Simulation of Accelerated AI Progress

A software intelligence explosion simulator

This tool is a simulation companion to How quick and big would a software intelligence explosion be? Alongside the ability to change parameterizations (in the sidebar), users have three main options:

  • Multiple Simulations: Run several simulations with uncertainty over key parameters. The output is a table showing the fraction of simulations where the average pace of software progress exceeds a specified threshold over some number of years.
  • Single Simulation: Run a single simulation under specific parameter values to illustrate the path of AI progress, including the evolution of diminishing research productivity and growth rates over time.
  • Speed-up calculator for when rcog<1: This calculator calculates how many times faster is the (balanced) growth path of software once ASARA is achieved if there is no SIE. This calculator explicitly considers a exponentially growing compute stock and constant value for rcog<1.

Results

Model Parameters and Estimates

The table below summarizes the model parameters:

Parameter Description Low Estimate Median Estimate High Estimate
Initial Speed Up (f) After AI Systems for AI R&D Automation (ASARA) is deployed, how many times faster does software progress become (compared to the recent pace of software progress)? 2832
Returns to Software R&D (rcog) Controls the degree of diminishing returns to research. Each time cognitive inputs to software R&D double, how many times does software double? (Note this parameter falls over time.) 0.41.23.6
Distance to effective limits on software At the start of the simulation, how far is software from effective limits? Measured in the years of AI progress at recent rates of progress. 61116
Diminishing returns to parallel labour (p) If cognitive inputs to software R&D instantaneously double, how many times does the pace of software progress double? 0.150.30.6
Fraction of AI progress due to better software. What fraction of observed progress in AI is due to software progress (rather than hardware improvements)? -0.5-

In addition to model parameters, you can select whether to enable additional model specifications:

  • Retraining Cost: Reduce the degree of acceleration as some software efficiency gains are spent making training happen more quickly.
  • Gradual Boost: The initial speed-up from ASARA ramps up gradually over 5 years.
  • Constant Diminishing Returns: Assumes that rcog is fixed at its initial value over time.

Sampling

When 'Multiple Simulations' is selected, randomization occurs over log-uniform distributions for f, rcog,0, and p, while the years until the ceiling are randomized over a uniform distribution. The bounds for these distributions come from the sidebar inputs. Note that we don't sample over the fraction of AI progress due to software progress, but instead leave this as a deterministic parameter.

Defining the inputs rcog and p

Consider the semi-endogenous software progress function:

Software Growth = (Lα C1−α)λ Sβ

where:

  • Research investment combines human labor (L) and compute (C)
  • L represents human labor input
  • C represents computational resources
  • S represents the existing software stock
  • α, λ, and β are parameters

From these parameters, we define our two key model inputs:

  • rcog = λα/β
  • p = λα

Additional details for this derivation as well as discussion of objections to our choice of model inputs are provided in the appendix of the post.