I'm interested in estimating how many 'OOMs of compute' span the human range. There are a lot of embedded assumptions there, but let's go with them for the sake of a thought experiment.

Cortical neuron counts in humans have a standard deviation of 10 - 17%, depending on which source you use. Neuron counts are a useful concrete anchor that I can relate to AI models. 

There are many other factors that account for intelligence variation among humans. I'd like to construct a toy model where those other factors are backed out. Put another way - if intelligence variation was entirely explained by neuron count differences, how much larger would the standard deviation of the neuron counts have to be to reproduce the same distribution we observe in intelligence?

From the literature, about 5-15% of the variance in intelligence test performance is attributable to intercranial volume differences. Intercranial volume differences also appear to be a reasonably close proxy for neuron count differences.

To be conservative, let's take the lower end (5%) as the variance in intelligence attributable to volume differences. The conclusions aren't sensitive to what you pick here.

Working this through, a neuron count standard deviation 4.47x larger would produce the same distribution, with the other sources of variaton removed.

Let's take the largest estimate of cortical neuron count standard deviation (17%) and inflate it by this multiplier. So our new standard deviation for "effective neuron count" is 76%

Next I want to estimate the gap between a median human and a world-historical genius. Let's take a population of 10b humans. Assuming Gaussianity, the maximum Z value you'd expect to observe in this sample is about 6.5.

So the maximum 'effective neuron count' for this hypothetical individual would be 1 + 0.76*6.5 = 5.9x

So roughly 6x "effective parameter count" spans the range from human median to world-historical genius. That's not very large.

There's no direct way to translate that into "OOMs of compute". But for what it's worth: for a Transformer, 6x the parameter count needs ~36x more compute for a loss-minimizing model. So we could say that if human brain scales on the same basis (probably wrong), rare human outliers would be equivalent to a model trained with 1.5 OOMs more FLOPs than baseline. That's slightly less than the gap from GPT-3 to GPT-4.

This is a toy model, it's easy to poke holes in, but I at least found the exercise interesting. It feels plausible, and would imply timelines from AGI to ASI of at most a few years on current trends.