AI Test - defeat Excel / me

I recently helped a client with a lovely annual return to our friendly local regulator.

It required some numbers (from a SQL query) to be put in some boxes (yep, in Excel 🤐) against a list of 251 countries.

The row heights had been shrunken and the column widths strangely embiggened so that a manual cross-referencing-and-copying into the spreadsheet was a time-consuming and error-prone task.

You could copy & paste into the relevant column (indeed the guidance encouraged it!) but for this to work you would need the full precise list of countries in the correct order… except you couldn’t extract this list from the spreadsheet since the entire workbook was locked down - locked cells and then password protected. No dice.

A couple of minutes Googling didn’t yield the exact list.

LLM time to shine

So. I politely asked some LLMs to be of assistance - could they extract the countries from the workbook?

Perplexity and Gemini flat-out refused to even try. Whilst ChatGPT could be coerced in to reading the file, it didn’t get much further in providing me with the list. The ChatGPT for Excel plugin got confused about what a column was 🤷‍♂️

Hmm. Taking a screenshot of the list and feeding that into Gemini did yield the list… partially - even on a 27" retina screen, 251 countries required multiple screenshots, so not without some faff.

I spent 10 minutes trying to understand CoPilot licensing and whether we had enough eligible prerequisite licenses on the correct upgrade channel. (We did not. I think. Still not sure.) ¹

Human power

Out of my way please. Everybody knows that an .xlsx file is really just a .zip file so by peering in to its contents and locating the relevant .xml file, with a little bit of regex could extract the 251 countries in the correct order. Pull those back in to the source query and in a few minutes the problem is solved, saving loads of manual re-keying.

So what

The tech will obviously get better. But for now at least, human ingenuity is still superior.

Obviously the point is: use the right tools for the right job. LLM was probably not the right tool, this time.

But in the grander scheme of things, it’s a test of the hitting everything with an AI-shaped hammer mantra. The fly in the elephant’s chardonnay² being that a 30 second human-to-human phone call asking for the unprotected workbook/list of countries would probably have got there even quicker.

Having said that, the free ride when you’ve already paid³ was that this bleddyawful spreadsheet was designed by a human in the first place.

So we are literally in a situation where an AI couldn’t fix a bloody awful human process. And this is why I, on occasion, feel bad for the AI, since it’s relentlessly tasked with imitating human behaviour which is, inherently, fucking stupid.