{"site":"https://rsi.zongxin.tech","description":"模型与RSI 证据库与判断台账。三个独立的轴：strength 证据强度（data 实测 / demo 演示 / argument 论证 / speculation 主张）、sourceKinds 来源类型（第三方 / 论文 / 官方口径 / 媒体）、verifiability 可核验性（reproduced 独立复现，人工确认并附 reproducedBy / evaluated 第三方评测 / registered 临床注册记录 / media 媒体实测 / checkable 可自查 / opaque 不可查）。artifact 是公开的制品（开源代码 / 开放权重），只影响 verifiability，不影响 strength。links 是可点击的原文。","evidence":[{"id":"rsi:anthropic-90-percent-merged","type":"rsi","title":"Anthropic：核心代码库九成以上的改动由 Claude 写","verdict":"年中报告里是「八成以上」，Fable 5.2 发布时升到「九成以上」（有人类指导）。官方口径，衡量的是代码量，不是研究产出。","href":"/rsi#anthropic-90-percent-merged","lines":["rsi"],"strength":"data","stance":"支持","tag":"RSI 阶段 3","sources":["官方口径（anthropic.com/claude-fable-and-mythos-5-2）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"anthropic.com/claude-fable-and-mythos-5-2","href":"https://anthropic.com/claude-fable-and-mythos-5-2"}],"addedIn":"2026-11-11","judgments":["thesis-coding"],"url":"https://rsi.zongxin.tech/evidence#rsi-anthropic-90-percent-merged","pageUrl":"https://rsi.zongxin.tech/rsi#anthropic-90-percent-merged"},{"id":"ai4s:pinal","type":"ai4s","title":"Pinal","verdict":"把蛋白设计的入口从「指定结构」换成「说出功能」；湿实验里有设计能用，但命中率低，对性质的精确控制仍有限。","href":"/ai4s/protein#pinal","lines":["ai4s"],"strength":"data","stance":"有争议","dispute":"荧光蛋白 29 条里只有 1 条发光，命中率低；作者也承认对蛋白性质的精确控制还有限。实验都由作者团队完成，尚无第三方复现。","level":2,"validation":"experiment","tag":"蛋白质与分子设计 · L2 · 作者实验","sources":["Nature DOI 10.1038/s41586-026-11205-z（原文需登录，本期未能直接打开；经 phys.org 转述）","经 phys.org 转述 Nature 论文（2026-11-03）","Hugging Face 模型卡 westlake-repl/Pinal"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"doi.org/10.1038/s41586-026-11205-z","href":"https://doi.org/10.1038/s41586-026-11205-z"},{"label":"phys.org/news/2026-11-ai-protein.html","href":"https://phys.org/news/2026-11-ai-protein.html"},{"label":"doi:10.1038/s41586-026-11205-z（原文需登录","href":"https://doi.org/10.1038/s41586-026-11205-z（原文需登录"},{"label":"huggingface.co/westlake-repl/Pinal","href":"https://huggingface.co/westlake-repl/Pinal"}],"addedIn":"2026-11-11","judgments":["ai4s-protein"],"url":"https://rsi.zongxin.tech/evidence#ai4s-pinal","pageUrl":"https://rsi.zongxin.tech/ai4s/protein#pinal"},{"id":"ai4s:proteintalks","type":"ai4s","title":"ProteinTalks","verdict":"虚拟细胞第一次以蛋白质组为底座进了《自然》；能预测药效与协同、给类器官排候选药，但摘要没说明预测有没有经过新的湿实验。","href":"/ai4s/cell#proteintalks","lines":["ai4s"],"strength":"data","stance":"有争议","dispute":"摘要里的优势限定在「所评估的协议下、相对所选基线」；预测是否经过新的湿实验验证，本期没能读到全文确认。","level":1,"validation":"benchmark","tag":"虚拟细胞与生物基础模型 · L1 · 基准评测","sources":["Nature DOI 10.1038/s41586-026-11001-9（摘要经 Europe PMC，PMID 42717098）","Nature 论文摘要（经 Europe PMC）"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"doi.org/10.1038/s41586-026-11001-9","href":"https://doi.org/10.1038/s41586-026-11001-9"},{"label":"doi:10.1038/s41586-026-11001-9（摘要经","href":"https://doi.org/10.1038/s41586-026-11001-9（摘要经"}],"addedIn":"2026-11-11","judgments":["ai4s-cell"],"url":"https://rsi.zongxin.tech/evidence#ai4s-proteintalks","pageUrl":"https://rsi.zongxin.tech/ai4s/cell#proteintalks"},{"id":"crossing:frontier-robot-policy","type":"crossing","title":"通用大模型直接当机器人策略","verdict":"通用大模型不经机器人专门训练，已能在模拟器里直接控制双臂机器人，第三方测出的成绩超过全部公开的机器人策略；但成绩几乎都在模拟器里，真机上它更像高层规划者，离稳定可靠的控制还远。","href":"/crossings#frontier-robot-policy","lines":["sota","wm"],"strength":"data","stance":"有争议","dispute":"成绩几乎都在模拟器里，Anthropic 的数字是自己跑的；第三方测出精细操作、动态控制与双臂协调都弱，真机上它多是配合专门控制器的高层规划者。","tag":"传导 · 大模型 →","sources":["arXiv:2609.24170","arXiv:2610.22471","官方口径（anthropic.com/claude-fable-and-mythos-5-2）","官方口径（developers.googleblog.com/gemini-4-argon-for-robotics）"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arXiv:2609.24170","href":"https://arxiv.org/abs/2609.24170"},{"label":"arXiv:2610.22471","href":"https://arxiv.org/abs/2610.22471"},{"label":"anthropic.com/claude-fable-and-mythos-5-2","href":"https://anthropic.com/claude-fable-and-mythos-5-2"},{"label":"developers.googleblog.com/gemini-4-argon-for-robotics","href":"https://developers.googleblog.com/gemini-4-argon-for-robotics"}],"addedIn":"2026-11-11","judgments":["thesis-physical-early"],"url":"https://rsi.zongxin.tech/evidence#crossing-frontier-robot-policy","pageUrl":"https://rsi.zongxin.tech/crossings#frontier-robot-policy"},{"id":"matrix:gemini-robotics-2-5:realworld","type":"matrix","title":"Gemini Robotics 2.5 · 真机迁移","verdict":"真机分任务成功率（如合上工具箱盖 0% → 93%、用夹子装糕点 5% → 100%）与奔驰产线试点 96%；均为厂商数字，没有试验次数，没有第三方复现。","href":"/models/gemini-robotics-2-5","lines":["wm"],"strength":"data","tag":"能力矩阵 · 真机迁移","sources":["官方口径（Google DeepMind 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence","href":"https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/"}],"addedIn":"2026-11-11","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-gemini-robotics-2-5-realworld","pageUrl":"https://rsi.zongxin.tech/models/gemini-robotics-2-5"},{"id":"matrix:groot-2:realworld","type":"matrix","title":"GR00T 2 · 真机迁移","verdict":"双臂 + 两只灵巧手的真机上，50 个没见过的任务给出零样本、1 个演示、5 个演示三档成功率；任务名称未逐一公布，尚无第三方复现。","href":"/models/groot-2","lines":["wm"],"strength":"data","tag":"能力矩阵 · 真机迁移","sources":["官方口径（NVIDIA GEAR 实验室）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"research.nvidia.com/labs/gear/gr00t-2","href":"https://research.nvidia.com/labs/gear/gr00t-2"}],"addedIn":"2026-11-11","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-groot-2-realworld","pageUrl":"https://rsi.zongxin.tech/models/groot-2"},{"id":"matrix:genie-4:controllable","type":"matrix","title":"Genie 4 · 动作可控性","verdict":"自然语言实时引导镜头、天气与角色，只有官方演示，没有量化可控性指标。","href":"/models/genie-4","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 动作可控性","sources":["官方口径（Google DeepMind 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"deepmind.google/blog/genie-4-a-new-generation-of-world-simulation","href":"https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/"}],"addedIn":"2026-11-11","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-genie-4-controllable","pageUrl":"https://rsi.zongxin.tech/models/genie-4"},{"id":"matrix:groot-2:planning","type":"matrix","title":"GR00T 2 · 规划与反事实","verdict":"同时预测未来视频与动作，先「想象」自己做这件事的画面再出动作；没有单独的规划指标。","href":"/models/groot-2","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 规划与反事实","sources":["官方口径（NVIDIA GEAR 实验室）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"research.nvidia.com/labs/gear/gr00t-2","href":"https://research.nvidia.com/labs/gear/gr00t-2"}],"addedIn":"2026-11-11","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-groot-2-planning","pageUrl":"https://rsi.zongxin.tech/models/groot-2"},{"id":"rsi:ace","type":"rsi","title":"ACE：让上下文自己演化","verdict":"不动权重，只让模型根据执行反馈不断改写自己的「操作手册」，智能体基准提升 10.6%。","href":"/rsi#ace","lines":["rsi"],"strength":"data","stance":"支持","tag":"RSI 阶段 1","sources":["arXiv:2510.04618"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arXiv:2510.04618","href":"https://arxiv.org/abs/2510.04618"}],"addedIn":"2026-09","judgments":["rsi-stage-1"],"url":"https://rsi.zongxin.tech/evidence#rsi-ace","pageUrl":"https://rsi.zongxin.tech/rsi#ace"},{"id":"rsi:voyager","type":"rsi","title":"Voyager：把技能存成代码库","verdict":"2023 年的早期样本：在 Minecraft 里把学会的技能写成代码存进技能库，后续任务直接调用。","href":"/rsi#voyager","lines":["rsi","wm"],"strength":"data","stance":"支持","tag":"RSI 阶段 1","sources":["arXiv:2305.16291"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arXiv:2305.16291","href":"https://arxiv.org/abs/2305.16291"}],"addedIn":"2026-09","judgments":["rsi-stage-1"],"url":"https://rsi.zongxin.tech/evidence#rsi-voyager","pageUrl":"https://rsi.zongxin.tech/rsi#voyager"},{"id":"rsi:darwin-godel-machine","type":"rsi","title":"达尔文-哥德尔机：智能体改写自己的代码","verdict":"智能体反复改写自己的脚手架代码并保留一个不断变大的变体档案库，SWE-bench 从 20% 升到 50%。","href":"/rsi#darwin-godel-machine","lines":["rsi"],"strength":"data","stance":"支持","tag":"RSI 阶段 2","sources":["arXiv:2505.22954"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arXiv:2505.22954","href":"https://arxiv.org/abs/2505.22954"}],"addedIn":"2026-09","judgments":["thesis-verifier","rsi-stage-2"],"url":"https://rsi.zongxin.tech/evidence#rsi-darwin-godel-machine","pageUrl":"https://rsi.zongxin.tech/rsi#darwin-godel-machine"},{"id":"rsi:meta-harness","type":"rsi","title":"Meta-Harness：端到端优化模型外面那一层","verdict":"自动搜索出的脚手架在 TerminalBench-2 上超过最好的人工设计，而且能迁移到没见过的模型上。","href":"/rsi#meta-harness","lines":["rsi"],"strength":"data","stance":"支持","tag":"RSI 阶段 2","sources":["arXiv:2603.28052"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arXiv:2603.28052","href":"https://arxiv.org/abs/2603.28052"}],"addedIn":"2026-09","judgments":["rsi-stage-2"],"url":"https://rsi.zongxin.tech/evidence#rsi-meta-harness","pageUrl":"https://rsi.zongxin.tech/rsi#meta-harness"},{"id":"rsi:alphaevolve","type":"rsi","title":"AlphaEvolve：加速了训练自己的那个模型","verdict":"Gemini 驱动的进化式编程智能体，找到的优化反过来加速了训练 Gemini 本身——阶段 3 最早的硬证据之一。","href":"/rsi#alphaevolve","lines":["rsi","ai4s"],"strength":"data","stance":"支持","tag":"RSI 阶段 3","sources":["arXiv:2506.13131"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arXiv:2506.13131","href":"https://arxiv.org/abs/2506.13131"}],"addedIn":"2026-09","judgments":["rsi-stage-3"],"url":"https://rsi.zongxin.tech/evidence#rsi-alphaevolve","pageUrl":"https://rsi.zongxin.tech/rsi#alphaevolve"},{"id":"rsi:openai-research-intern","type":"rsi","title":"OpenAI 宣布达成「自动化研究实习生」","verdict":"OpenAI 称研究团队每个人工工作日对应约 3.1 个智能体工作日；数字全部自测自评。","href":"/rsi#openai-research-intern","lines":["rsi"],"strength":"data","stance":"支持","tag":"RSI 阶段 3","sources":["官方口径（openai.com/index/research-acceleration-view-inside-openai）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"openai.com/index/research-acceleration-view-inside-openai","href":"https://openai.com/index/research-acceleration-view-inside-openai"}],"addedIn":"2026-09","judgments":["rsi-stage-3"],"url":"https://rsi.zongxin.tech/evidence#rsi-openai-research-intern","pageUrl":"https://rsi.zongxin.tech/rsi#openai-research-intern"},{"id":"rsi:anthropic-when-ai-builds-itself","type":"rsi","title":"Anthropic：《当 AI 开始构建自己》","verdict":"合入代码库的代码 80% 以上由 Claude 编写；代码优化任务上的提速从约 3 倍升到约 52 倍。均为官方口径。","href":"/rsi#anthropic-when-ai-builds-itself","lines":["rsi"],"strength":"data","stance":"支持","tag":"RSI 阶段 3","sources":["官方口径（anthropic.com/institute/recursive-self-improvement）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"anthropic.com/institute/recursive-self-improvement","href":"https://anthropic.com/institute/recursive-self-improvement"}],"addedIn":"2026-09","judgments":["thesis-coding","rsi-stage-3"],"url":"https://rsi.zongxin.tech/evidence#rsi-anthropic-when-ai-builds-itself","pageUrl":"https://rsi.zongxin.tech/rsi#anthropic-when-ai-builds-itself"},{"id":"rsi:ai-scientist-nature","type":"rsi","title":"AI Scientist 登上《自然》：全自动写出的论文通过同行评审","verdict":"Sakana 的系统从想法到写论文全自动完成，一篇通过了顶会工作坊首轮评审——但只是工作坊，接收率约 70%。","href":"/rsi#ai-scientist-nature","lines":["rsi","ai4s"],"strength":"data","stance":"支持","tag":"RSI 阶段 3","sources":["Nature s41586-026-10265-5"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"s41586-026-10265-5","href":"https://www.nature.com/articles/s41586-026-10265-5"}],"addedIn":"2026-09","judgments":["rsi-stage-3","ai4s-scientist"],"url":"https://rsi.zongxin.tech/evidence#rsi-ai-scientist-nature","pageUrl":"https://rsi.zongxin.tech/rsi#ai-scientist-nature"},{"id":"rsi:princeton-open-ended-research","type":"rsi","title":"反方：智能体在开放式研究上「明确不行」","verdict":"普林斯顿团队让智能体复做两篇未发表的 NeurIPS 2026 论文：工程执行很强，研究判断很弱。","href":"/rsi#princeton-open-ended-research","lines":["rsi"],"strength":"data","stance":"反方","tag":"RSI 阶段 3","sources":["technologyreview.com/2026/08/18/1142188/ai-recursive-self-improvement/"],"sourceKinds":["媒体"],"officialOnly":false,"verifiability":"opaque","links":[{"label":"technologyreview.com/2026/08/18/1142188/ai-recursive-self-improvement","href":"https://technologyreview.com/2026/08/18/1142188/ai-recursive-self-improvement/"}],"addedIn":"2026-09","judgments":["thesis-ai4s-two-years","rsi-stage-3","ai4s-scientist"],"url":"https://rsi.zongxin.tech/evidence#rsi-princeton-open-ended-research","pageUrl":"https://rsi.zongxin.tech/rsi#princeton-open-ended-research"},{"id":"ai4s:alphafold-2","type":"ai4s","title":"AlphaFold 2","verdict":"AI4S 的起点事件：CASP14 上的精度接近实验解析水平，随后开放的结构数据库成了生物学的公共基础设施。","href":"/ai4s/protein#alphafold-2","lines":["ai4s"],"strength":"data","level":1,"validation":"independent","tag":"蛋白质与分子设计 · L1 · 独立检验","sources":["Nature 596, 583–589 (2021)","nobelprize.org"],"sourceKinds":["第三方","论文"],"officialOnly":false,"verifiability":"reproduced","artifact":"开源代码","links":[{"label":"nature.com/articles/s41586-021-03819-2","href":"https://www.nature.com/articles/s41586-021-03819-2"},{"label":"nobelprize.org/prizes/chemistry/2024/press-release","href":"https://www.nobelprize.org/prizes/chemistry/2024/press-release/"},{"label":"doi.org/10.1038/s41592-024-02272-z","href":"https://doi.org/10.1038/s41592-024-02272-z"}],"reproducedBy":{"value":"OpenFold 团队用自己的代码从头重训 AlphaFold 2，精度与原版相当，并公开了训练代码与权重","source":"Nature Methods 21, 1514–1524 (2024)","href":"https://doi.org/10.1038/s41592-024-02272-z"},"addedIn":"2026-09","judgments":["thesis-verifier"],"url":"https://rsi.zongxin.tech/evidence#ai4s-alphafold-2","pageUrl":"https://rsi.zongxin.tech/ai4s/protein#alphafold-2"},{"id":"ai4s:alphafold-3","type":"ai4s","title":"AlphaFold 3","verdict":"预测这一层的标杆，已是药物研发的日常工具；但它回答的是「长什么样」，不回答「怎么设计」。","href":"/ai4s/protein#alphafold-3","lines":["ai4s"],"strength":"data","level":1,"validation":"benchmark","tag":"蛋白质与分子设计 · L1 · 基准评测","sources":["官方口径；Nature 630, 493–500 (2024)","官方口径"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"nature.com/articles/s41586-024-07487-w","href":"https://www.nature.com/articles/s41586-024-07487-w"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#ai4s-alphafold-3","pageUrl":"https://rsi.zongxin.tech/ai4s/protein#alphafold-3"},{"id":"ai4s:rfdiffusion","type":"ai4s","title":"RFdiffusion","verdict":"让蛋白质设计从少数专家的手艺变成可规模化的流程，设计结果经实验解析的结构与计算设计高度吻合。","href":"/ai4s/protein#rfdiffusion","lines":["ai4s"],"strength":"data","level":2,"validation":"experiment","tag":"蛋白质与分子设计 · L2 · 作者实验","sources":["Nature 620, 1089–1100 (2023)","nobelprize.org"],"sourceKinds":["第三方","论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开源代码","links":[{"label":"nature.com/articles/s41586-023-06415-8","href":"https://www.nature.com/articles/s41586-023-06415-8"},{"label":"nobelprize.org/prizes/chemistry/2024/press-release","href":"https://www.nobelprize.org/prizes/chemistry/2024/press-release/"}],"addedIn":"2026-09","judgments":["ai4s-protein"],"url":"https://rsi.zongxin.tech/evidence#ai4s-rfdiffusion","pageUrl":"https://rsi.zongxin.tech/ai4s/protein#rfdiffusion"},{"id":"ai4s:esm3","type":"ai4s","title":"ESM3","verdict":"设计出一个与已知荧光蛋白差异很大、但真的会发光的蛋白——语言模型路线做「设计」的代表作。","href":"/ai4s/protein#esm3","lines":["ai4s"],"strength":"data","level":2,"validation":"experiment","tag":"蛋白质与分子设计 · L2 · 作者实验","sources":["Science 387, 850–858 (2025)；「5 亿年」为官方口径","官方口径"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"science.org/doi/10.1126/science.ads0018","href":"https://www.science.org/doi/10.1126/science.ads0018"},{"label":"doi.org/10.1126/science.ads0018","href":"https://doi.org/10.1126/science.ads0018"}],"addedIn":"2026-09","judgments":["ai4s-protein"],"url":"https://rsi.zongxin.tech/evidence#ai4s-esm3","pageUrl":"https://rsi.zongxin.tech/ai4s/protein#esm3"},{"id":"ai4s:alphaproteo","type":"ai4s","title":"AlphaProteo","verdict":"在七个靶点上湿实验验证过，并首次为 VEGF-A 设计出可用的结合剂；但模型未开放，外部无法复现。","href":"/ai4s/protein#alphaproteo","lines":["ai4s"],"strength":"data","level":2,"validation":"experiment","tag":"蛋白质与分子设计 · L2 · 作者实验","sources":["arXiv:2409.08022"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arxiv.org/abs/2409.08022","href":"https://arxiv.org/abs/2409.08022"}],"addedIn":"2026-09","judgments":["ai4s-protein"],"url":"https://rsi.zongxin.tech/evidence#ai4s-alphaproteo","pageUrl":"https://rsi.zongxin.tech/ai4s/protein#alphaproteo"},{"id":"ai4s:rentosertib","type":"ai4s","title":"Rentosertib（ISM001-055）","verdict":"第一个进入三期临床的「靶点和分子都由生成式 AI 发现」的药物。二期数据积极但样本小，三期结果要等到 2027 年以后。","href":"/ai4s/drug#rentosertib","lines":["ai4s"],"strength":"data","stance":"有争议","dispute":"二期 a 规模小（71 人）、周期短（12 周），主要终点是安全性而非疗效，FVC 改善只是探索性结果。「AI 发现」的说法主要依据公司自述，外界无法独立检验 AI 在其中的具体贡献。","level":2,"validation":"clinical","tag":"药物发现 · L2 · 临床试验","sources":["Nature Medicine (2025), s41591-025-03743-2","官方口径（insilico.com 新闻稿）；clinicaltrials.gov NCT07687459"],"sourceKinds":["第三方","论文","官方口径"],"officialOnly":false,"verifiability":"registered","links":[{"label":"nature.com/articles/s41591-025-03743-2","href":"https://www.nature.com/articles/s41591-025-03743-2"},{"label":"NCT07687459","href":"https://clinicaltrials.gov/study/NCT07687459"},{"label":"insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3","href":"https://insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3"}],"addedIn":"2026-09","judgments":["thesis-ai4s-two-years","ai4s-drug"],"url":"https://rsi.zongxin.tech/evidence#ai4s-rentosertib","pageUrl":"https://rsi.zongxin.tech/ai4s/drug#rentosertib"},{"id":"ai4s:boltz-2","type":"ai4s","title":"Boltz-2","verdict":"把药物筛选里最贵的一步（算亲和力）提速上千倍，而且权重与训练代码全开放、可商用——开源阵营的 AlphaFold 3 替代品。","href":"/ai4s/drug#boltz-2","lines":["ai4s"],"strength":"data","level":1,"validation":"benchmark","tag":"药物发现 · L1 · 基准评测","sources":["bioRxiv 10.1101/2025.06.14.659707；「1000 倍」为官方口径","官方口径"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","artifact":"开源代码","links":[{"label":"biorxiv.org/content/10.1101/2025.06.14.659707v1","href":"https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1"},{"label":"doi:10.1101/2025.06.14.659707","href":"https://doi.org/10.1101/2025.06.14.659707"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#ai4s-boltz-2","pageUrl":"https://rsi.zongxin.tech/ai4s/drug#boltz-2"},{"id":"ai4s:evo-2","type":"ai4s","title":"Evo 2","verdict":"生物序列的「大模型」标杆，完全开放；能预测突变影响、生成基因组片段，但生成的长序列是否真能在细胞里工作仍需逐一验证。","href":"/ai4s/cell#evo-2","lines":["ai4s"],"strength":"data","level":2,"validation":"benchmark","tag":"虚拟细胞与生物基础模型 · L2 · 基准评测","sources":["Nature (2026), s41586-026-10176-5","Arc 研究所官方"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","artifact":"开源代码","links":[{"label":"nature.com/articles/s41586-026-10176-5","href":"https://www.nature.com/articles/s41586-026-10176-5"}],"addedIn":"2026-09","judgments":["ai4s-cell"],"url":"https://rsi.zongxin.tech/evidence#ai4s-evo-2","pageUrl":"https://rsi.zongxin.tech/ai4s/cell#evo-2"},{"id":"ai4s:xtrimo-sc-perturb","type":"ai4s","title":"xTrimoSCPerturb","verdict":"拿下 Arc 首届虚拟细胞挑战赛第一名的中国团队模型；证明了单细胞基础模型的可用性，但比赛本身只覆盖有限的细胞类型。","href":"/ai4s/cell#xtrimo-sc-perturb","lines":["ai4s"],"strength":"data","stance":"有争议","dispute":"Arc 2026 年的第二届挑战赛把题目改成「在模型没见过的细胞环境中预测扰动响应」——这恰恰说明首届比赛的设定还不足以回答泛化问题。","level":1,"validation":"independent","tag":"虚拟细胞与生物基础模型 · L1 · 独立检验","sources":["Arc 研究所官方（arcinstitute.org/news/virtual-cell-challenge-2025-wrap-up）"],"sourceKinds":["第三方","官方口径"],"officialOnly":false,"verifiability":"evaluated","links":[{"label":"arcinstitute.org/news/virtual-cell-challenge-2025-wrap-up","href":"https://arcinstitute.org/news/virtual-cell-challenge-2025-wrap-up"}],"addedIn":"2026-09","judgments":["ai4s-cell"],"url":"https://rsi.zongxin.tech/evidence#ai4s-xtrimo-sc-perturb","pageUrl":"https://rsi.zongxin.tech/ai4s/cell#xtrimo-sc-perturb"},{"id":"ai4s:gnome","type":"ai4s","title":"GNoME","verdict":"把「预测稳定」这一层做到了极致；但「稳定」不等于「做得出来」也不等于「有用」，这之间的距离是材料 AI 的核心问题。","href":"/ai4s/materials#gnome","lines":["ai4s"],"strength":"data","stance":"有争议","dispute":"有材料学者指出，大量预测结构只是已知结构的元素替换，化学上的新颖性有限；「新材料」这个说法在材料学界有争议。","level":1,"validation":"benchmark","tag":"材料研发 · L1 · 基准评测","sources":["Nature 624, 80–85 (2023)"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"nature.com/articles/s41586-023-06735-9","href":"https://www.nature.com/articles/s41586-023-06735-9"}],"addedIn":"2026-09","judgments":["ai4s-materials"],"url":"https://rsi.zongxin.tech/evidence#ai4s-gnome","pageUrl":"https://rsi.zongxin.tech/ai4s/materials#gnome"},{"id":"ai4s:mattergen","type":"ai4s","title":"MatterGen","verdict":"从「筛选」跨到「按需生成」，并有一例实验合成；但只有一例，离常规流程还远。","href":"/ai4s/materials#mattergen","lines":["ai4s"],"strength":"data","level":2,"validation":"experiment","tag":"材料研发 · L2 · 作者实验","sources":["Nature (2025), s41586-025-08628-5"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开源代码","links":[{"label":"nature.com/articles/s41586-025-08628-5","href":"https://www.nature.com/articles/s41586-025-08628-5"}],"addedIn":"2026-09","judgments":["ai4s-materials"],"url":"https://rsi.zongxin.tech/evidence#ai4s-mattergen","pageUrl":"https://rsi.zongxin.tech/ai4s/materials#mattergen"},{"id":"ai4s:mattersim","type":"ai4s","title":"MatterSim","verdict":"把材料的第一级验证器（DFT 计算）变便宜：同一个模型跨元素、跨温度压力可用，是生成模型之后的筛选工具。","href":"/ai4s/materials#mattersim","lines":["ai4s"],"strength":"data","level":1,"validation":"benchmark","tag":"材料研发 · L1 · 基准评测","sources":["arXiv:2405.04967","github.com/microsoft/mattersim"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","artifact":"开源代码","links":[{"label":"arxiv.org/abs/2405.04967","href":"https://arxiv.org/abs/2405.04967"},{"label":"github.com/microsoft/mattersim","href":"https://github.com/microsoft/mattersim"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#ai4s-mattersim","pageUrl":"https://rsi.zongxin.tech/ai4s/materials#mattersim"},{"id":"ai4s:uma","type":"ai4s","title":"UMA","verdict":"原子模拟也开始讲规模定律：用约 5 亿个三维结构训练，跨化学领域通用；但仍是预测层，离设计与闭环还远。","href":"/ai4s/materials#uma","lines":["ai4s"],"strength":"data","level":1,"validation":"benchmark","tag":"材料研发 · L1 · 基准评测","sources":["arXiv:2506.23971"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2506.23971","href":"https://arxiv.org/abs/2506.23971"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#ai4s-uma","pageUrl":"https://rsi.zongxin.tech/ai4s/materials#uma"},{"id":"ai4s:a-lab","type":"ai4s","title":"A-Lab","verdict":"自动实验室的标志性案例，也是最好的反面教材：「新材料」的说法被同行质疑后，论文在 2026 年 2 月作了更正，连标题里的「新材料」都改成了「无机材料」。","href":"/ai4s/materials#a-lab","lines":["ai4s","rsi"],"strength":"data","stance":"有争议","dispute":"伦敦大学学院的 Robert Palgrave 与普林斯顿的 Leslie Schoop 团队分析后认为，这批化合物中没有真正意义上的新材料，自动化的 X 射线衍射解析也存在误判。这是「自动实验室的结论也需要验证器」的现成例子。","level":3,"validation":"experiment","tag":"材料研发 · L3 · 作者实验","sources":["Nature 624, 86–91 (2023)","Nature 650 (2026-02-05) 作者更正 s41586-025-09992-y","Nature 作者更正 s41586-025-09992-y"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"nature.com/articles/s41586-023-06734-w","href":"https://www.nature.com/articles/s41586-023-06734-w"},{"label":"s41586-025-09992-y","href":"https://www.nature.com/articles/s41586-025-09992-y"}],"addedIn":"2026-09","judgments":["thesis-ai4s-two-years","ai4s-materials"],"url":"https://rsi.zongxin.tech/evidence#ai4s-a-lab","pageUrl":"https://rsi.zongxin.tech/ai4s/materials#a-lab"},{"id":"ai4s:robin","type":"ai4s","title":"Robin","verdict":"第一次把「实验室在环」的完整科研循环交给 AI，找到一个老药的新用途；但实验仍由人来做，候选药离临床还很远。","href":"/ai4s/scientist#robin","lines":["ai4s","sota","rsi"],"strength":"data","stance":"有争议","dispute":"ripasudil 本身是已上市药物，这是「重定位」而不是全新分子；细胞实验到临床疗效之间还有很长的距离。","level":3,"validation":"experiment","tag":"AI 科学家 · L3 · 作者实验","sources":["Nature 655 (2026-05-19), s41586-026-10652-y；arXiv:2505.13400","官方口径（Incyte 新闻稿）"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arxiv.org/abs/2505.13400","href":"https://arxiv.org/abs/2505.13400"},{"label":"s41586-026-10652-y","href":"https://www.nature.com/articles/s41586-026-10652-y"},{"label":"futurehouse.org/research-announcements/demonstrating-end-to-end-scientific-discovery-with-robin-a-multi-agent-system","href":"https://www.futurehouse.org/research-announcements/demonstrating-end-to-end-scientific-discovery-with-robin-a-multi-agent-system"}],"addedIn":"2026-09","judgments":["ai4s-scientist"],"url":"https://rsi.zongxin.tech/evidence#ai4s-robin","pageUrl":"https://rsi.zongxin.tech/ai4s/scientist#robin"},{"id":"ai4s:co-scientist","type":"ai4s","title":"Co-Scientist（Gemini）","verdict":"通用大模型直接给出实验配方并在真实实验室一次成功——SOTA 牵引 AI4S 最直接的一例；但仍是半自动，结构确认也还不完整。","href":"/ai4s/scientist#co-scientist","lines":["ai4s","sota"],"strength":"data","stance":"有争议","dispute":"论文自己承认，MXene 类材料的合成结果「需要进一步实验确认原子结构」；这是 Google 自家团队的论文，尚无第三方复现。","level":3,"validation":"experiment","tag":"AI 科学家 · L3 · 作者实验","sources":["arXiv:2608.26701","arXiv:2608.26701","arXiv:2608.26701"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arxiv.org/abs/2608.26701","href":"https://arxiv.org/abs/2608.26701"}],"addedIn":"2026-09","judgments":["thesis-ai4s-two-years","ai4s-scientist"],"url":"https://rsi.zongxin.tech/evidence#ai4s-co-scientist","pageUrl":"https://rsi.zongxin.tech/ai4s/scientist#co-scientist"},{"id":"crossing:harness-gap","type":"crossing","title":"同一个基准，两个都算数的分数","verdict":"同一个模型、同一套题，换一层评测外壳分数从 62.7% 跳到 99.9%——世界模型的官方数字经不起同样这一问。","href":"/crossings#harness-gap","lines":["sota","wm","rsi"],"strength":"data","stance":"有争议","dispute":"有人认为 62.7% 才是模型的真实能力、99.9% 是注水；ARC Prize 自己不这么表述。争的是「要衡量的到底是模型，还是模型加上它的那层壳」，这个问题对世界模型同样没有答案。","tag":"传导 · 大模型 →","sources":["ARC Prize 官方博客 arcprize.org/blog/astra","官方口径（openai.com/index/gpt-6-astra/）","官方口径（World Labs 博客 2026-09-01）"],"sourceKinds":["第三方","官方口径"],"officialOnly":false,"verifiability":"evaluated","links":[{"label":"arcprize.org/blog/astra","href":"https://arcprize.org/blog/astra"},{"label":"openai.com/index/gpt-6-astra","href":"https://openai.com/index/gpt-6-astra/"}],"addedIn":"2026-09","judgments":["thesis-coding","rsi-stage-2"],"url":"https://rsi.zongxin.tech/evidence#crossing-harness-gap","pageUrl":"https://rsi.zongxin.tech/crossings#harness-gap"},{"id":"matrix:atlas:controllable","type":"matrix","title":"Atlas · 动作可控性","verdict":"相机位姿是原生输入格式，人类偏好评测在五个对手上给出 75%–94% 的胜率。需要注意口径：对手模型不接受相机作为原生输入，只能用文本描述相机运动，World Labs 自己也承认更精细的提示工程可能改善对手表现。","href":"/models/atlas","lines":["wm"],"strength":"data","stance":"有争议","dispute":"胜率是厂商自己组织的人类偏好评测，对手模型在不利口径下参评（不接受相机原生输入），World Labs 也承认更精细的提示工程可能改善对手表现；没有第三方用同一口径重测过。","tag":"能力矩阵 · 动作可控性","sources":["官方口径（World Labs 博客，人类偏好评测）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"worldlabs.ai/blog/atlas","href":"https://www.worldlabs.ai/blog/atlas"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-atlas-controllable","pageUrl":"https://rsi.zongxin.tech/models/atlas"},{"id":"crossing:science-workbench","type":"crossing","title":"通用大模型进入科研工作台","verdict":"三家都在 2026 年把通用模型推进了科研流程；目前加速的是文献、分析、写作与实验设计，验证环节——湿实验与临床——一步也没有变快。","href":"/crossings#science-workbench","lines":["sota","ai4s"],"strength":"demo","stance":"有争议","dispute":"量化数字大多是用户自述而非对照实验，Co-Scientist 的结果是 Google 自家论文、尚无第三方复现；药企试点多落地少是老问题，诺和诺德的合作公告里没有可事后核对的目标。","tag":"传导 · 大模型 →","sources":["官方口径（anthropic.com/news/claude-science-ai-workbench）","官方口径（Novo Nordisk 新闻稿，BioSpace 转载）","官方口径（openai.com/index/gpt-6-astra/）","arXiv:2608.26701"],"sourceKinds":["论文","官方口径","媒体"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"anthropic.com/news/claude-science-ai-workbench","href":"https://anthropic.com/news/claude-science-ai-workbench"},{"label":"openai.com/index/gpt-6-astra","href":"https://openai.com/index/gpt-6-astra/"},{"label":"arXiv:2608.26701","href":"https://arxiv.org/abs/2608.26701"}],"addedIn":"2026-09","judgments":["thesis-ai4s-two-years"],"url":"https://rsi.zongxin.tech/evidence#crossing-science-workbench","pageUrl":"https://rsi.zongxin.tech/crossings#science-workbench"},{"id":"matrix:cosmos-3:physics","type":"matrix","title":"Cosmos 3 · 物理合理性","verdict":"官方称在 Physics-IQ、R-Bench、PAI-Bench 这几个物理类榜单上居首，但发布博客没有给出任何具体分数，也没有第三方复现——这是「有主张、无可核对数字」的典型。","href":"/models/cosmos-3","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 物理合理性","sources":["官方口径（NVIDIA 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"blogs.nvidia.com/blog/cosmos-3-physical-ai","href":"https://blogs.nvidia.com/blog/cosmos-3-physical-ai/"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-cosmos-3-physics","pageUrl":"https://rsi.zongxin.tech/models/cosmos-3"},{"id":"matrix:atlas:temporal","type":"matrix","title":"Atlas · 时序一致性","verdict":"官方展示了 1 分钟、1440p 的连贯长视频，场景在整条相机路径上保持一致，但没有给出任何量化的长程一致性指标。","href":"/models/atlas","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 时序一致性","sources":["官方口径（World Labs 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"worldlabs.ai/blog/atlas","href":"https://www.worldlabs.ai/blog/atlas"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-atlas-temporal","pageUrl":"https://rsi.zongxin.tech/models/atlas"},{"id":"matrix:atlas:memory","type":"matrix","title":"Atlas · 长程记忆","verdict":"官方称同一组输入图像下无论换多少次相机路径，场景都保持一致；这是空间一致性的演示，没有量化的场景保真度数字。","href":"/models/atlas","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 长程记忆","sources":["官方口径（World Labs 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"worldlabs.ai/blog/atlas","href":"https://www.worldlabs.ai/blog/atlas"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-atlas-memory","pageUrl":"https://rsi.zongxin.tech/models/atlas"},{"id":"matrix:atlas:realworld","type":"matrix","title":"Atlas · 真机迁移","verdict":"官方演示用手机拍摄的 24 帧视频重建大场景，再模拟机器人沿不同路径行走并生成其机身相机会看到的 RGB 与深度；但没有公开任何真机策略迁移的成功率数字。","href":"/models/atlas","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 真机迁移","sources":["官方口径（World Labs 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"worldlabs.ai/blog/atlas","href":"https://www.worldlabs.ai/blog/atlas"}],"addedIn":"2026-09","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-atlas-realworld","pageUrl":"https://rsi.zongxin.tech/models/atlas"},{"id":"rsi:weng-harness","type":"rsi","title":"翁荔：近期的自我改进走的是 Harness，而不是改权重","verdict":"一篇系统梳理：从提示词优化到「优化器代码自己演化」的五级阶梯，主张近期 RSI 主要经由 Harness 发生。","href":"/rsi#weng-harness","lines":["rsi"],"strength":"argument","stance":"支持","tag":"RSI 阶段 2","sources":["lilianweng.github.io/posts/2026-07-04-harness/"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"lilianweng.github.io/posts/2026-07-04-harness","href":"https://lilianweng.github.io/posts/2026-07-04-harness/"}],"addedIn":"2026-09","judgments":["rsi-stage-2"],"url":"https://rsi.zongxin.tech/evidence#rsi-weng-harness","pageUrl":"https://rsi.zongxin.tech/rsi#weng-harness"},{"id":"crossing:persistent-context","type":"crossing","title":"从压缩摘要到可检索的持久上下文","verdict":"对照组这个月在补「长程记忆」，但补的是自己会话历史的记忆，不是环境状态的记忆——两件事只共享一个名字。","href":"/crossings#persistent-context","lines":["sota","wm","rsi"],"strength":"argument","stance":"有争议","dispute":"Codex 记的是自己写过的会话文本，世界模型要记的是连续高维、从没被显式写下来的环境状态；检索一段自己写过的笔记与重建一个自己没写过的世界，是两个问题。真正在传导的可能只是「为长上下文做的系统工程」。","tag":"传导 · 大模型 →","sources":["官方口径（openai.com/index/gpt-6-astra/）","ARC Prize 官方博客 arcprize.org/blog/astra","官方口径（World Labs 博客 2026-09-01）"],"sourceKinds":["第三方","官方口径"],"officialOnly":false,"verifiability":"evaluated","links":[{"label":"openai.com/index/gpt-6-astra","href":"https://openai.com/index/gpt-6-astra/"},{"label":"arcprize.org/blog/astra","href":"https://arcprize.org/blog/astra"}],"addedIn":"2026-09","judgments":["rsi-stage-1"],"url":"https://rsi.zongxin.tech/evidence#crossing-persistent-context","pageUrl":"https://rsi.zongxin.tech/crossings#persistent-context"},{"id":"rsi:stage4-stated-goals","type":"rsi","title":"阶段 4 目前只有时间表","verdict":"OpenAI 把「自动化 AI 研究员」定在 2028 年 3 月；Anthropic 的时间线上这一格写的是「20XX?」。","href":"/rsi#stage4-stated-goals","lines":["rsi"],"strength":"speculation","stance":"支持","tag":"RSI 阶段 4","sources":["官方口径（openai.com/index/research-acceleration-view-inside-openai）","官方口径（anthropic.com/institute/recursive-self-improvement）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"openai.com/index/research-acceleration-view-inside-openai","href":"https://openai.com/index/research-acceleration-view-inside-openai"},{"label":"anthropic.com/institute/recursive-self-improvement","href":"https://anthropic.com/institute/recursive-self-improvement"}],"addedIn":"2026-09","judgments":["rsi-stage-4"],"url":"https://rsi.zongxin.tech/evidence#rsi-stage4-stated-goals","pageUrl":"https://rsi.zongxin.tech/rsi#stage4-stated-goals"},{"id":"ai4s:isomorphic-pipeline","type":"ai4s","title":"Isomorphic Labs 管线","verdict":"AI4S 声量最大的公司之一，但截至本期还没有药物进入临床，首个临床时间点已推迟一次。","href":"/ai4s/drug#isomorphic-pipeline","lines":["ai4s"],"strength":"speculation","level":2,"validation":"none","tag":"药物发现 · L2 · 未公开","sources":["路透社报道（2026-01-20，雅虎财经转载）","官方口径"],"sourceKinds":["官方口径","媒体"],"officialOnly":false,"verifiability":"opaque","links":[{"label":"finance.yahoo.com/news/google-backed-ai-drug-discovery-195423147.html","href":"https://finance.yahoo.com/news/google-backed-ai-drug-discovery-195423147.html"},{"label":"isomorphiclabs.com","href":"https://www.isomorphiclabs.com/"}],"addedIn":"2026-09","judgments":["thesis-ai4s-two-years","ai4s-drug"],"url":"https://rsi.zongxin.tech/evidence#ai4s-isomorphic-pipeline","pageUrl":"https://rsi.zongxin.tech/ai4s/drug#isomorphic-pipeline"},{"id":"ai4s:periodic-labs","type":"ai4s","title":"Periodic Labs 自动实验室","verdict":"资本投入最重的「AI 科学家 + 自动实验室」公司；截至本期没有公开发表的材料发现，只能标官方口径。","href":"/ai4s/materials#periodic-labs","lines":["ai4s","sota"],"strength":"speculation","level":3,"validation":"none","tag":"材料研发 · L3 · 未公开","sources":["官方口径","官方口径（a16z 投资公告）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[{"label":"a16z.com/announcement/investing-in-periodic-labs","href":"https://a16z.com/announcement/investing-in-periodic-labs/"}],"addedIn":"2026-09","judgments":["ai4s-materials"],"url":"https://rsi.zongxin.tech/evidence#ai4s-periodic-labs","pageUrl":"https://rsi.zongxin.tech/ai4s/materials#periodic-labs"},{"id":"crossing:rsi","type":"crossing","title":"递归自我改进","verdict":"方向和直觉相反：不是 RSI 顺带产生世界模型，而是世界模型是 RSI 走出数字域的前置条件——因为物理世界没有便宜的验证器。","href":"/crossings#rsi","lines":["rsi","wm"],"strength":"data","stance":"有争议","dispute":"杨立昆一派认为，数字域的自我改进靠的是可验证的奖励信号，本质是搜索；物理域的困难在表征，因此 RSI 在代码上的成功不构成迁移证据。另外 World Action Verifier 验证的是「自我改进能让世界模型变好」，反方向「世界模型让 RSI 进入物理域」仍只有论证。","tag":"传导 · RSI →","sources":["arXiv:2607.07663","world-action-verifier.github.io（作者项目页）","官方口径（iclr.cc/virtual/2026/workshop/10000796）"],"sourceKinds":["论文","官方口径"],"officialOnly":false,"verifiability":"checkable","links":[{"label":"arXiv:2607.07663","href":"https://arxiv.org/abs/2607.07663"},{"label":"world-action-verifier.github.io","href":"https://world-action-verifier.github.io/"},{"label":"iclr.cc/virtual/2026/workshop/10000796","href":"https://iclr.cc/virtual/2026/workshop/10000796"}],"addedIn":"2026-08","judgments":["thesis-verifier"],"url":"https://rsi.zongxin.tech/evidence#crossing-rsi","pageUrl":"https://rsi.zongxin.tech/crossings#rsi"},{"id":"matrix:world-models-2018:temporal","type":"matrix","title":"World Models · 时序一致性","verdict":"控制器 C 完全在 M 生成的想象序列里训练、全程未接触真实环境，转移到真实环境后 VizDoom 平均存活约 1100 步，说明长程想象没有从一开始就崩坏。","href":"/models/world-models-2018","lines":["wm"],"strength":"data","tag":"能力矩阵 · 时序一致性","sources":["arXiv:1803.10122"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/1803.10122","href":"https://arxiv.org/abs/1803.10122"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-world-models-2018-temporal","pageUrl":"https://rsi.zongxin.tech/models/world-models-2018"},{"id":"matrix:world-models-2018:planning","type":"matrix","title":"World Models · 规划与反事实","verdict":"策略完全在 M 生成的想象环境中训练、不接触真实环境，CarRacing 均分 906±21、VizDoom 平均存活约 1100 步，证明想象训练可迁移为真实决策。","href":"/models/world-models-2018","lines":["wm"],"strength":"data","tag":"能力矩阵 · 规划与反事实","sources":["arXiv:1803.10122"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/1803.10122","href":"https://arxiv.org/abs/1803.10122"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-world-models-2018-planning","pageUrl":"https://rsi.zongxin.tech/models/world-models-2018"},{"id":"matrix:dreamer-v3:temporal","type":"matrix","title":"DreamerV3 · 时序一致性","verdict":"单套超参数在 150+ 个多样任务上稳定训练，其中 Minecraft 采钻石任务连续跑约 3000 万环境步（约 17 天游戏内时长）未发散。","href":"/models/dreamer-v3","lines":["wm"],"strength":"data","tag":"能力矩阵 · 时序一致性","sources":["arXiv:2301.04104"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2301.04104","href":"https://arxiv.org/abs/2301.04104"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-dreamer-v3-temporal","pageUrl":"https://rsi.zongxin.tech/models/dreamer-v3"},{"id":"matrix:dreamer-v3:planning","type":"matrix","title":"DreamerV3 · 规划与反事实","verdict":"核心机制是让智能体在世界模型生成的想象场景中学习并改进策略，据此在 150+ 任务上追平或超过各领域专用方法。","href":"/models/dreamer-v3","lines":["wm"],"strength":"data","tag":"能力矩阵 · 规划与反事实","sources":["arXiv:2301.04104"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2301.04104","href":"https://arxiv.org/abs/2301.04104"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-dreamer-v3-planning","pageUrl":"https://rsi.zongxin.tech/models/dreamer-v3"},{"id":"matrix:v-jepa-2:temporal","type":"matrix","title":"V-JEPA 2 · 时序一致性","verdict":"在跨机械臂零样本抓取放置任务中，多步动作规划全程使用同一世界模型的想象序列，任务整体成功率约 65%–80%，说明想象没有在任务时程内明显发散。","href":"/models/v-jepa-2","lines":["wm"],"strength":"data","tag":"能力矩阵 · 时序一致性","sources":["arXiv:2506.09985"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2506.09985","href":"https://arxiv.org/abs/2506.09985"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-v-jepa-2-temporal","pageUrl":"https://rsi.zongxin.tech/models/v-jepa-2"},{"id":"matrix:v-jepa-2:controllable","type":"matrix","title":"V-JEPA 2 · 动作可控性","verdict":"V-JEPA 2-AC 用动作条件建模，在表征空间里对不同候选动作序列的后果做预测并据此选择动作；在两个实验室从未训练过的 Franka 机械臂上零样本部署，不同目标/物体对应的 pick-and-place 成功率约 65%–80%，说明模型确实按不同目标选出了不同且大多正确的动作序列。","href":"/models/v-jepa-2","lines":["wm"],"strength":"data","tag":"能力矩阵 · 动作可控性","sources":["arXiv:2506.09985"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2506.09985","href":"https://arxiv.org/abs/2506.09985"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-v-jepa-2-controllable","pageUrl":"https://rsi.zongxin.tech/models/v-jepa-2"},{"id":"matrix:v-jepa-2:planning","type":"matrix","title":"V-JEPA 2 · 规划与反事实","verdict":"给定目标图像，模型在表征空间里对候选动作序列做「脑内试错」并挑选最接近目标的一条执行；这套机制在两个实验室的零样本抓取放置任务上取得约 65%–80% 的成功率，单步规划耗时约 16 秒（比 Cosmos 基线快一个数量级）。","href":"/models/v-jepa-2","lines":["wm"],"strength":"data","tag":"能力矩阵 · 规划与反事实","sources":["arXiv:2506.09985"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2506.09985","href":"https://arxiv.org/abs/2506.09985"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-v-jepa-2-planning","pageUrl":"https://rsi.zongxin.tech/models/v-jepa-2"},{"id":"matrix:v-jepa-2:realworld","type":"matrix","title":"V-JEPA 2 · 真机迁移","verdict":"在两个实验室此前未训练过的 Franka 机械臂上零样本部署，pick-and-place 成功率约 65%–80%（因物体而异）。","href":"/models/v-jepa-2","lines":["wm"],"strength":"data","tag":"能力矩阵 · 真机迁移","sources":["arXiv:2506.09985"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2506.09985","href":"https://arxiv.org/abs/2506.09985"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-v-jepa-2-realworld","pageUrl":"https://rsi.zongxin.tech/models/v-jepa-2"},{"id":"matrix:pi-0-5:temporal","type":"matrix","title":"π0.5 · 时序一致性","verdict":"在从未见过的新家庭/卧室里执行 10–15 分钟的长程多阶段收拾任务（任务类型据论文），未见执行中期明显退化的报告。","href":"/models/pi-0-5","lines":["wm"],"strength":"data","tag":"能力矩阵 · 时序一致性","sources":["arXiv:2504.16054"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2504.16054","href":"https://arxiv.org/abs/2504.16054"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-pi-0-5-temporal","pageUrl":"https://rsi.zongxin.tech/models/pi-0-5"},{"id":"matrix:pi-0-5:controllable","type":"matrix","title":"π0.5 · 动作可控性","verdict":"flow-matching 动作专家依据语言指令与视觉观测，在 17 维动作空间（双臂+全向移动底座、14 维状态）连续输出不同动作序列。","href":"/models/pi-0-5","lines":["wm"],"strength":"data","tag":"能力矩阵 · 动作可控性","sources":["arXiv:2504.16054"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2504.16054","href":"https://arxiv.org/abs/2504.16054"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-pi-0-5-controllable","pageUrl":"https://rsi.zongxin.tech/models/pi-0-5"},{"id":"matrix:pi-0-5:realworld","type":"matrix","title":"π0.5 · 真机迁移","verdict":"双臂 ARX 机械臂+全向移动底座，在训练时从未见过的新家庭/卧室完成打扫、整理等真机任务；官方口径整体成功率约 94%。","href":"/models/pi-0-5","lines":["wm"],"strength":"data","tag":"能力矩阵 · 真机迁移","sources":["arXiv:2504.16054"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2504.16054","href":"https://arxiv.org/abs/2504.16054"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-pi-0-5-realworld","pageUrl":"https://rsi.zongxin.tech/models/pi-0-5"},{"id":"matrix:world-models-2018:controllable","type":"matrix","title":"World Models · 动作可控性","verdict":"M（MDN-RNN）以动作为条件预测下一时刻隐状态，论文对此是架构层面的定性描述，没有单独的可控性量化指标。","href":"/models/world-models-2018","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 动作可控性","sources":["arXiv:1803.10122"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/1803.10122","href":"https://arxiv.org/abs/1803.10122"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-world-models-2018-controllable","pageUrl":"https://rsi.zongxin.tech/models/world-models-2018"},{"id":"matrix:dreamer-v3:controllable","type":"matrix","title":"DreamerV3 · 动作可控性","verdict":"同一套配置同时适配离散动作（Atari 等）与连续动作控制（DeepMind Control 等），论文对此是架构层面的定性描述，没有单独衡量「动作跟随准确度」一类可控性的量化指标。","href":"/models/dreamer-v3","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 动作可控性","sources":["arXiv:2301.04104"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2301.04104","href":"https://arxiv.org/abs/2301.04104"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-dreamer-v3-controllable","pageUrl":"https://rsi.zongxin.tech/models/dreamer-v3"},{"id":"matrix:sora:temporal","type":"matrix","title":"Sora · 时序一致性","verdict":"官方技术报告展示长达 1 分钟、镜头运动与光照保持较高一致性的精选视频，但无第三方可复现的定量连贯性指标。","href":"/models/sora","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 时序一致性","sources":["官方口径"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-sora-temporal","pageUrl":"https://rsi.zongxin.tech/models/sora"},{"id":"matrix:sora:physics","type":"matrix","title":"Sora · 物理合理性","verdict":"OpenAI 官方承认「模拟复杂物理的能力有限，理解因果关系和分辨左右的能力也有限」；另有研究（arXiv:2411.02385，测试的是自建的同类扩散视频模型而非 Sora 本身）发现规模扩大不足以让此类模型学到可泛化的物理规律。","href":"/models/sora","lines":["wm"],"strength":"demo","stance":"有争议","dispute":"「Sora 是不是学到了物理」本身就是争议点：生成派认为像素级预测隐含地学到了物理规律，反方（arXiv:2411.02385）发现同类模型规模扩大也没有泛化的物理规律，官方自己也把物理模拟列为已知限制。","tag":"能力矩阵 · 物理合理性","sources":["官方口径"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-sora-physics","pageUrl":"https://rsi.zongxin.tech/models/sora"},{"id":"matrix:genie-3:temporal","type":"matrix","title":"Genie 3 · 时序一致性","verdict":"官方称环境一致性可维持数分钟；至今没有发表技术论文，24fps/720p 为 2025-08 研究预览版规格，非独立可复现的测量结果。","href":"/models/genie-3","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 时序一致性","sources":["官方口径（Google DeepMind 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-genie-3-temporal","pageUrl":"https://rsi.zongxin.tech/models/genie-3"},{"id":"matrix:genie-3:controllable","type":"matrix","title":"Genie 3 · 动作可控性","verdict":"支持用文字指令实时向场景「注入」事件（如天气变化、角色出现），为官方演示描述，无量化可控性指标。","href":"/models/genie-3","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 动作可控性","sources":["官方口径（Google DeepMind 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-genie-3-controllable","pageUrl":"https://rsi.zongxin.tech/models/genie-3"},{"id":"matrix:genie-3:memory","type":"matrix","title":"Genie 3 · 长程记忆","verdict":"官方称视觉记忆可回溯约 1 分钟前的画面（转身再回头场景大概率还在），该规格适用于 2025-08 研究预览版，2026-01 后的开放版本未见重新公布。","href":"/models/genie-3","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 长程记忆","sources":["官方口径（Google DeepMind 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-genie-3-memory","pageUrl":"https://rsi.zongxin.tech/models/genie-3"},{"id":"matrix:lingbot-world-2:temporal","type":"matrix","title":"LingBot-World 2 · 时序一致性","verdict":"论文摘要称因果预训练配合 MoBA 注意力可支持『unbounded interaction horizon』的小时级持续生成，但 60fps 衡量的是吞吐而非时序一致性本身，论文未给出独立的场景一致性量化指标。","href":"/models/lingbot-world-2","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 时序一致性","sources":["arXiv:2607.07534"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2607.07534","href":"https://arxiv.org/abs/2607.07534"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-lingbot-world-2-temporal","pageUrl":"https://rsi.zongxin.tech/models/lingbot-world-2"},{"id":"matrix:lingbot-world-2:controllable","type":"matrix","title":"LingBot-World 2 · 动作可控性","verdict":"论文摘要描述 pilot agent 负责规划角色行为、director agent 负责合成新环境元素，并支持多名玩家同时进入同一世界互动，但这是架构层面的定性描述，没有单独的可控性量化指标。","href":"/models/lingbot-world-2","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 动作可控性","sources":["arXiv:2607.07534"],"sourceKinds":["论文"],"officialOnly":false,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2607.07534","href":"https://arxiv.org/abs/2607.07534"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-lingbot-world-2-controllable","pageUrl":"https://rsi.zongxin.tech/models/lingbot-world-2"},{"id":"matrix:lingbot-world-2:physics","type":"matrix","title":"LingBot-World 2 · 物理合理性","verdict":"第三方实测认为可玩性与物理表现令人印象深刻，但为主观定性评价，没有公开的物理合规性/碰撞基准数字。","href":"/models/lingbot-world-2","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 物理合理性","sources":["第三方实测（爱范儿 ifanr.com/1671406）"],"sourceKinds":["第三方","媒体"],"officialOnly":false,"verifiability":"media","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2607.07534","href":"https://arxiv.org/abs/2607.07534"},{"label":"ifanr.com/1671406","href":"https://ifanr.com/1671406"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-lingbot-world-2-physics","pageUrl":"https://rsi.zongxin.tech/models/lingbot-world-2"},{"id":"matrix:lingbot-world-2:memory","type":"matrix","title":"LingBot-World 2 · 长程记忆","verdict":"官方展示多人同时进入同一持续生成的世界互动，但第三方实测记录了转身返回后出现此前不存在的墙这一空间不一致问题（官方称为体验版工程限制），暂无量化场景一致性数字。","href":"/models/lingbot-world-2","lines":["wm"],"strength":"demo","stance":"有争议","dispute":"官方演示与第三方实测结论相反：实测遇到转身返回后凭空出现的墙，官方解释为体验版的工程限制而非模型能力问题。两边都没有量化的场景一致性数字。","tag":"能力矩阵 · 长程记忆","sources":["arXiv:2607.07534；第三方实测（爱范儿 ifanr.com/1671406）"],"sourceKinds":["第三方","论文","媒体"],"officialOnly":false,"verifiability":"media","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2607.07534","href":"https://arxiv.org/abs/2607.07534"},{"label":"ifanr.com/1671406","href":"https://ifanr.com/1671406"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-lingbot-world-2-memory","pageUrl":"https://rsi.zongxin.tech/models/lingbot-world-2"},{"id":"matrix:pi-0-5:planning","type":"matrix","title":"π0.5 · 规划与反事实","verdict":"官方博客消融实验称，去掉高层子任务预测（语义层规划）后开放世界成功率从约 94% 降到约 31%；具体百分比来自官网博客转述，未在 arXiv 论文正文逐一核对到对应表格。","href":"/models/pi-0-5","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 规划与反事实","sources":["官方口径（Physical Intelligence 官网博客 pi05）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"checkable","artifact":"开放权重","links":[{"label":"arxiv.org/abs/2504.16054","href":"https://arxiv.org/abs/2504.16054"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-pi-0-5-planning","pageUrl":"https://rsi.zongxin.tech/models/pi-0-5"},{"id":"matrix:marble:physics","type":"matrix","title":"Marble · 物理合理性","verdict":"官方称同一模型同时输出高斯泼溅与「可供物理引擎运算」的碰撞网格，但官方自己也承认 AI 生成几何有时会自相交或比例失真，导致「不合理的物理」，无公开的物理准确度基准数字。","href":"/models/marble","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 物理合理性","sources":["官方口径（World Labs 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-marble-physics","pageUrl":"https://rsi.zongxin.tech/models/marble"},{"id":"matrix:marble:memory","type":"matrix","title":"Marble · 长程记忆","verdict":"官方演示可从不同角度绕回同一位置，看到的仍是同一批物体、同样的位置，即「持久性」，但没有量化的场景保真度指标。","href":"/models/marble","lines":["wm"],"strength":"demo","tag":"能力矩阵 · 长程记忆","sources":["官方口径（World Labs 博客）"],"sourceKinds":["官方口径"],"officialOnly":true,"verifiability":"opaque","links":[],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#matrix-marble-memory","pageUrl":"https://rsi.zongxin.tech/models/marble"},{"id":"crossing:mhs","type":"crossing","title":"物理版 MCP：Model Hardware Standard","verdict":"它一行动力学预测能力都不提供，但它降低的恰恰是能力矩阵「真实世界」那一列最贵的东西——把模型接到真机上的成本。","href":"/crossings#mhs","lines":["sota","wm","ai4s"],"strength":"argument","stance":"有争议","dispute":"推理链有三处可能断：标准之争不由技术优劣决定，实验室仪器的动作空间与接触丰富的操作任务并不重合，而且接线便宜不等于数据变好——世界模型缺的还有数据的多样性与覆盖面。","tag":"传导 · 大模型 →","sources":["官方口径（anthropic.com/news/model-hardware-standard-research-preview）","公开报道（CNBC、Fortune、The Register 2026-08-27/28）"],"sourceKinds":["官方口径","媒体"],"officialOnly":false,"verifiability":"opaque","links":[{"label":"anthropic.com/news/model-hardware-standard-research-preview","href":"https://anthropic.com/news/model-hardware-standard-research-preview"}],"addedIn":"2026-08","judgments":[],"url":"https://rsi.zongxin.tech/evidence#crossing-mhs","pageUrl":"https://rsi.zongxin.tech/crossings#mhs"}],"judgments":[{"id":"thesis-verifier","question":"谁先被 AI 接手，是不是由「验证一个结果有多便宜」决定？","value":"成立","summary":"RSI 最先在代码上发生、AI4S 最先在蛋白质与材料计算上成熟，都符合「验证器越便宜越早」。反例目前没有出现，但样本仍少。","confidence":"中","stability":"稳定","support":[{"kind":"crossing","slug":"rsi"},{"kind":"rsi","slug":"darwin-godel-machine"},{"kind":"ai4s","slug":"alphafold-2"}],"contra":[],"basis":[{"label":"Jason Wei：验证的不对称与「验证者定律」——一项任务越容易验证，AI 越快学会","source":"jasonwei.net，2025-07","url":"https://www.jasonwei.net/blog/asymmetry-of-verification-and-verifiers-law"},{"label":"OpenAI GDPval：职场任务的自动评分与专家评分一致率只有约 66%，办公任务的验证器仍然昂贵","source":"OpenAI，2025-09（官方口径）；arXiv:2510.04374"}],"wouldChangeIf":["某个验证很贵的领域（如药物临床、真机操作）在没有新验证手段的情况下，进展速度反超验证便宜的领域。","代码领域的自我改进在测试充足的条件下长期停滞，说明瓶颈不在验证器。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"成立","reason":"RSI 最先在代码上发生，AI4S 最先在蛋白质与材料计算上成熟，都符合这条规律。"}],"href":"/about","url":"https://rsi.zongxin.tech/judgments/thesis-verifier"},{"id":"thesis-coding","question":"编程被 AI 接手到什么程度了？","value":"已进入生产","summary":"测试与编译器让编程的验证几乎免费，所以它最先被接手：头部实验室称合入的代码九成以上由模型写成。但「写得出」不等于「省时间」，独立实验里资深开发者用 AI 反而变慢。","confidence":"中","stability":"稳定","support":[{"kind":"rsi","slug":"anthropic-when-ai-builds-itself"},{"kind":"rsi","slug":"anthropic-90-percent-merged"}],"contra":[{"kind":"crossing","slug":"harness-gap"}],"basis":[{"label":"METR 随机对照实验：16 位资深开源开发者，允许用 AI 时完成任务的时间反而多了 19%（2025 年初的工具）","source":"arXiv:2507.09089"}],"wouldChangeIf":["第三方对照实验测出 AI 让资深开发者的交付速度稳定提升，并被多家复现——置信度升为「高」。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"已进入生产","reason":"测试与编译器让编程的验证几乎免费；头部实验室公开称八成以上代码由模型写成。"}],"href":"/frontier","url":"https://rsi.zongxin.tech/judgments/thesis-coding"},{"id":"thesis-office","question":"办公（写报告、做表、操作电脑）被 AI 接手到什么程度了？","value":"进行中","summary":"办公任务大多没有测试可跑：GDPval 上前沿模型对行业专家的胜率接近一半，但自动评分与专家评分的一致率只有约 66%——验证本身还贵，所以它落后于编程。","confidence":"低","stability":"观察中","support":[],"contra":[],"basis":[{"label":"OpenAI GDPval：44 个职业、1,320 个任务，专家盲评；首版前沿模型胜率约 48%，自动评分与专家一致率约 66%","source":"OpenAI，2025-09（官方口径）；arXiv:2510.04374"}],"wouldChangeIf":["出现与专家评分一致率超过 90% 的自动评审，并被用来规模化训练办公智能体——会提前。","第三方在真实企业环境里测出计算机操作智能体的端到端完成率接近人类——改为「已进入生产」。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"进行中","reason":"办公没有便宜的验证器，按主轴应当晚于编程，单独成一级。"}],"href":"/frontier","url":"https://rsi.zongxin.tech/judgments/thesis-office"},{"id":"thesis-ai4s-two-years","line":"ai4s","question":"科研（AI4S）会在约两年内成熟吗？","value":"约两年内","summary":"预测与设计两层已经可用，蛋白质方向有湿实验验证；但真正的硬终点（临床、第三方复现）目前只有零星几例，这一判断靠的是趋势，不是结果。","criterion":"2028 年 9 月对账。算「成熟」要满足其一：至少两个方向出现 L3 闭环、且结果经过独立检验；或首个以 AI 设计为主的药物在二期达到主要疗效终点。","forecast":{"due":"2028-09","scope":"「成熟」指至少两个方向出现经独立检验的 L3 闭环，或 AI 设计的药物在二期达到主要终点。"},"milestone":"Rentosertib 后续试验读出结果；自动实验室的结果被第三方复现","confidence":"低","stability":"观察中","support":[{"kind":"ai4s","slug":"rentosertib"},{"kind":"ai4s","slug":"co-scientist"},{"kind":"crossing","slug":"science-workbench"}],"contra":[{"kind":"ai4s","slug":"a-lab"},{"kind":"rsi","slug":"princeton-open-ended-research"},{"kind":"ai4s","slug":"isomorphic-pipeline"}],"wouldChangeIf":["自动实验室把合成与表征的周期从几周压到几天，并有第三方复现——会提前。","Rentosertib 三期失败，或首批 AI 设计药物普遍在二期折戟——会推后。","AI 科学家在开放式研究判断上一年内仍无可测进步——会推后。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"约两年内","reason":"起始判断，依据是预测与设计两层已经可用。"}],"href":"/ai4s","url":"https://rsi.zongxin.tech/judgments/thesis-ai4s-two-years"},{"id":"thesis-physical-early","line":"wm","question":"理解并操控物理世界（世界模型）走到哪了？","value":"行业早期","summary":"证据主要在能力矩阵：四条技术路线都有演示，少有跨场景的实测。2026 年秋，世界动作模型与 VLA 给出了真机上少样本学新任务的成功率，通用大模型也能在模拟器里直接控制机器人，但真机数字都是厂商自报、还没有第三方复现。物理世界只有真机能验证，又慢又贵，所以它仍在阶梯的最后一级。","milestone":"世界模型训练的机器人策略在没见过的真实场景里稳定泛化，并被第三方复现","confidence":"中","stability":"观察中","support":[],"contra":[{"kind":"crossing","slug":"frontier-robot-policy"}],"wouldChangeIf":["世界模型生成的仿真数据训练出的机器人策略，在没见过的真实场景里稳定泛化，且有第三方复现。","某个通用机器人在开放家庭或工厂环境中规模化部署并公开失败率。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"行业早期","reason":"起始判断：多数能力只有演示，真机实测稀少。"}],"href":"/matrix","url":"https://rsi.zongxin.tech/judgments/thesis-physical-early"},{"id":"rsi-stage-1","line":"rsi","question":"RSI 阶段 1「经验自进化」在发生吗？","value":"已规模化","summary":"只改经验、不改权重的方法已进产品，门槛从「能不能」转成「经验会不会越积越乱」。","milestone":"长期运行中记忆会不会退化，有没有被系统性测出来","confidence":"高","stability":"稳定","support":[{"kind":"rsi","slug":"ace"},{"kind":"rsi","slug":"voyager"},{"kind":"crossing","slug":"persistent-context"}],"contra":[],"wouldChangeIf":["长期运行中记忆污染导致的退化被系统性测出，主流产品回退跨会话记忆。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"已规模化","reason":"ACE 等方法有实测提升，前沿产品把跨会话记忆做成默认能力。"}],"href":"/rsi#stage-1","url":"https://rsi.zongxin.tech/judgments/rsi-stage-1"},{"id":"rsi-stage-2","line":"rsi","question":"RSI 阶段 2「Harness 自进化」在发生吗？","value":"进行中","summary":"改写脚手架有扎实的基准数字，但生产中是否常态化，外界看不到。","milestone":"有实验室公开：生产中的智能体脚手架主要由智能体自己迭代","confidence":"中","stability":"观察中","support":[{"kind":"rsi","slug":"darwin-godel-machine"},{"kind":"rsi","slug":"meta-harness"},{"kind":"rsi","slug":"weng-harness"}],"contra":[{"kind":"crossing","slug":"harness-gap"}],"wouldChangeIf":["有实验室公开：生产环境里的智能体脚手架主要由智能体自己迭代——会升为「已规模化」。","自动搜索出的脚手架被证明普遍过拟合基准、换任务即失效——会降级。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"进行中","reason":"达尔文-哥德尔机与 Meta-Harness 在基准上超过人工设计。"}],"href":"/rsi#stage-2","url":"https://rsi.zongxin.tech/judgments/rsi-stage-2"},{"id":"rsi-stage-3","line":"rsi","question":"RSI 阶段 3「AI 加速 AI 研究」在发生吗？","value":"进行中","summary":"工程层已大面积自动化，研究判断层没有；关键数字全是实验室自测，没有第三方复核。","milestone":"METR、Epoch 等第三方独立测出 AI 对前沿模型研发的加速倍数","confidence":"中","stability":"观察中","support":[{"kind":"rsi","slug":"alphaevolve"},{"kind":"rsi","slug":"openai-research-intern"},{"kind":"rsi","slug":"anthropic-when-ai-builds-itself"},{"kind":"rsi","slug":"ai-scientist-nature"}],"contra":[{"kind":"rsi","slug":"princeton-open-ended-research"}],"wouldChangeIf":["METR、Epoch 等第三方独立测出 AI 对前沿模型研发周期的加速倍数——置信度会升。","AI 提出的研究方向（不只是执行）被用于训练已部署的前沿模型并公开——会升为「已规模化」。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"进行中","reason":"两家头部实验室公开了自测数字；普林斯顿的独立实验给出反方。"}],"href":"/rsi#stage-3","url":"https://rsi.zongxin.tech/judgments/rsi-stage-3"},{"id":"rsi-stage-4","line":"rsi","question":"RSI 阶段 4「AI 端到端训练下一代 AI」在发生吗？","value":"未发生","summary":"没有公开证据表明它已发生，这一格目前只有实验室的目标与时间表。","milestone":"有实验室公开：某个已部署模型主要由上一代模型自主研发并训练","confidence":"高","stability":"稳定","support":[{"kind":"rsi","slug":"stage4-stated-goals"}],"contra":[],"wouldChangeIf":["有实验室公开承认某个已部署模型主要由上一代模型自主研发并训练。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"未发生","reason":"只有目标与时间表（OpenAI 定在 2028 年 3 月），没有已发生的证据。"}],"href":"/rsi#stage-4","url":"https://rsi.zongxin.tech/judgments/rsi-stage-4"},{"id":"ai4s-protein","line":"ai4s","question":"蛋白质与分子设计：AI 在科研闭环里走到了哪一环？","value":"最远到 L2","summary":"预测已是基础设施，并经过 CASP 盲测这样的独立检验；设计有作者自己的湿实验验证，还没有闭环系统。","confidence":"高","stability":"观察中","support":[{"kind":"ai4s","slug":"rfdiffusion"},{"kind":"ai4s","slug":"esm3"},{"kind":"ai4s","slug":"alphaproteo"}],"contra":[{"kind":"ai4s","slug":"pinal"}],"wouldChangeIf":["出现由 AI 决定下一轮设计、机器人执行表达与测定的蛋白质闭环系统——升到 L3。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"最远到 L2","reason":"设计出的蛋白已在实验室表达并按预期工作。"}],"href":"/ai4s/protein","url":"https://rsi.zongxin.tech/judgments/ai4s-protein"},{"id":"ai4s-drug","line":"ai4s","question":"药物发现：AI 在科研闭环里走到了哪一环？","value":"最远到 L2","summary":"AI 在药物里的角色到「设计」为止（靶点与分子）；检验走得最远：一例已进入三期临床，结果要等到 2027 年以后。","confidence":"中","stability":"观察中","support":[{"kind":"ai4s","slug":"rentosertib"}],"contra":[{"kind":"ai4s","slug":"isomorphic-pipeline"}],"wouldChangeIf":["Rentosertib 三期读出失败——角色仍是 L2，但本方向的置信度会降。","第二、第三个 AI 发现的分子进入二期以上——置信度会升。","药物研发出现 AI 自主提出假设、设计并迭代实验的闭环，且产出进入临床前——升到 L3。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"最远到 L4","reason":"Rentosertib 进入三期临床。"},{"ym":"2026-09","from":"最远到 L4","to":"最远到 L2","reason":"尺子改了，事实没改：阶梯拆成「AI 的角色」和「结果的检验」两个轴，L4 不再表示「外部检验」。Rentosertib 的角色是 L2 设计，检验是临床三期——它仍是 AI4S 检验走得最远的条目。","evidence":[{"kind":"ai4s","slug":"rentosertib"}]}],"href":"/ai4s/drug","url":"https://rsi.zongxin.tech/judgments/ai4s-drug"},{"id":"ai4s-cell","line":"ai4s","question":"虚拟细胞与生物基础模型：AI 在科研闭环里走到了哪一环？","value":"最远到 L2","summary":"有了第一个公开统一评测，但模型在没见过的细胞类型上是否还准，是最大的问号。","confidence":"中","stability":"稳定","support":[{"kind":"ai4s","slug":"evo-2"},{"kind":"ai4s","slug":"xtrimo-sc-perturb"}],"contra":[{"kind":"ai4s","slug":"proteintalks"}],"wouldChangeIf":["虚拟细胞模型的预测被用来指导一轮扰动实验，并在未见过的细胞类型上被实验证实——升到 L3。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"最远到 L2","reason":"Evo 2 生成了可用的基因组片段；Arc 挑战赛建立了公开评测。"}],"href":"/ai4s/cell","url":"https://rsi.zongxin.tech/judgments/ai4s-cell"},{"id":"ai4s-materials","line":"ai4s","question":"材料研发：AI 在科研闭环里走到了哪一环？","value":"最远到 L3","summary":"自动实验室已经在闭环跑，但标志性案例 A-Lab 的「新材料」结论被更正，闭环本身的可信度仍在被检验。","confidence":"低","stability":"观察中","support":[{"kind":"ai4s","slug":"a-lab"},{"kind":"ai4s","slug":"mattergen"}],"contra":[{"kind":"ai4s","slug":"gnome"},{"kind":"ai4s","slug":"periodic-labs"}],"wouldChangeIf":["自动实验室发现的材料被第三方复现并确认新颖——检验升到「独立检验」，置信度会升。","Periodic Labs 等公司长期没有公开发表的发现——置信度会降。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"最远到 L3","reason":"A-Lab、Periodic Labs 在做自动实验闭环。"}],"href":"/ai4s/materials","url":"https://rsi.zongxin.tech/judgments/ai4s-materials"},{"id":"ai4s-scientist","line":"ai4s","question":"AI 科学家：AI 在科研闭环里走到了哪一环？","value":"最远到 L3","summary":"完整科研循环已有演示，但实验仍靠人或半自动完成；收录的发现还没有经过第三方复现。","confidence":"中","stability":"观察中","support":[{"kind":"ai4s","slug":"robin"},{"kind":"ai4s","slug":"co-scientist"},{"kind":"rsi","slug":"ai-scientist-nature"}],"contra":[{"kind":"rsi","slug":"princeton-open-ended-research"}],"wouldChangeIf":["AI 科学家独立提出的发现被第三方复现——检验升到「独立检验」。","AI 自己选定研究问题并提出新方法，被同行认可为原创贡献——升到 L4。"],"reviewedIn":"2026-11-11","history":[{"ym":"2026-09","to":"最远到 L3","reason":"Robin 与 Co-Scientist 跑通了「实验室在环」的循环。"}],"href":"/ai4s/scientist","url":"https://rsi.zongxin.tech/judgments/ai4s-scientist"}]}