AI递归自我改进的早期进展：GPT-5.3-Codex、AlphaEvolve等系统初露锋芒
英文摘要
Over the past year, several projects have taken early steps toward recursive self-improvement (RSI), where AI systems help design better AI. OpenAI's GPT-5.3-Codex, released in February 2026, deeply participates in its own training debugging, deployment management, and evaluation analysis. Anthropic disclosed that most of its internal code is now written by its Claude Code tool. Google DeepMind's AlphaEvolve (2025) leverages large models to explore and optimize algorithms for chip design and neural networks. The Darwin Gödel Machines project (2025) enables AI agents to modify their own code via evolutionary algorithms. An 'AI Scientist' system, published in Nature (March 2026), automates the entire research pipeline—idea generation, experimentation, paper writing, and peer review. While these systems mark progress, experts stress that full RSI—where AIs autonomously invent, evaluate, and improve without human guidance—remains out of reach, and human oversight is still required.
中文摘要
过去一年，多项AI系统在递归自我改进（RSI）方面迈出了早期步伐。OpenAI于2026年2月推出的GPT-5.3-Codex能深度参与自身训练的调试、部署管理和评估分析；Anthropic透露其内部大部分代码由Claude Code编写。谷歌DeepMind的AlphaEvolve（2025年）利用大模型探索优化芯片设计和神经网络算法。2025年发布的达尔文-哥德尔机项目使AI智能体能够通过进化算法修改自身代码。2026年3月《自然》报道的‘AI科学家’系统则实现了从构思想法、运行实验、撰写论文到同行评审的全流程自动化。尽管如此，专家认为完全自主的RSI尚未实现，人类指引与核验仍不可或缺。
关键要点
OpenAI's GPT-5.3-Codex (Feb 2026) actively participates in its own training debugging, deployment management, and evaluation analysis.
OpenAI的GPT-5.3-Codex（2026年2月）深度参与自身训练调试、部署管理和评估分析。
Anthropic says that most of its internal code is now written by its Claude Code tool.
Anthropic表示其内部大多数代码现由Claude Code编写。
Google DeepMind's AlphaEvolve (2025) uses large models to automatically optimize algorithms for chip design and neural network improvements.
谷歌DeepMind的AlphaEvolve（2025年）利用大模型自动优化芯片设计和神经网络改良算法。
The Darwin Gödel Machines project (2025) allows AI agents to modify their own code through evolutionary algorithms, though the underlying LLM remains fixed.
达尔文-哥德尔机项目（2025年）使AI智能体能够通过进化算法修改自身代码，但底层大模型保持不变。
An 'AI Scientist' system (Nature, March 2026) automates the full research cycle: generating ideas, running experiments, writing papers, and performing peer review.
‘AI科学家’系统（2026年3月《自然》报道）能自动化完成从构思课题、实验、撰写论文到同行评审的完整科研流程。
Startup Ricursive Intelligence, founded by former DeepMind chip design leads, aims to compress chip design cycles from 1-2 years to days through recursive AI-driven improvements.
由前DeepMind芯片设计负责人创立的递归智能公司，计划通过递归AI驱动改进将芯片设计周期从1-2年压缩至数天。