卡帕西开源Agent自进化训练框架，5分钟一轮实验，48h内揽星9.5k

Core Insights - The article discusses a new open-source project called autoresearch, developed by Karpathy, which enables AI to autonomously conduct research by following instructions written in Markdown documents [2][5]. - The framework is designed to be lightweight, consisting of only 630 lines of code, and can run on a single GPU [3][16]. - The project has gained significant attention, achieving over 9.5k stars on GitHub within two days of release [6]. Project Overview - Autoresearch automates the AI training loop, allowing the AI to modify code, run short experiments lasting five minutes, and evaluate results to determine the next steps [13][14]. - The system operates under two main rules: each experiment has a fixed training time of five minutes, and it evaluates based on the val_bpb metric, where lower values indicate better model performance [15]. Technical Structure - The project relies on three core files: prepare.py for setting constants and tools, train.py for AI modifications, and program.md for human-written instructions [20][24]. - The AI modifies train.py based on instructions from program.md, runs experiments, and decides whether to keep or discard changes based on performance metrics [30][32]. Future Aspirations - Karpathy envisions a future where thousands of AI agents collaborate asynchronously across numerous branches, enhancing research efficiency through collective intelligence [5][35]. - He draws parallels to the SETI@home project, aiming to create a decentralized, distributed exploration model for AI research [38][41]. Research Methodology - The autoresearch process involves AI iterating through modifications, training, evaluation, and decision-making, achieving a high efficiency that surpasses human capabilities [29][32]. - The project aims to shift the research paradigm from a linear, centralized approach to a more flexible, experience-based model that accommodates diverse research paths [49].