|In * * the prompt with very conservative parameters. This prompts demonstrates that GPT-3 can sum list of numbers, i.e. do iterative reasoning.|
One might wonder, why is all this fluffy stuff with variable names necessary?
Here is the story how I came up with the prompt above.
I first tried to do the same without these variables, and it fails, and I think I know why.
*GPT-3 can sum lists of up to 10 numbers.
Problem: Simulate the execution of 87+18+95+59+13+30+35+25+40+36
Let a=87, b=18, c=95, d=59, e=13, f=30, g=35, h=25, i=40, j=36
Problem: Simulate the execution of
Let a=5, b=74, c=84, d=29, e=59, f=41, g=25, h=1, i=15, j=38
It fails to add the penultimate element (the 9th element of a list of 10 elements), constantly. When I click it, I see its probability was 40% and the probability of last was 60% roughly. It's very similar to how we humans reasons when we are told a list of 10 numbers. We remember the first ones (a sequence) the last one (because being the last is an event), but we might forget about some elements in between, or think we have to finish the computation quickly.
I thought, how do we humans augment our memory abilities, like to remember 100 items? From experimenting it myself, I knew that associative memories augment the capacity to remember things. For example, to remember 100 items, I can think of walking through a place of 100 rooms I know, and imagine an item in each room.
What would be the equivalent for GPT-3? Associate the numbers to add with something else. I thought, let's put variable names.