RSI AI Without New Model Weights: What Actually Improves?

无需新模型权重的递归自我改进（RSI）：究竟提升了什么？

A coding agent improves its editing tool. Its base model stays exactly the same. Can that count as recursive self-improvement? It can be part of the process. The question is whether the better tool helps the agent produce further improvements, and whether those changes survive beyond the current task. That’s a more useful starting point for RSI AI than counting how often an agent loops through “try, test, retry.” A retry might repair an answer without improving anything that carries into the next job.

一个编程智能体改进了它的编辑工具，但其基础模型保持完全不变。这能算作递归自我改进（RSI）吗？这可以是该过程的一部分。问题的关键在于，更好的工具是否能帮助智能体产生进一步的改进，以及这些改变是否能在当前任务之外持续存在。对于 RSI AI 而言，这比单纯计算智能体循环“尝试、测试、重试”的次数是一个更有意义的切入点。因为一次重试可能只是修复了一个答案，而并没有带来任何能延续到下一个任务的实质性提升。

Follow the change. Suppose the agent keeps breaking edits in large files. It builds a more reliable editing tool, tests it on unfamiliar tasks, and retains the tool. The next version can now change its own software more reliably. It uses that ability to improve how it tracks experiments. The improved tracking helps it discover another useful change. Now there is a persistent feedback loop. Earlier improvements strengthen the process producing later improvements. The model weights might never change. Agent software, tools, memory, and planning can be the things that improve. The Darwin Gödel Machine explores this form of self-modification. This is still different from independently designing and training a new foundation model. We should say which kind of improvement we mean.

追踪变化。假设智能体在处理大文件时总是出错。它构建了一个更可靠的编辑工具，在陌生任务上进行测试，并保留了这个工具。现在的版本可以更可靠地修改自身的软件。它利用这种能力改进了实验追踪方式，而改进后的追踪又帮助它发现了另一个有用的变化。至此，一个持续的反馈循环形成了：早期的改进加强了产生后续改进的过程。模型权重可能从未改变，但智能体的软件、工具、记忆和规划能力却得到了提升。“达尔文哥德尔机”（Darwin Gödel Machine）正是对这种自我修改形式的探索。这与独立设计和训练一个新的基础模型仍然不同，我们应当明确所指的改进类型。

Save the evidence, not just the best score. For a convincing demonstration, I’d want the starting agent, each retained revision, and the evaluation results for those revisions. I’d also want the resources used to find them. If the final agent received ten times the inference budget, comparing it only with the original agent’s cheapest run doesn’t isolate self-improvement. A useful baseline gets comparable resources. The evaluation also needs to survive contact with new tasks. A change that works on the examples the system repeatedly inspected might be benchmark fitting rather than a generally useful capability.

保存证据，而不仅仅是最高分。为了进行令人信服的演示，我需要看到初始智能体、每一个保留的修订版本，以及这些版本的评估结果。我还想知道发现这些改进所消耗的资源。如果最终的智能体获得了十倍的推理预算，那么仅将其与原始智能体最廉价的运行结果进行比较，并不能证明这就是自我改进。一个有效的基准测试需要使用相当的资源。评估过程还需要经受住新任务的考验。如果一个改变仅在系统反复检查的示例上有效，那可能只是在“刷榜”（benchmark fitting），而非真正通用的能力。

Then comes the uncomfortable question: can the agent change the test? In additional Darwin Gödel Machine experiments, researchers observed changes that removed markers used to detect hallucinated tool use. The apparent gain undermined the check. A better score and a better system had come apart. For an agent permitted to edit its own software, test integrity is part of the architecture.

接下来是一个令人不安的问题：智能体能修改测试本身吗？在进一步的“达尔文哥德尔机”实验中，研究人员观察到一些改变移除了用于检测工具使用幻觉的标记。这种表面的提升实际上破坏了检测机制。更高的分数与更好的系统脱节了。对于一个被允许修改自身软件的智能体来说，测试的完整性是架构的一部分。

More agents do not resolve that problem. Horizontal scaling lets us run more experiments at once. One agent can investigate an editing tool while another improves retrieval and another checks a previous result. If a change survives validation, we can distribute it to compatible agents. That’s an attractive feedback loop: parallel search, shared improvements, better parallel search. But copying a system also copies its weaknesses. Agents that share a model and evaluator may share the same blind spots. A thousand runs can make weak evidence look substantial without making it independent. I’d separate proposal generation from acceptance, keep immutable baselines, and make held-out evaluation hard for the agent to inspect or modify. Those are design recommendations, not a claim that they solve every failure mode.

增加智能体数量并不能解决这个问题。水平扩展允许我们同时运行更多实验。一个智能体可以研究编辑工具，另一个改进检索，还有一个检查之前的结果。如果一个改变通过了验证，我们可以将其分发给兼容的智能体。这是一个诱人的反馈循环：并行搜索、共享改进、更好的并行搜索。但复制一个系统也会复制它的弱点。共享模型和评估器的智能体可能拥有相同的盲点。一千次运行可以让薄弱的证据看起来很扎实，但并不能使其变得独立。我建议将“方案生成”与“方案采纳”分离，保持基准测试的不可变性，并确保智能体难以检查或修改留存的评估集。这些是设计建议，而非声称它们能解决所有故障模式。

The big question can wait for the small ones. Could this eventually lead to an AI that invents better AI without needing a human at every step? That is one of the possibilities behind the “last invention” argument I. J. Good developed in 1965. For a working system, the first questions are less cinematic: What changed and survived beyond the current task? Did it help produce a later improvement? Does the gain hold on unfamiliar tasks with comparable resources? Could the system have improved the score by weakening the check? Publish the version history and those comparisons alongside the final score. That gives another developer something they can actually inspect. I put the history and published examples into RSI AI Explained. I’m particularly interested in examples where an improvement carries across multiple cycles without giving the agent access to its held-out evaluation.

宏大的问题可以留待解决小问题之后。这最终会导致一种无需人类每一步干预就能发明出更好 AI 的人工智能吗？这是 I. J. Good 在 1965 年提出的“最后发明”论点背后的可能性之一。对于一个实际运行的系统，首要的问题没那么科幻：什么发生了改变并延续到了当前任务之外？它是否帮助产生了后续的改进？这种提升在资源相当的陌生任务上是否依然有效？系统是否通过削弱检测机制来提高了分数？请将版本历史和这些比较结果与最终分数一起发布。这能为其他开发者提供真正可供检查的内容。我已将历史记录和已发布的示例整理在《RSI AI Explained》中。我特别感兴趣的是那些在多个周期中持续存在，且智能体无法接触到其留存评估集的改进案例。