# Nathan Lambert：人类仍深嵌 AI 研究，RSI 尚未到来

- 来源：Nathan Lambert (@natolambert)
- 发布时间：2026-09-12 19:50
- AIHOT 分数：26
- AIHOT 链接：https://aihot.news/items/cmtybyh4206o5rojh27skys1v
- 原文链接：https://x.com/natolambert/status/2098741127398482387

## AI 摘要

Nathan Lambert 认为，前沿实验室里构建 LLM 的人仍在高强度、窄任务上拼命工作，这说明我们可能只是"接近"而非处于 RSI（递归自我改进）。若真进入 RSI，人类投入应逐步减少；只要人类仍深度嵌入研究流程，最极端的 RSI 形态就不会出现，实验室人员的倦怠程度可作为判断代理指标。他因此认为当前是 AI 研究自动化带来的加速，但并非立即爆发式。

## 正文

What is the expectations for how long human effort will last in the center of AI research tasks if we are in the early stages of RSI?

One of my observations that makes me think we may be *approaching* a form of RSI, rather than in it, is how everyone who builds LLMs at a frontier lab is in a total grindfest. They work SO hard and often on fairly narrow tasks that accumulate to the model.

In an era of RSI, I would expect the human effort to progressively decrease. Is the contention that there was so much work to be done for humans AND AI assistants, that AI's are doing so much more and humans are doing the same amount -- both with more things they could do?

Seems like so long as humans are that deeply embedded, the most extreme forms of RSI aren't on the table, and the burnout of our mutuals at labs is a fairly easy proxy to read, if they can't share more technical details.

Thoughts? This is one of a few reasons I think we're in an acceleration by AI research automation, but one that isn't immediately explosive.