# AI-Build-AI

**Repository Path**: ktwu/AI-Build-AI

## Basic Information

- **Project Name**: AI-Build-AI
- **Description**: AIBuildAI Inc  
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-08-19
- **Last Updated**: 2026-09-05

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

<h1 align="center"><small>AIBuildAI – An AI agent that automatically builds AI models</small></h1>

<h1 align="center"><sub>🏆 #1 on OpenAI <a href="https://github.com/openai/mle-bench">MLE-Bench</a></sub></h1>

<p align="center">
  <img src="assets/downloads.svg" alt="Downloads">
  <a href="https://arxiv.org/abs/2604.14455">
  <img alt="AIBuildAI v1 arXiv 2604.14455" src="https://img.shields.io/badge/AIBuildAI%20v1-2604.14455-b31b1b.svg?logo=arxiv&logoColor=white">
  </a>
  <a href="https://arxiv.org/abs/2605.27873">
  <img alt="AIBuildAI v2 arXiv 2605.27873" src="https://img.shields.io/badge/AIBuildAI%20v2-2605.27873-b31b1b.svg?logo=arxiv&logoColor=white">
  </a>
</p>

---

## News

- **[6/17/2026]** The **AIBuildAI Agent 2.5** version is made available to [Max users](https://www.aibuildai.io/#products): a live build dashboard, run replay, cross-run memory, MCP research tools, and DeepSeek support. [Release notes and download](https://github.com/aibuildai/AI-Build-AI/releases/tag/v2.5.1).
- **[5/26/2026]** The **AIBuildAI Agent 2.0** version is made available to [Pro users](https://www.aibuildai.io/#products).
- **[5/1/2026]** In the [TGS Salt Identification Challenge](https://www.kaggle.com/competitions/tgs-salt-identification-challenge) hosted by Kaggle, the model automatically developed by our AIBuildAI Agent ranked in the top 5.7%. Among 3,219 teams composed of human experts, this performance reaches the level of top-tier human AI experts. Model and code developed by the Agent: [tasks/tgs-salt-identification-challenge](https://github.com/aibuildai/AI-Build-AI/tree/main/tasks/tgs-salt-identification-challenge).
- **[4/27/2026]** Excited to announce **AIBuildAI Agent 2.0**! It has once again achieved #1 on OpenAI's MLE-Bench, reaching a score of 70.7% and substantially outperforming the agents ranked 2nd through 5th. Compared to version 1.0, Agent 2.0 introduces several technical advancements, which we will detail in an upcoming technical report. The 2.0 version will be available to Pro users soon.
- **[4/24/2026]** In a [heart disease prediction competition](https://www.kaggle.com/competitions/playground-series-s6e2/overview) hosted by Kaggle, the model automatically developed by our AIBuildAI Agent ranked in the top 6.6%. Among 4,370 teams composed of human experts, this performance reaches the level of top-tier human AI experts. Model and code developed by the agent: [tasks/playground-series-s6e2](https://github.com/aibuildai/AI-Build-AI/tree/main/tasks/playground-series-s6e2).

---

https://github.com/user-attachments/assets/b6043d39-43df-464a-8e25-d24006ba99c8

---


## Introduction

AIBuildAI is an AI agent that automatically builds AI models. Given a task, it runs an agent loop that analyzes the problem, designs models, writes code to implement them, trains them, tunes hyperparameters, evaluates model performance, and iteratively improves the models. By automating the model development workflow, AIBuildAI reduces much of the manual effort required to build AI models.

<p align="center">
  <img src="assets/workflow.png" width="70%" alt="AIBuildAI Architecture">
</p>

---

## Current Results

On OpenAI [MLE-Bench](https://github.com/openai/mle-bench), AIBuildAI ranked #1, demonstrating strong performance on real-world AI model building tasks.

<p align="center">
  <img src="assets/results.png" width="50%" alt="MLE-Bench Results">
</p>

---

## Quick Start

AIBuildAI requires a **Linux x86_64** machine (Ubuntu 20.04 or newer).

There are three versions. **V2.5 (Max)** is the current, most capable version. **V2 (Pro)** and **V1 (free)** remain available.

| Version | Plan | To run it |
|---|---|---|
| **V2.5 (Max)** — current | Max subscription | `aibuildai login` (Max plan) + a Claude Code login or model API key |
| **V2 (Pro)** | Pro subscription | `aibuildai login` (Pro plan) + a Claude Code login or Anthropic API key |
| **V1** | free | a Claude Code login or Anthropic API key (no account) |

Subscriptions are managed at [accounts.aibuildai.io](https://accounts.aibuildai.io); see the account page for the available plans.

### V2.5 (Max) — current

1. **Subscribe.** Create an account at [accounts.aibuildai.io/sign-up](https://accounts.aibuildai.io/sign-up) and switch to the **Max** plan.

2. **Install.**

   ```bash
   curl -fsSL https://github.com/aibuildai/AI-Build-AI/releases/download/v2.5-latest/aibuildai-linux-x86_64.tar.gz | tar xz && ./aibuildai-linux-x86_64-*/install.sh
   ```

3. **Log in** (required before running):

   ```bash
   aibuildai login      # opens a browser to sign in
   aibuildai whoami     # should show an active Max plan
   ```

4. **Sign in to Claude Code or set your API key.**

   If Claude Code is already signed in on this machine, no Anthropic API key
   is needed. AIBuildAI automatically uses the local Claude Code credentials.

   If Claude Code is not signed in, run:

   ```bash
   claude auth login
   ```

   You can also use an Anthropic API key instead:

   ```bash
   export AIBUILDAI_API_KEY=your-anthropic-api-key
   ```

   To use DeepSeek, set a `deepseek-*` model in the config and put your
   DeepSeek API key in `AIBUILDAI_API_KEY`.

5. **Run.** V2.5 is driven by a YAML config:

   ```bash
   aibuildai config > task.yaml    # writes a starter config with every field
   # edit task.yaml: set run.task_name, run.data_root, run.instruction, run.playground_root
   aibuildai run task.yaml
   ```

   Other commands: `aibuildai memorize` (summarize past runs into memory), `aibuildai replay <run-dir>` (replay a finished run), `aibuildai --help`.

### V2 (Pro)

1. **Subscribe** to the **Pro** plan at [accounts.aibuildai.io/sign-up](https://accounts.aibuildai.io/sign-up).

2. **Install.**

   ```bash
   curl -fsSL https://github.com/aibuildai/AI-Build-AI/releases/download/v2.0-latest/aibuildai-linux-x86_64.tar.gz | tar xz && ./aibuildai-linux-x86_64-*/install.sh
   ```

3. **Log in.**

   ```bash
   aibuildai login
   aibuildai whoami     # should show an active Pro plan
   ```

4. **Sign in to Claude Code or set your API key.**

   If Claude Code is already signed in on this machine, no Anthropic API key
   is needed. AIBuildAI automatically uses the local Claude Code credentials.

   If Claude Code is not signed in, run:

   ```bash
   claude auth login
   ```

   You can also use an Anthropic API key instead:

   ```bash
   export ANTHROPIC_API_KEY=your-api-key
   ```

5. **Run.** V2 is driven by command-line flags:

   ```bash
   aibuildai run --task-name <name> --data-dir <path> \
     --playground-dir <path> --instruction "$(cat task.md)" --no-form
   ```

   Or run `aibuildai` with no flags to fill in the parameters in an interactive form.

### V1 (free)

No account or subscription required.

1. **Install.**

   ```bash
   curl -fsSL https://github.com/aibuildai/AI-Build-AI/releases/download/v1.0-latest/aibuildai-linux-x86_64.tar.gz | tar xz && ./aibuildai-linux-x86_64-*/install.sh
   ```

2. **Sign in to Claude Code or set your API key.**

   If Claude Code is already signed in on this machine, no Anthropic API key
   is needed. AIBuildAI automatically uses the local Claude Code credentials.

   If Claude Code is not signed in, run:

   ```bash
   claude auth login
   ```

   You can also use an Anthropic API key instead:

   ```bash
   export ANTHROPIC_API_KEY=your-api-key
   ```

3. **Run** with command-line flags:

   ```bash
   aibuildai --task-name <name> --data-dir <path> \
     --playground-dir <path> --instruction "$(cat task.md)" --no-form
   ```

   Or run `aibuildai` with no flags to fill in the parameters in an interactive form.

**Important:** run the command directly in your terminal. Do not wrap it in a `.sh`/`.bash` script — running it through a script may cause the TUI (Text User Interface) to crash.

### Results

After a run completes, the output directory usually looks like (structure may slightly vary by task):

```
├── candidate_1/  candidate_2/  candidate_3/  # Auto-generated training scripts and model checkpoints
├── checkpoint.pth       # Best model checkpoint
├── inference.py         # Standalone inference script for the final model
├── submission.csv       # Test predictions (if test inputs are provided)
└── progress.pdf         # Visual progress report
```

The main outputs of an AIBuildAI run are the model checkpoints and the script `inference.py`, which runs predictions with the final model on any data. When the task data folder includes unlabeled test inputs, AIBuildAI also writes a predicted-label file `submission.csv`.

### Tasks

We provide ready-to-run task markdowns and datasets in the `tasks/` folder of this repository. You can also write your own task description and point the run at your own dataset.

```bash
git clone https://github.com/aibuildai/AI-Build-AI.git
```

---

## License

This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.

---

## Citation

```bibtex
@article{zhang2026aibuildai,
    title={AIBuildAI: An AI Agent for Automatically Building AI Models},
    author={Ruiyi Zhang and Peijia Qin and Qi Cao and Li Zhang and Pengtao Xie},
    year={2026},
    journal={arXiv},
    url={https://arxiv.org/abs/2604.14455}
}
@article{zhang2026aibuildai2,
    title={AIBuildAI-2: A Knowledge-Enhanced Agent for Automatically Building AI Models},
    author={Ruiyi Zhang and Peijia Qin and Qi Cao and Li Zhang and Pengtao Xie},
    year={2026},
    journal={arXiv},
    url={https://arxiv.org/abs/2605.27873}
}
```