AI 主导开发出 27B 工业编码模型 iCoder，成绩超同尺寸基线
AI 已经能主导训练出高性能工业编码模型，但人类仍把控目标和验证环节，离完全自主递归改进尚有距离。
上海交大等团队让 AI Agent 主导训练了一个 27B 参数的工业编码模型 iCoder-27B。Agent 负责选择实验、诊断失败并调整策略，人类只设定目标和验证标准。在 RTL（寄存器传输级硬件描述）和 GPU Kernel 等任务上，该模型表现超越同尺寸 Qwen 基线，为递归自我改进提供了实例。
正文摘录
Can AI Build Better AI? The Debate Is Over — Someone Actually Did It Can AI build a better AI? Let's cut the speculation — a team just did it, and the results are in. This is the story of how AI took the helm in developing a production-grade model, and the hard-won lessons hidden in the training failures. This might be the most sci-fi phrase in the AI world: Recursive Self-Improvement (RSI). At its most powerful, you simply tell a system "make a better version of yourself," and it proposes new ideas, sets up training pipelines, verifies results, and delivers a more capable successor. Then, the new model repeats the process, and intelligence enters a self-accelerating loop. This vision hasn't become reality yet. But over the past few months, discussions about RSI have shifted from "will it …