AGI Timelines Model (METR)

This Streamlit app allows you to interactively forecast AGI timelines using the METR model, based on the paper "Measuring AI Ability to Complete Long Tasks" and insights from Forecaster Reacts to METR's bombshell.

Features

• Interactive Parameter Controls:

  • Set start task length, AGI task length, doubling time, acceleration, and shift.
  • Choose from preset models (e.g., o3, Claude, GPT-4o, DeepSeek, etc.) for start task length.
  • Advanced mode unlocks additional parameters: elicitation boost, reliability, task complexity, and reference date.
  • Select from preset doubling time models or use a custom value.

• Results Visualization:

  • Histogram of AGI arrival years.
  • Exponential growth plot of task length over time, showing:
    • Median, 10% earliest, and 90% latest curves (with AGI achievement markers).
  • Computed AGI dates for median, 10% earliest, and 90% latest scenarios, all in long date format.
  • Yearly probability table for AGI arrival.

• Robust Error Handling:

  • Handles NaN values and outliers in model results.
  • Displays warnings and fallback results if model execution fails.

Setup Instructions

1. Clone the Repository

git clone <repo-url>
cd agi_timelines

2. Install Dependencies

It is recommended to use a virtual environment:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Required packages include:

• streamlit
• squigglepy
• numpy
• matplotlib

2. Install dependencies using Poetry:

   # Install Poetry if you don't have it
   pip install poetry
   
   # Install dependencies
   poetry install
   

3. Run Jupyter notebook:

   poetry run jupyter notebook
   

If you don't have a requirements.txt, create one with:

streamlit
squigglepy
numpy
matplotlib

3. Run the App

streamlit run streamlit_app.py

The app will open in your browser at http://localhost:8501.

Usage

• Adjust parameters in the sidebar to explore different AGI timeline scenarios.
• Use "Advanced Mode" for more detailed control and to select reference models or adjust reliability, elicitation, and task complexity.
• View the results in the main panel:
  • Histogram of AGI arrival years
  • Exponential growth plot (with three percentile curves)
  • Computed AGI dates for median, 10% earliest, and 90% latest
  • Yearly probability table

Notes

• The app uses 100,000 samples for model runs by default for robust results.
• The exponential growth plot is a simplified visualization and may not reflect all stochasticity in the full model, but is parameterized to match the sidebar settings.
• All results update live as you change parameters and rerun the model.

References

• Measuring AI Ability to Complete Long Tasks (METR paper)
• Forecaster Reacts to METR's bombshell

---

For questions or issues, please open an issue on GitHub or contact the project maintainer.