autoresearch
AI agents running research on single-GPU nanochat training automatically
freevoice-agents
62.2k
Stars
+29588
Stars/month
0
Releases (6m)
Star Growth
+4.8k (8.2%)
Overview
autoresearch 是一个创新的自主AI研究工具，让AI代理在单GPU环境下自动进行LLM训练实验。该工具基于简化的nanochat实现，允许AI代理在夜间自主运行：修改训练代码、进行5分钟训练、评估结果改进、保留或丢弃更改，然后重复这个过程。你醒来时会看到完整的实验日志和（希望）改进的模型。核心理念是通过编程`program.md`文件来指导AI代理，而不是直接修改Python代码。代理可以自由修改架构、超参数、优化器、批量大小等所有训练要素。系统使用固定的5分钟时间预算和val_bpb（验证每字节比特数）作为评估指标，确保不同架构变更之间的公平比较。这代表了自主AI研究的新范式，where研究不再需要人类持续监督。
Deep Analysis
Key Differentiator
Karpathy's pioneering concept of AI agents autonomously running ML experiments overnight — vs traditional hyperparameter search tools that don't modify architecture or code
⚡ Capabilities
- • Autonomous AI-driven LLM training experiments
- • Automated code modification, training, evaluation loop
- • Fixed 5-minute time-budgeted experiments for fair comparison
- • Experiment logging with val_bpb metric tracking
- • Agent-editable single-file training setup (train.py)
- • Markdown-based agent programming (program.md)
🔗 Integrations
Claude CodeCodexPyTorchNVIDIA CUDA
✓ Best For
- ✓ Researchers exploring autonomous ML experiment iteration
- ✓ Learning about AI-driven research automation
- ✓ Overnight autonomous hyperparameter/architecture search
✗ Not Ideal For
- ✗ Production model training or deployment
- ✗ Teams without GPU access
Languages
Python
Deployment
Local single-GPU setupuv package manager
Pricing Detail
Free: Open source MIT, free to run locally
Paid: N/A — requires GPU hardware + AI agent API costs
⚠ Known Limitations
- ⚠ Requires NVIDIA GPU (tested on H100, works on smaller GPUs with forks)
- ⚠ Single-GPU only, no distributed training
- ⚠ Narrow scope: only LLM training optimization experiments
- ⚠ Results not comparable across different hardware
Pros
- + 完全自主的夜间实验能力，无需人工干预即可进行数百次训练迭代
- + 简洁的三文件架构设计，降低复杂性同时保持实验灵活性
- + 固定时间预算确保不同实验配置之间的公平比较和评估
Cons
- - 限制为单GPU环境，无法扩展到大规模分布式训练
- - 5分钟的固定训练窗口可能限制复杂模型或大数据集的充分训练
- - 需要NVIDIA GPU硬件支持，增加了使用门槛
Use Cases
- • 自动超参数调优，让AI代理探索最佳学习率、批量大小和优化器设置
- • 神经网络架构搜索，自主试验不同的模型设计和层配置
- • 夜间无人值守的研究实验，充分利用计算资源进行持续优化
Getting Started
1. 准备单个NVIDIA GPU环境并克隆仓库；2. 编辑program.md文件设置AI代理的研究指令和目标；3. 启动自主代理循环，让其开始自动实验和迭代训练过程