    },
    {
     "label": "AlphaGenome API 文档",
     "url": "https://www.alphagenomedocs.com/",
     "type": "official"
    },
    {
     "label": "GitHub: alphagenome",
     "url": "https://github.com/google-deepmind/alphagenome",
     "type": "code"
    }
   ]
  },
  {
   "id": "alphagenome-atlas",
   "name": "AlphaGenome Atlas",
   "date": "2026-09-08",
   "discipline": "genomics",
   "tags": [
    "genomics"
   ],
   "status": "active",
   "tagline": "人类基因组全部 90 亿种单碱基变异的预测图谱",
   "problem": "约 90 亿种可能的单字母突变不可能逐一实验检测。",
   "result": "预先计算 AlphaGenome 对全部 90 亿个单核苷酸变异的分子效应预测，并发布结合 AlphaGenome 与 AlphaMissense 的 AVI（AlphaGenome Variant Impact）评分，学术研究可免费网页使用。",
   "metric": "90 亿 SNV 全覆盖",
   "applications": [
    "变异优先级排序",
    "罕见病研究"
   ],
   "openSource": "免费门户（学术研究）",
   "parents": [
    "alphagenome",
    "alphamissense"
   ],
   "children": [],
   "links": [
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphagenome-atlas-a-predictive-map-of-every-possible-dna-letter-change-in-the-human-genome/",
     "type": "blog"
    },
    {
     "label": "The Verge 报道",
     "url": "https://www.theverge.com/ai-artificial-intelligence/991180/google-launches-alpha-genome-atlas",
     "type": "news"
    }
   ]
  },
  {
   "id": "alphagenome-2",
   "name": "AlphaGenome 2",
   "date": "2026-11-11",
   "discipline": "genomics",
   "tags": [
    "genomics"
   ],
   "status": "active",
   "tagline": "细胞类型级的变异效应预测",
   "problem": "同一 DNA 在心肌、神经、免疫细胞中读法不同，组织平均模型会漏掉疾病相关信号。",
   "result": "重构为条件模型：输入 DNA 序列与指定细胞类型/状态；基于 68 种组织超 1 亿个单细胞图谱训练，识别出改变基因表达的变异数量约为上一代最佳模型的两倍，可预测未见过的细胞状态。",
   "metric": "有效变异识别数约 2 倍",
   "applications": [
    "疾病细胞机制研究",
    "API 非商业可用，商业研究私测"
   ],
   "openSource": "API（非商业）",
   "parents": [
    "alphagenome"
   ],
   "children": [],
   "links": [
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/",
     "type": "blog"
    },
    {
     "label": "AlphaGenome API 文档",
     "url": "https://www.alphagenomedocs.com/",
     "type": "official"
    }
   ]
  },
  {
   "id": "alphaearth",
   "name": "AlphaEarth Foundations",
   "date": "2025-07-30",
   "discipline": "earth",
   "tags": [
    "earth"
   ],
   "status": "active",
   "tagline": "像“虚拟卫星”一样刻画整个地球陆地",
   "problem": "卫星、雷达、激光测高等地球观测数据多源异构，难以统一利用。",
   "result": "将多源数据融合为 10×10 米分辨率的紧凑嵌入，存储比其他测试系统少 16 倍、平均错误率低 24%；年度嵌入以 Satellite Embedding 数据集发布于 Google Earth Engine。2026 年 7 月推出 Custom Satellite Embeddings 企业私测。",
   "metric": "错误率平均低 24%，存储省 16 倍",
   "applications": [
    "粮食安全、毁林监测",
    "城市扩张、水资源",
    "森林碳估算"
   ],
   "openSource": "数据集在 Earth Engine 开放",
   "parents": [],
   "children": [],
   "links": [
    {
     "label": "DeepMind 博客",
     "url": "https://deepmind.google/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/",
     "type": "blog"
    },
    {
     "label": "arXiv 论文",
     "url": "https://arxiv.org/abs/2507.22291",
     "type": "paper"
    },
    {
     "label": "Custom Satellite Embeddings（2026）",
     "url": "https://mapsplatform.google.com/resources/blog/introducing-custom-satellite-embeddings-powered-by-alphaearth-foundations/",
     "type": "official"
    }
   ]
  }
 ],
 "frontiers": [
  {
   "title": "从“会下棋”到“会证明”",
   "en": "From games to proofs",
   "body": "AlphaZero 的搜索+自我对弈思想，被移植到 Lean 形式证明（AlphaProof），再到研究级开放问题（AlphaProof Nexus）与真实软件验证（AlphaProof Code）。核心是：可验证的环境让 AI 能从经验中可靠学习。",
   "ids": [
    "alphazero",
    "alphaproof",
    "alphaproof-nexus",
    "alphaproof-code"
   ]
  },
  {
   "title": "算法发现的自动化",
   "en": "Automating algorithm discovery",
   "body": "从 AlphaTensor（矩阵乘法）、AlphaDev（排序）到 AlphaEvolve（通用进化编码智能体），AI 正在为数学、量子电路、基因组学和数据中心直接产出可部署的新算法。",
   "ids": [
    "alphatensor",
    "alphadev",
    "alphaevolve"
   ]
  },
  {
   "title": "读写生命的语言",
   "en": "Reading and writing biology",
   "body": "“读”：AlphaFold 预测结构，AlphaMissense/AlphaGenome 解读变异；“写”：AlphaProteo 设计结合体，AlphaProtein Novo 从头设计酶。2026 年的 AlphaGenome 2 已下沉到单细胞层面。",
   "ids": [
    "alphafold-3",
    "alphagenome-2",
    "alphaproteo",
    "alphaprotein-novo"
   ]
  },
  {
   "title": "AI 设计 AI 的硬件",
   "en": "AI designing its own substrate",
   "body": "AlphaChip 为 TPU 布局，AlphaEvolve 优化芯片电路与训练内核——形成“AI 改进训练 AI 的硬件与软件”的反馈回路。",
   "ids": [
    "alphachip",
    "alphaevolve"
   ]
  },
  {
   "title": "通向容错量子计算",
   "en": "Toward fault-tolerant quantum",
   "body": "AlphaQubit 证明神经解码器的精度优势，AlphaQubit 2 进一步解决实时性与可扩展性；AlphaTensor-Quantum 与 AlphaEvolve 则用于优化量子电路。",
   "ids": [
    "alphaqubit",
    "alphaqubit-2",
    "alphatensor-quantum"
   ]
  },
  {
   "title": "行星尺度的感知",
   "en": "Planetary-scale perception",
   "body": "AlphaEarth Foundations 把地球观测数据压缩为统一嵌入，使毁林、农业与水资源监测的成本大幅降低。",
   "ids": [
    "alphaearth"
   ]
  }
 ],
 "changelog": [
  {
   "date": "2026-11-13",
   "text": "网站上线，收录 32 个项目/条目；同步 AlphaGenome 2（11-11）、AlphaProof Code（11-10）、AlphaFold 3 v3.0.5（10-30）等最新进展。"
  }
 ]
}