ChatGPT saved me an hour of spreadsheet work today. It wrote a complex sorting script beyond my ability, explained how it works, and helped me deploy it.
The party tricks have been fun, but this was a whoa moment for me similar to the first time I used Google Search.
Some are asking if I could share the prompt that I used. But that’s just it—I’m not an expert in Excel and it wasn’t a single well-written prompt. It was more like a long chat with an Excel expert who was eager to help.
I think I started with something like “Help me write a script in Google Sheets.” Then just started describing the data and everything I wanted it to do. When something didn’t work, I’d feed the whole code back to it and ask what changes I need to make in order to do X.
I asked for rows within a range to be sorted based on values in several columns in order of priority. When certain conditions are met, values in a range are summed and added to another sheet. There’s a timestamp for new data. That sort of thing. It’s best to work in chunks.
Some issues came up, and I’d just post the error code or describe how it wasn’t doing what I expected and ask for some potential reasons why it wasn’t working. That was very effective—it would give me a clear list of possible problems, and often it was number 3 or 4.
Sorry I can’t be more specific—it’s trial and error. The key is ChatGPT’s natural language abilities. Just have a conversation like you would with a colleague on Slack—no need for a perfect prompt.
I’ve had some VBA macros stop working after recent Office updates.
Maybe I’ll see if ChatGPT can rewrite them for me. Or come up with a better solution.
Okay, I actually did not use ChatCPT for anything myself
HOWEVER
I really liked the discussion over ChatGPT in education:
One of the uses the guy in education talked about was using it to generate examples for use in lesson plans, which I really liked. Using it to generate text on the same topic but with different tones, as it were.
Also, I’ve used AI tools for a couple things lately - not ChatGPT, but something called Lex, and if you’re interested in it, I’ve got 11 invites currently.
I’ve used Lex to generate titles/subtitles for stuff I’m writing. It supposedly helps kick people out of writer’s block, but writer’s block is far from my problem.
Generating titles is a problem for me, so on a recent piece I was working on, I took the executive summary, fed it into Lex, and asked it to generate 10 titles for me, and then asked my colleagues to pick the best one.
The unwritten part at the end of numbers’ sentence was “as your own”, that is not the case for GPS, auto-pilot and a fuel gauge. The black box model at work, there’s an ASOP for that.
Black boxes are in a lot of places that we just take for granite (solidly, like a rock) that they work properly (“they” wouldn’t allow them if they weren’t tested and approved, right?). I can’t seem to think of any at the moment.
OK: Google will often show business hours of businesses. What is the source of that info? The proprietor? Other users whose word we must take? How old is that info? Sure, it would be nice if the source was legitimate.
Same for Waze or Google Maps: says there is some slow-down, and on Waze, someone has piped in to say there is an accident and left lane is closed. Why should we blindly trust that info? How many drivers are taking that alternative route that is now also slowing down? Fun question: how do these apps know the speed of traffic? Via ground sensors? Via turned-on cell phones in cars? Via cars’ computers sending info?
I love how people are all “OMG AI’S GONNA KILL US!!!” and I’m thinking, why the hell are you giving it access to those sorts of systems… you’re not even putting controls on your spreadsheets.
…oh yeah
Seriously, humans are lazy and stupid, and we lie. We created AI in our own image.
AI is not going to be used anytime soon over here as there are too many issues still to iron out.
At most we will only be rolling out AI-assisted MSFT apps once they come out with something workable.
It will definitely improve productivity as a lot of manual work (not sensitive) can be automated. Minutes of meetings for example. Much easier with AI-assisted audio transcription (the current one they have is not very good).
The text below came from an abstract via EuSprIG (European Spreadsheet Risks Interest Group)
Bringing generative AI to the Excel grid: from research to practice (Keynote)
Andrew D. Gordon and Jack Williams,
Calc Intelligence Microsoft Research
Team link: https://aka.ms/CalcIntel
Abstract
Learn how to transform your textual data within the Excel grid using the new LABS.GENERATIVEAI() function. It empowers Excel users with OpenAI’s pre-trained language models. It’s part of the new add-in Excel Labs, a Microsoft Garage project that also delivers new formula editing and reuse features, including easy editing of LAMBDAs. We’ll showcase the productivity benefits of using LABS.GENERATIVEAI() in conjunction with LAMBDA-defined functions. These features build on our research in the Calc Intelligence team at Microsoft Research in Cambridge and arise from a long-term partnership with Excel
This paper adopts a critical approach to GPT-4, showing how its huge reach makes it a useful tool for people with simple requirements but a bad, even misleading guide to those with more complex problems which are more rarely present in the training data and even more rarely have straightforward solutions.
It concludes with a practical guide for how to add an Excelscript button, with system and user prompts, to the GPT-4 API into the Excel desktop environment, supported by a blog post giving the technical details for those interested.
Experimenting with ChatGPT3 for Spreadsheet Formulae Generation: The Risks of AI Generated Spreadsheets
Simon Thorne
Cardiff Metropolitan University
Abstract
Large Language Models (LLM) have become sophisticated enough that complex computer programs can be created through interpretation of plain English sentences and implemented in a variety of modern languages such as Python, Java Script, C++ and Spreadsheets. These tools are powerful and relatively accurate and therefore provide broad access to computer programming regardless of the background or knowledge of the individual using them. This paper presents a series of experiments with ChatGPT to explore it’s ability to produce valid spreadsheet formulae and related computational outputs in situations where ChatGPT has to deduce, infer and problem solve the answer. The results show that in certain circumstances, ChatGPT can produce correct spreadsheet formulae with correct reasoning, deduction and inference. However, when information is limited, uncertain or the problem is too complex, the accuracy of ChatGPT breaks down as does its ability to reason, infer and deduce. This can also result in false statements and “hallucinations” that all subvert the process of creating spreadsheet formulae.
