使用本地 LLM 运行 Karpathy 的 autoresearch — 零 API 成本的自主 AI 研究

📄 中文摘要

Andrej Karpathy 发布了 autoresearch,这是一个实验,利用 LLM 自主修改 GPT 训练脚本,进行 5 分钟的实验,保留提升 val_bpb 的结果,丢弃不符合的结果。原始版本使用 Claude Code(云 API)作为研究者。SohniSwatantra 的分支项目用本地运行的 Qwen 3.5 9B 替代了 Claude Code,采用单 GPU 运行,完全自主且无需 API 成本。

📄 English Summary

Running Karpathy's autoresearch with Local LLM — Zero API Cost Autonomous AI Research

Andrej Karpathy released autoresearch, an experiment where an LLM autonomously modifies a GPT training script, conducts 5-minute experiments, retains improvements in val_bpb, and discards the rest. The original version utilizes Claude Code (cloud API) as the researcher. SohniSwatantra's fork replaces Claude Code with Qwen 3.5 9B running locally, leveraging a single GPU for fully autonomous operation with zero API costs.

Powered by Cloudflare Workers + Payload CMS + Claude 3.5

数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等