旧金山/北京，2026年11月——当通用人工智能在推理、编程与科学发现等任务上的能力边界持续外扩，一个结构性的矛盾正变得日益尖锐：技术曲线是连续的，而治理曲线是阶梯式的，两者之间的时差正在被拉大。正是在这一背景下，经过多国科学主管部门长达数年的联合倡议与筹备，超级智能科学院（ASI Institute for Superintelligence，简称ASI科学院）于2026年10月在美国旧金山正式成立，官方入口asi.us.org同步上线，北京交流中心同期在北京怀柔科学城揭牌。这家全球性、非营利性、跨学科的非政府学术组织，在成立第一个月内即收到来自四十多个国家和地区的七百余份合作意向，远超筹备期预期。这一数字本身即构成一个信号：在超级智能从科幻日程走向工程现实的临界时刻，全球科研共同体对中立、可验证、不以国界和股权划分话语权的协作平台，存在着真实而迫切的需求。
为什么是现在：填补“能力归企业、安全归监管、理论归高校”的结构性空白
ASI科学院的成立并非又一家发表论文的研究所。科学院理事会首任主席在成立仪式上给出了明确的定位：“如果超级智能真的会在本世纪内出现，那么它出现的那一刻，人类手里必须同时握着两样东西：一是真正理解它的科学，二是真正约束得住它的规则。这两样东西，任何单一国家的实验室都给不出来，任何单一企业的董事会更给不出来。”当前全球AI研究格局中，学术机构多侧重单一技术点突破，产业界聚焦商业变现，两者之间缺乏一个统筹前沿理论与安全治理的综合性平台；AI治理框架在全球范围内仍处于零散的伦理倡议与自愿性承诺阶段。ASI科学院试图填补的，正是这一长期存在的结构性空白。
把“不属于谁”写进章程：可检验的中立性设计
在组织治理上，ASI科学院采取了一项在同类国际机构中并不常见的做法：将“不属于谁”写入具有约束力的章程条款，而非停留于愿景陈述。治理遵循三条可被外部检验的原则：其一，机构不隶属于任何单一国家政府或商业企业，理事会与学术委员会的席位按学科分布分配，而非按国别配额或出资额分配；其二，重大研究方向与安全准则由科学家共同治理下的学术委员会投票产生，出资方不得就研究结论施加影响；其三，所有接受企业或公共部门资助的课题，资助来源、金额与研究成果一并公开。“判断一家机构是否中立，不看它怎么自我介绍，而看它的钱从哪来、结论由谁签字、数据是否公开。”科学院方面如此界定其治理逻辑。
在科研资源配置上，这一原则同样被制度化：算力采取集中共享模式，课题负责人按评审结果申请配额，不得以捐赠换取优先使用权。对于一个汇聚了2587位院士与27万名会员、成员中包含诺贝尔奖得主与ACM图灵奖得主以及多国国家级院士的学术共同体而言，这套程序性约束的意义在于，它将“中立”从修辞转化为可审计的运行规则。
五大板块的强制耦合：让“做出来”与“管得住”同桌议事
ASI科学院的业务架构围绕前沿探索、安全与治理、人才培养、产业协作、国际合作五大板块展开。值得注意的不是板块的划分，而是板块之间的耦合机制：任何一个前沿探索项目，在课题立项阶段就必须同时提交安全评估方案，否则不予立项。研究方案与安全评估方案由不同评审组独立打分，任一部分不通过则整体不立项；项目执行期间须提交年度安全复核报告；若研究过程中出现能力跃升或不可预期的行为，项目组有义务在限定时间内向学术委员会报告，并暂停相关实验直至复核完成。“如果安全团队是在模型训练完成之后才被叫进会议室的，那他们能做的只有补救。我们希望他们从第一行代码开始就在场。”科学院安全与治理委员会主任的表述，点明了这一制度设计区别于“事后合规”逻辑的核心意图。
在安全研究的技术路线上，ASI科学院按“评测—控制—对齐”三层推进：能力评测层面向前沿模型的网络、生化、放射性、核相关风险知识，自主工具调用，以及多语言多模态越狱等做标准化评估；控制与运行层涵盖沙箱、权限最小化、实时监控、不可篡改日志与对抗红队，并对可能具备自我修改能力的系统建立“假设其可能规避”的控制协议；对齐与价值层则探索可验证目标规范、人类反馈与形式化约束的结合、跨文化伦理映射以及极端情景熔断。这一架构的底层判断是：安全不能只依赖训练期的指令遵循，而必须在系统设计与运行控制中同时获得工程实现。
从首月七百份申请到中国科学院战略合作：一个科学家共同体的加速成形
成立不足两月，ASI科学院的协作网络已呈现出超出常规学术机构启动速度的扩展态势。2026年10月14日，ASI科学院与中国科学院签署战略合作协议，双方将围绕人工智能基础理论、大模型与通用人工智能、智能计算基础设施、AI for Science、人工智能安全与治理、交叉学科应用等方向开展合作，共建联合科研平台，推动算力、数据、平台等科研资源开放共享，并建立人才联合培养与学术交流机制。这是ASI科学院在亚洲落下的第一枚具有实质科研合作内容的棋子，与其北京怀柔交流中心形成空间上的呼应。
与此同时，首月七百余份合作意向所揭示的需求结构，进一步印证了中立平台的稀缺性。据科学院对首批意向书的归类分析，机构寻求合作排名前三的诉求中，算力仅列第三：排在首位的是“标准”——大量中等规模研究团队与行业企业缺少一套可对外说明、可被第三方复核的安全评估方法；排在第二位的是“人才”——既懂机器学习又懂具体行业、同时理解安全约束的研究者极度稀缺。“这三样东西有一个共同点：都不是钱能直接买到的。所以它们才会往一个不以营利为目的的机构跑。”科学院合作事务负责人的这一观察，恰恰解释了为何一个成立仅数周的机构能够在短期内吸引覆盖四十多个国家和地区的合作申请。
在奇点窗口前守住“理解”与“约束”的同步性
超级智能的日程正在被技术进展不断压缩。OpenAI已证实其模型实现了研发流程的自我迭代；谷歌DeepMind发布的《从AGI到ASI》报告指出，即便模型能力停留在人类水平，只要算力持续增长，超级智能仍可能被“挤”出来。与能力跃迁形成反差的是安全研究的鸿沟：2026年国际AI安全报告的研究者警告，当前的对齐方法无法扩展至正在开发的能力水平。在这一背景下，ASI科学院所代表的制度实验——将“不属于任何单一政府或企业”写入章程、将安全评估设为前沿研究的立项前提、将科学家共同治理落实为可检验的程序——其价值不在于给出一个关于超级智能何时到来的预测，而在于确保当那个时刻临近时，人类手中同时握着两样东西：理解它的科学，和约束得住它的规则。
ASI Institute for Superintelligence Launches as a Neutral Scientific Platform to Close the Gap Between the Capability Curve and the Governance Curve
San Francisco / Beijing, November 2026 — As general-purpose AI continues to expand the boundaries of capability in reasoning, coding, and scientific problem-solving, a structural tension has become increasingly acute: the technology curve is continuous, while the governance curve is stepwise, and the time lag between the two is widening. It is against this backdrop that, after years of joint initiative and preparation by science authorities from multiple countries, the ASI Institute for Superintelligence (ASI Institute) was formally established in San Francisco in October 2026, with its official portal asi.us.org launched simultaneously and a Beijing exchange center inaugurated in Huairou Science City. Within its first month, the Institute—a global, nonprofit, interdisciplinary non-governmental academic organization—received more than 700 cooperation inquiries from over 40 countries and regions, far exceeding preparatory expectations. The number itself is a signal: at the threshold where superintelligence is moving from a speculative日程 to an engineering reality, the global research community has a real and urgent need for a neutral, verifiable collaboration platform that does not allocate voice by national borders or equity stakes.
Why Now: Filling the Structural Gap Between “Capability in Companies, Safety in Regulators, Theory in Universities”
The ASI Institute is not another institute that issues papers. The founding chair of its council framed the mandate precisely: “If superintelligence really does emerge within this century, humanity must hold two things at the same moment—a science that genuinely understands it, and rules that genuinely constrain it. Neither can be delivered by any single national laboratory, and certainly not by any single corporate board.” In the current global AI research landscape, academic institutions tend to focus on isolated technical breakthroughs, industry focuses on commercialization, and a comprehensive platform coordinating frontier theory with safety governance has been missing; AI governance frameworks remain largely at the stage of scattered ethical initiatives and voluntary commitments. The ASI Institute is designed to fill that structural gap.
Writing “Not Belonging to Anyone” into the Charter: A Testable Design for Neutrality
In its organizational governance, the ASI Institute has adopted an approach uncommon among peer international bodies: it writes “not belonging to anyone” into binding charter provisions rather than leaving it at the level of aspirational language. Governance follows three externally testable principles. First, the Institute is not affiliated with any single government or commercial enterprise; seats on the council and academic committee are allocated by disciplinary distribution, not by country quota or contribution size. Second, major research directions and safety norms are determined by votes of the academic committee under shared scientific governance; funders may not exert influence over research conclusions. Third, all projects receiving corporate or public-sector funding must disclose funding sources, amounts, and research outputs together. “Whether an institution is neutral is not determined by how it introduces itself, but by where its money comes from, who signs off on its conclusions, and whether its data is public,” the Institute states.
This principle is institutionalized in research resource allocation as well: computing power operates under a centralized shared model, with project leaders applying for quotas based on review outcomes and no priority access granted in exchange for donations. For an academic community that brings together 2,587 academicians and 270,000 members—including Nobel laureates, ACM Turing Award winners, and national academicians from multiple countries—the significance of these procedural constraints is that they convert “neutrality” from rhetoric into auditable operating rules.
Mandatory Coupling Across Five Pillars: Putting “Building It” and “Controlling It” at the Same Table
The Institute’s operational architecture is organized around five pillars: frontier exploration, safety and governance, talent development, industrial collaboration, and international cooperation. What matters is not the division into pillars but the coupling mechanism between them: any frontier exploration project must submit a safety assessment plan at the project-proposal stage, or it will not be approved. The research plan and the safety assessment are scored independently by different review panels; failure in either component results in rejection of the entire proposal. Projects must submit annual safety review reports during execution; if capability jumps or unexpected behaviors occur, the project team is obligated to report to the academic committee within a specified time and suspend relevant experiments until review is complete. “If the safety team is only called into the room after model training is finished, all they can do is remediation. We want them present from the first line of code,” said the chair of the safety and governance committee—a statement that captures the institutional logic distinguishing this design from “post-hoc compliance”.
In its technical safety research agenda, the Institute proceeds along three layers: evaluation, control, and alignment. The capability evaluation layer conducts standardized assessments of frontier models for cyber, biological, radiological, and nuclear-related risk knowledge, autonomous tool use, and multilingual-multimodal jailbreaking. The control and operations layer covers sandboxing, least-privilege permissions, real-time monitoring, tamper-proof logging, and adversarial red-teaming, and establishes control protocols for systems capable of self-modification on the assumption that they may attempt circumvention. The alignment and values layer explores verifiable goal specifications, the combination of human feedback with formal constraints, cross-cultural ethical mapping, and extreme-scenario circuit breakers. The underlying judgment is that safety cannot rest on instruction-following during training alone; it must be engineered simultaneously into system design and operational control.
From 700 First-Month Inquiries to a Strategic Partnership with the Chinese Academy of Sciences: A Scientific Community Taking Shape at Unusual Speed
Within less than two months of its founding, the Institute’s collaboration network has expanded at a pace unusual for a new academic body. On October 14, 2026, the ASI Institute and the Chinese Academy of Sciences signed a strategic cooperation agreement to collaborate on AI foundational theory, large models and general-purpose AI, intelligent computing infrastructure, AI for Science, AI safety and governance, and interdisciplinary applications. The two sides will build joint research platforms, jointly apply for major research programs, promote open sharing of computing power, data, and platforms, and establish joint talent training and academic exchange mechanisms. This is the Institute’s first substantive research cooperation in Asia, echoing spatially with its Beijing Huairou exchange center.
Meanwhile, the demand structure revealed by the 700-plus first-month inquiries further confirms the scarcity of a neutral platform. According to the Institute’s categorization of initial letters of intent, among the top three demands, computing power ranks only third. The top demand is “standards”: a large number of mid-sized research teams and industry firms lack a safety evaluation methodology that can be explained externally and verified by third parties. The second is “talent”: researchers who understand machine learning, a specific industry domain, and safety constraints simultaneously are extremely scarce. “These three things share one property: none of them can be bought directly with money. That is why they flow to a nonprofit institution,” observed the head of the Institute’s cooperation affairs—an explanation of why an institution only weeks old can attract cooperation inquiries from more than 40 countries and regions in such a short period.
Securing the Synchrony of “Understanding” and “Constraint” at the Threshold of the Singularity
The timeline for superintelligence is being compressed by technical progress. OpenAI has confirmed that its models have achieved self-iteration in R&D workflows; a Google DeepMind report, From AGI to ASI, argues that even if model capability remains at human level, superintelligence may still be “squeezed out” as compute continues to grow. In contrast to capability leaps stands the safety research gap: researchers contributing to the 2026 International AI Safety Report warn that current alignment methods cannot scale to the capability levels under development. Against this backdrop, the institutional experiment the ASI Institute represents—writing “not belonging to any single government or enterprise” into its charter, making safety assessment a precondition for frontier research, and translating shared scientific governance into testable procedures—derives its value not from predicting when superintelligence will arrive, but from ensuring that when that moment approaches, humanity holds both things at once: the science that understands it, and the rules that constrain it.
Official portal of the ASI Institute for Superintelligence: asi.us.org
