Core Insights - The article discusses the evolution of AI research from human-driven efforts to autonomous AI agents capable of conducting research independently, marking a significant shift in the field [1][5]. Group 1: AI Research Evolution - AI research has transitioned from human researchers to autonomous AI agents, with the current generation being the 10,205th iteration of this technology [1]. - A system named FARS has been developed, which produces a research paper approximately every two hours, generating 244 research hypotheses and 100 short papers during the Spring Festival [1]. Group 2: Autonomy in AI Research - The project "autoresearch" allows AI agents to iteratively modify their training code based on human-provided prompts, potentially unlocking self-iteration capabilities for AI [2][3]. - The core idea of the autoresearch project is to provide a small but functional LLM training environment for AI agents to conduct experiments autonomously overnight, leading to improved model performance [5]. Group 3: New Research Paradigms - A new research paradigm is emerging where AI conducts experiments while humans design the research systems, indicating a shift from competition based on models and data to competition based on "research organization code" [9]. - The training code for this project is derived from the previously open-sourced nanochat, which serves as a simplified framework for large model training [10]. Group 4: Training Efficiency - Nanochat can now train a GPT-2 capability model on a single 8XH100 node in just 2 hours, reducing the training time by approximately 3 hours compared to a month ago [13].