/ftm

Work in progress! I don't recommend looking at the code right now.

Primary LanguageHTML

Full Takeoff Model

Work in progress! I don't recommend looking at the code right now. It needs to be cleaned up.

For now, you might be interested in the core file.

Instructions to execute the model:

  • Clone the github repo
  • Set up the environment (this project uses Poetry) with poetry install
  • Create a copy of the Task_based_model_inputs and set the parameters as you want, and make it public
  • Run poetry run python -m ftm.core.model -w YOUR_SHEET_URL to run the model with best_guess parameters
  • Run poetry run python -m ftm.analysis.exploration_analysis -w YOUR_SHEET_URL to run a comparison of the model with aggressive, best_guess and conservative parameter choices
    • You can run poetry run python -m ftm.analysis.exploration_analysis -w YOUR_SHEET_URL -t PARAMETER_NAME instead to produce a detailed comparison of what happens when you set said parameter at aggressive, best_guess and conservative, with all other parameters held at best_guess value
  • Run poetry run python -m ftm.analysis.sensitivity_analysis -w YOUR_SHEET_URL to see a high level comparison of what happens when you change each parameter between their aggressive, best_guess and conservative values, having all other parameters fixed at their best_guess value
    • The parameters will appear ordered by most sensitive to least sensitive. Concretely, the difference between the takeoff_length of the conservative and aggressive value, for a complex definition of takeoff_length
  • Run poetry run python -m ftm.analysis.mc_analysis -w YOUR_SHEET_URL to run a MC sampling
    • The aggressive, best guess and conservative values correspond to percentiles 5%, 50% and 95% of a distribution
    • You can adjust the correlation between parameters using the rank_correlation_between_buckets tab in the sheet
  • Run poetry run python -m ftm.analysis.timelines_report -w YOUR_SHEET_URL to run nine scenarios corresponding to the conservative, best_guess and aggresive choices, conditioned on short, best_guess and long AI timelines
    • The full_automation_requirement and flop_gap parameters are governed by the sheet Guess FLOP gap and timelines
  • Run poetry run python -m ftm.analysis.megareport -w YOUR_SHEET_URL to run the three previous analysis at once