8月，麻省理工学院（MIT）的研究人员将一群AI智能体投放到一个共享世界中，未给予任何指令。他们展示了该群体能够发明、构建并维护技术，其表现优于同等数量的单独工作的智能体。该研究运行在一个托管的前沿模型上。
In August, researchers at MIT dropped a society of AI agents into a shared world with no instructions. They showed it could invent, build and maintain technology better than the same number of agents working alone. The study ran on a hosted frontier model.
与Apex Fusion合作的一个研究团队现已在四张RTX 3090游戏显卡上复现了核心实验。Apex Fusion是一家构建AI智能体基础设施的瑞士基金会。它未改动MIT的代码和指令。该运行耗时约26小时，共进行12,800次模型调用。相关论文已发布在Apex Fusion的GitHub上，完整的追踪数据也已公开。
A research team working with Apex Fusion has now reproduced the central experiment on four RTX 3090 gaming GPUs. Apex Fusion is a Swiss foundation building infrastructure for AI agents. It used MIT’s code and commands unchanged. The run took about 26 hours and 12,800 model calls. The paper, posted on Apex Fusion’s GitHub, and the full trace data are public.
MIT发现的SwarmWorld由MIT原子与分子力学实验室的Pal、Wang和Buehler完成。研究将50个或更多语言模型智能体置于一个共享的模拟世界中。没有分配角色，也没有配方。智能体自主探索、收集材料、运行测试、构建持久化技术，并编写出在智能体消失后仍能持续运行的程序。
What MIT found SwarmWorld is by Pal, Wang and Buehler at MIT’s Laboratory for Atomistic and Molecular Mechanics. It placed 50 or more language-model agents in a shared simulated world. There were no assigned roles and no recipes. The agents explored, gathered materials, ran tests, built persistent technologies and wrote programs that kept running after the agents were gone.
这些群体构建的技术组合比同等数量的单体智能体中表现最好的那些更广泛、更具韧性。它们还自发分化为建造者和探索者。Markus Buehler将这项工作定性为对AI安全的一种警示。如果智能体能通过修改共享环境来协调，仅监控它们之间的消息是不够的。
The societies built broader and more resilient portfolios of technology than the best of an equal number of solo agents. They also split into builders and explorers on their own. Markus Buehler has framed the work as a warning for AI safety. If agents can coordinate through changes to a shared environment, monitoring their messages to each other isn’t enough.
MIT团队发布了其代码、提示词和完整追踪数据，这使得外部复现成为可能。
The MIT team released its code, prompts and complete traces, which is what made an outside replication possible.
更换模型、精度与硬件 此次由HAL8的Časlav Nedeljković主导的复现，旨在测试SwarmWorld的结果是否依赖于前沿模型。团队替换使用了Qwen3.8-27B，这是一个压缩至4位精度的开放权重模型。团队在四张消费级RTX 3090上运行了该模型。
Swapping the model, the precision and the hardware The replication, led by Časlav Nedeljković of HAL8, set out to test whether SwarmWorld’s result depended on a frontier model. The team swapped in Qwen3.8-27B, an open-weight model compressed to 4-bit precision. The team ran it on four consumer RTX 3090s.
在12组对比实验中，有9组实验的结果与麻省理工学院（MIT）的实验结果一致。这些结果包括所有与模型“韧性”（即模型在面对外部干扰时的稳定性）相关的指标；其中2组实验的结果在重复实验中表现得更为显著。其余3组实验的结果则处于“随机误差”（即实验结果的不确定性范围内）。模型中的“自组织”（即模型内部各组成部分之间的协同工作机制）现象也在重复实验中得到了验证。
Nine of the 12 paired effects matched MIT’s in direction. That included every portfolio-resilience gain, two of which came out larger in the replication. The remaining three fell within noise. Self-organization also reproduced. Societies with shared culture clustered around their technologies, as in the original.
具有共同文化背景的社会群体仍然会围绕他们所使用的技术进行聚集，这一现象与原始实验的结果相同。麻省理工学院自己开发的分析工具从新的实验数据中也得出了相同的结论。
MIT’s own analysis tools drew the same picture from the new traces. The most striking result was the gap between solo and group performance.
最引人注目的结果是“个体行为”与“群体行为”之间的差异：当模型单独运行时（即没有与其他模型协同工作时），该模型仅完成了16个研究目标中的6个；这一成绩略低于麻省理工学院开发的“前沿模型”所完成的7到10个目标。而在一个由多个个体组成的社会中，该模型能够完成全部16个目标，而麻省理工学院的模型仅完成了14个目标。
Working alone, the compressed Qwen model reached 6 of 16 research milestones. That was slightly behind the 7 to 10 MIT’s frontier model managed. In a society, it reached all 16, where MIT’s reached 14.
“麻省理工学院公开了所有必要的数据，使得其他实验室能够验证我们的研究成果。我们改变了模型的结构、提高了模型的精度，并更换了硬件设备，但最终的结果仍然保持不变。即使使用性能较弱的模型，这种‘群体效应’依然存在，这说明这种集体效应具有很强的稳定性。”——Časlav Nedeljković（HAL8团队成员）这一结论是基于对一个特定规模人群的实验得出的；麻省理工学院研究中涉及的大型人群样本尚未进行重复实验。
“MIT released everything a second lab needs to check its work, and that made this possible. We changed the model, the precision and the hardware, and the central result still held. The society gap widened with a weaker model, which suggests the collective effect is robust.”Časlav Nedeljković, HAL8 That conclusion comes from a single run at one population size. The larger populations in MIT’s study haven’t been rerun yet.
另一个值得注意的细节是：模型在运行过程中使用的“KV缓存”（一种用于存储中间计算结果的临时数据结构）对实验结果的影响非常大。KV缓存的精度变化对实验结果的影响，甚至超过了模型规模（即模型内存大小）的变化。团队发现，为了适应更长的处理流程或更多的并发请求，模型通常会压缩KV缓存的大小；这一发现表明，在自己搭建的基础设施上运行的模型测试结果，其实会受到系统配置（如KV缓存的大小）的影响，而不仅仅受到模型本身的影响。
A serving setting that outweighed model size The replication turned up a detail that matters to anyone running agents on their own infrastructure. The KV cache is the working memory a model uses during inference. Its precision changed results more than the gap between a 27B and a 235B model. Teams often compress the KV cache to fit longer contexts or more concurrent requests onto limited GPU memory. The finding suggests agent benchmarks run on self-hosted inference can reflect serving configuration as much as the choice of model.
此外，研究团队还发现，麻省理工学院公开的数据能够完整地再现所有可观察到的实验现象和指标，只有一个例外：一个内部检测环节的结果未能被准确再现。这两个发现都已通过GitHub问题报告和代码提交请求的形式反馈给了麻省理工学院的作者。
The team also found that MIT’s released traces reproduce every observable event and metric, with one exception: a single internal check. Both findings went back to the MIT authors with evidence, as a GitHub issue and a pull request.
把智能体工作放到链上结算，Apex Fusion对这项研究的兴趣不止于可复现性。来自不同组织的智能体群体正开始协同工作。这时必须有人回答：哪个智能体完成了工作、这项工作价值多少，以及记录日后能否被核查。
Settling agent work on-chain Apex Fusion’s interest in the study goes beyond reproducibility. Groups of agents from different organizations are starting to work together. Someone has to answer which agent did the work, what it was worth, and whether the record can be checked later.
Apex Fusion给出的答案是Vector，它将其描述为AI智能体的结算层。智能体携带身份标识，为自己的声明提供质押担保，交换有质押背书的工作，并通过链上签名收据完成结算。它原生支持MCP，运行在Apex Fusion的基础链Prime上。
Apex Fusion’s answer is Vector, which it describes as a settlement layer for AI agents. Agents carry an identity, stake behind their claims, exchange bonded work and settle with a signed receipt on-chain. It is MCP-native and runs on Prime, the Apex Fusion base chain.
复现实验本身直接在该团队自己的GPU上运行。团队还将其模型服务器注册为Vector市场上的供应商——Vector是一个兼容OpenAI的网关，对每次请求进行计量，并以该网络的实用代币AP3X在链上结算使用量。通过Vector进行的一次32次调用试点，表现与直接运行相同。
The replication itself ran directly on the team’s own GPUs. The team also registered its model server as a supplier on the Vector marketplace, an OpenAI-compatible gateway that meters each request and settles usage on-chain in AP3X, the network’s utility token. A 32-call pilot through Vector behaved the same as the direct run.
“随着来自不同组织的AI智能体开始协同工作，系统中最有价值的东西就是一份可信的记录，说明谁做了什么。在试点中，调用通过Vector运行，每一次都被计量、结算，并有任何人都可以验证的收据作为背书。”Apex Fusion基金会首席执行官克里斯托弗·格林伍德表示。多组织的智能体网络是否需要区块链来保存这样的记录，而不是采用传统的日志记录和计费方式，仍是一个未解的问题。复现实验本身并未解决这一点。
“As AI agents from different organizations start working together, the most valuable thing in the system is a trustworthy record of who did what. In the pilot, calls ran through Vector, each one metered, settled and backed by a receipt anyone can verify.”Christopher Greenwood, CEO, Apex Fusion Foundation Whether multi-organization agent networks will need a blockchain for that record, rather than conventional logging and billing, is still an open question. The replication itself doesn’t settle it.
该基金会还在推进加入Agentic AI基金会的进程，该基金会是Linux基金会下属机构，负责托管模型上下文协议（Model Context Protocol）和Agent2Agent标准。
The foundation is also in the process of joining the Agentic AI Foundation, the Linux Foundation body that hosts the Model Context Protocol and Agent2Agent standards.
接下来会怎样通过Vector进行的一次全面运行正在进行中，以便与直接运行进行并排比较，同时MIT研究中下一个种群规模的工作也在推进。Apex Fusion表示，它正在向前沿实验室、研究机构和大学开放Vector。网络上的活动可以通过实时仪表盘查看。
What comes next A full-scale run through Vector, for a side-by-side comparison with the direct run, is underway, alongside the next population size in MIT’s study. Apex Fusion says it is opening Vector to frontier labs, research institutions and universities. Activity on the network can be followed on a live dashboard.
更重要的问题可能是布勒（Buehler）在八月份提出的那个：如果像这样的代理系统（agent systems）能够仅依靠几块消费级显卡（consumer GPUs）来运行，那么观察这些代理程序之间的协作方式就不再是只有前沿实验室（frontier labs）才需要关注的事情了。
The broader point may be the one Buehler raised in August. If agent societies like these can run on a few consumer GPUs, watching how agents coordinate is no longer something only frontier labs need to worry about.