Speed-up calculator: How much will automating AI R&D speed up AI software progress, absent a software intelligence explosion?

We’ve previously argued that once AI R&D is automated, the feedback loop of AI improving AI might be powerful enough to sustain accelerating progress, even holding compute fixed. We call this possibility a software intelligence explosion (SIE).

But, as Ryan Greenblatt points out, automating AI R&D will speed up AI progress even if there’s no SIE.

We’ve created a tool for calculating how much faster. The calculator assumes that training and inference compute continues to grow at a constant exponential rate after automation.

Two factors drive the speed-up, according to this calculator:

1. Faster researcher growth. Post-automation, the quality and quantity of researchers will grow faster than today due to exponential compute growth. The tool captures this via the following parameters:

   1. g_L gives the growth rate of human AI researchers today, the baseline for the comparison.

   2. g_C gives the growth rate of compute (both today and after AI R&D automation). This is proportional to the number of automated researchers that can be run post automation.

   3. γ gives the rate at which more training compute transfers to more productive researchers. The productivity of a researcher is proportional to training_compute^γ.

   4. So the old rate of researcher growth is g_L; the new rate is γ × g_C. The impact of this faster growth is mediated by α, the labour share of AI software R&D.

2. The new feedback loop. The classic software feedback loop (better AI → more software R&D → better AI) can significantly increase the pace of progress even if it does not drive accelerating progress.

   1. The condition for accelerating progress is r_cog > 1.

   2. When r_cog < 1, the speed-up from this feedback loop is proportional to a 1 / (1 - r_cog).

My (Tom Davidson’s) quick best-guess inputs were:​

This resulted in a 3.3x speed-up in software progress. If software accounts for half of total AI progress (with increasing compute responsible for the other half), then the total AI progress speeds up by 0.5 + 0.5 × 3.3 = 2.1x. ​

Try the calculator yourself!