We’ve been interested in looking at how AI can process things other than silly images. That’s why the “Free AI Bot that Generates the Excel Formula for Any Problem” caught our eye. Based on GPT-3, it supposedly transforms your problem description into a formula suitable for Excel or Google Sheets.
Our first prompt didn’t work out very well. But that was sort of our fault. When they say “Excel formula” they mean that quite literally. So trying to describe the actual result you want in terms of columns or rows seems to be beyond it. Not realizing that, we asked:
If the sum of column H is greater than 50, multiply column A by 0.33
And got:
=IF(SUM(H:H)>50,A*0.33,0)
A Better Try
Which is close, but not really how anyone even mildly proficient with Excel would interpret that request. But that’s not fair. It really needs to be a y=f(x) sort of problem, we suppose.
Here’s a different prompt:
Negative 1 if the sum of column H is greater than 50, otherwise 1 unless cell A4 is negative and then the result is 42
And the result:
=IF(SUM(H:H)>50, -1, IF(A4<0, 42, 1))
So where’s the intelligence? The result is correct, of course, but it isn’t that hard to figure out yourself.
Part of the problem with these sorts of things is that they are inexact. The website has disclaimers about “there may be imperfections in the model” and “verify this output before putting to use.” Verifying this is correct implies you could have written it yourself and it probably wouldn’t take much more time.
Looking Up Formulas
We were impressed that it seems to know how to find certain relationships. For example:
Find the power if the voltage is in cell B1 and the current in cell C1 =B1*C1
Or…
The current for a resistor in R1 when the voltage is in E1 =E1/R1
Here’s what’s funny, though. It must be keying on the actual names of the cells because try this:
The current for a resistor in C5 when the voltage is in C6
That flips the equation over for some reason:
=C5/C6
It just goes to show that the neural net doesn’t think about the problem the way we do. We think this one was right:
The amount of interest on a V1 balance at a compound interest rate in P1 compounded annually for V2 years =V1*(1+P1)^V2
But we were pretty sure this one was wrong:
The time it takes a ball to fall to the ground from Q1 feet =Q1/32
Our Opinion
So, in our opinion, this is little more than a parlor trick. You could have had some keyword-matching templates and gotten as good a result that would also be more repeatable. You could even ask for clarification or note impossible-seeming situations.
Is there a market for this? Sure, when it becomes Star Trek-level computer interaction. (“Computer: Run an analysis of all known Romulan transmission format…”). But at the level of GPT-3 guessing our formulas from natural language, it is just a toy. Even Wolfram Alpha which does a better job is still not up to the task, although we wonder how hard it would be for them to push out Excel formulae? At least it is better than the “prompt engineering” we’ve been hearing about.
That’s not to say that GPT-3 isn’t useful. It clearly is for some problems. This implementation of using it for spreadsheets just isn’t one of them. Still neat, though.
