Back to News

AI Agent Autonomously Develops 27B Industrial Coding Model iCoder-27B

#ai-agent#coding-model#recursive-self-improvement#industrial-coding

A joint team from Shanghai Jiao Tong University, DP Technology, and other institutions used an AI agent (Codex GPT-5.6-Sol) to autonomously develop a 27B industrial coding model named iCoder-27B. The agent handled data evolution, SFT cold start, self-distillation, and RL training, while humans provided goals, permissions, SOPs, and validators. The model ranks among top industrial coding models, contributing to the recursive self-improvement debate.

Coverage timeline

  1. 机器之心机器之心
    AI到底能不能自己造AI？别吵了，有人做出来了

    RSI，可能是 AI 世界里最像科幻的一个词。 在最强的版本里，你只需要对一套系统说一句 “做一个更好的自己”，它就能提出新想法、搭建训练流程、验证结果，最终交付一个能力更强的继任者。随后，新模型再次重复这个过程，智能由此进入自我加速的循环。 这个图景还没有成为现实。但过去几个月，围绕递归自我改进（Recursive Self-Improvement，RSI）的讨论，已经从 “会不会发生” 转向一个更具体的问题： 今天的 AI，究竟已经接管了 AI 研发循环中的哪些部分？ Anthropic 联合创始人 Jack Clark 在今年 5 月给出过一个激进判断：到 2028 年底，无人参与的 AI 研发有 60% 以上概率出现；他同时承认，近期更可能先看到非前沿模型层面的端到端训练样例。Allen Institute for AI 的 Nathan Lambert 则提出 “有损自我改进”：AI 会进入模型研发核心，但数据、算力、组织协调和系统复杂性会让飞轮持续损耗，很难形成无摩擦的指数爆发。 现在， 一支来自上海交通大学人工智能学院、深势科技、无尽前沿与新加坡国立大学的联合团队，给出了一个很适合放进这场争论的样本 。 团队让 Agent（Codex GPT-5.6-Sol）主导一个 27B 工业 Coding 模型的开发：从数据演化、SFT 冷启动，到同策略自蒸馏和可验证奖励强化学习，Agent 负责选择实验、诊断失败，并反复修改训练策略。同时，人类专家提供目标、权限边界、训练 SOP 和可信验证器。最终得到了一个在工业 Coding 能力排名前列的模型，名为 iCoder-27B 。 论文标题：Recursive AI-Led Development of Frontier Industrial Coding Model 项目主页：liunanfu1992.github.io/iCoder-site 论文: huggingface.co/i-Coder/iCoder-27B/blob/main/Coder_Tech_Report.pdf GitHub：github.com/bingreeky/iCoder Hugging Face：huggingface.co/i-Coder 27B，先挤进工业 Coding 第一梯队 工业 Coding 与日常所说的代码生成，并

Related stories

  • Naive AI open-sources first model built with AI-led R&D

    Naive AI, founded by Dai Jifeng, has open-sourced its first model, developed from day one with AI models handling coding, experiments, and analysis while humans set goals and constraints. The company focuses on post-training and agent development based on open-source models, echoing recent disclosures from Anthropic and OpenAI about AI-led research proportions.

    1 reports · 1 sources

  • 7 PhD students train 7B model from scratch in 3 months, open-sourcing full pipeline

    Seven PhD students at Beijing Zhongguancun Academy trained the 7B model ZGCM-1 from scratch over a summer, releasing training data, recipes, weights, code, checkpoints, and logs. They employed hundreds of AI agents to handle data processing, experiments, log review, and evaluation, practicing an 'AI4AI' development paradigm. In general benchmarks, ZGCM-1 performs comparably to models like Qwen3-8B, and in some math reasoning and search tasks it rivals larger models such as Qwen3-235B-A22B and GLM-5.1.

    1 reports · 1 sources

  • Researchers: Suspected Chinese hackers used open-source AI agents to build autonomous hacking tool

    According to a Financial Times report, researchers say suspected Chinese hackers used open-source AI agents to develop an autonomous hacking tool that compromised Taiwanese government websites in July. The AI agents reportedly conducted simultaneous reconnaissance and break-ins, marking a new phase in cyberwarfare.

    1 reports · 1 sources

  • Apple trains China-specific LLM with Alibaba support, sources say

    Reuters reports, citing three sources, that Apple has trained a large language model specifically for the China market with support from Alibaba. If confirmed, Apple would become the first foreign company to offer a proprietary AI model in China, marking a departure from its previous strategy of using domestic Chinese models.

    2 reports · 2 sources

News updates every hour. Subscribe in your RSS reader to get new stories automatically.

Subscribe