AI两天推翻20年工作习惯！Karpathy百行代码开源项目“封神”，AI替你通宵肝研究、战绩可查

Core Insights - The article discusses the autoresearch project launched by Andrej Karpathy, which allows AI agents to autonomously conduct deep learning research without human intervention, achieving significant improvements in training efficiency [2][4][5]. Group 1: Project Overview - The autoresearch project has gained significant attention, with 36.9k stars on GitHub and over 10.6 million views, aiming to create AI agents that can continuously advance research at high speed without human involvement [4][12]. - The project consists of only 630 lines of Python code, where the AI agent modifies code, trains for 5 minutes, evaluates results, and iterates autonomously [7][8]. - The design allows the AI agent to complete approximately 12 experiments per hour, totaling around 100 experiments overnight, with a focus on maintaining comparability across different model modifications [8][9]. Group 2: Performance and Results - In a recent experiment, the autoresearch project successfully reduced the training time for a model from 2.02 hours to 1.80 hours, achieving an 11% improvement in performance [15][16]. - The AI agent autonomously identified about 20 modifications that lowered the model's validation loss, demonstrating the effectiveness of the autonomous tuning process [15][16]. Group 3: Future Aspirations - Karpathy envisions the next step for autoresearch to involve asynchronous large-scale collaboration among multiple agents, simulating a complete research community rather than just a single researcher [12][13]. - The project aims to explore new collaborative models where agents can independently contribute to various research directions and share their findings, potentially revolutionizing the way AI research is conducted [13][17].