Money quote:
Excel requires some skill to use (to the point where high-level Excel is a competitive sport), and AI is mostly an exercise in deskilling its users and humanity at large.
Money quote:
Excel requires some skill to use (to the point where high-level Excel is a competitive sport), and AI is mostly an exercise in deskilling its users and humanity at large.
OK, I’m not really mad at this. I already used Copilot to design a table for me in Excel and it worked really well. It did everything for me, and I just had to copy-paste the formulas into their appropriate spots. If it’s built-in, possibly will work better.
Not everybody needs to be an Excel expert, after all. Having that functionality might be actually beneficial.
How do you know those formulas are correct?
I’m talking about using it when you’re “not great at Excel”, not when “you can’t do basic math”.
Always verify the results given to you by LLMs.
By verifying that they’re correct…? 🤔
I think the concern is that you can come up with a number of formulas that will get correct answers for some combinations of values and not others.
If you do not understand the logic of the formula, and what each function does, how do you verify they are correct and will always give you the results you think they will? Double check every result in its entirety?
I think you’re completely missing the point here.
I’m not great at Excel. That doesn’t mean I can’t do basic math, it means I struggle designing an
xlookuporhlookup.If AI does that for me, I’ll be a happy bunny. And then run a dozen different iterations of data to verify that the results I’m getting are correct.
This is what this integration is for - it’s not a replacement for a human brain, it’s an assistant. As are all LLMs.
This is what I think AI and automation is generally good at and should be used for - mitigating unpleasant or repetitive work so that the focus of the user is productivity/creativity.
The context is something we disagree on wholeheartedly. Those funding and fundraising for AI and an enormous subset of those using are not looking to use AI in the way we are talking about. The prior are hoping to use AI to extract value from it at the expense of people who would otherwise need to be paid, or they and claim it can do anything and everything. Those using it, many of them, do not have a sufficient understanding to comprehend the solution. They are basically “vibe coding”. Tell the LLM to do something they aren’t knowledgeable about, then keep telling it to fix the problems until they don’t see problems anymore. Yes, spreadsheet formulas are likely simpler than an app but I know people who use AI for Google Sheets and they rarely test any results, let alone rigorously.
Anecdotal, sure, but I don’t have enough faith in humanity to presume everyone else is doing something wildly different.
Edit: To expand, LLMs specifically, are what I consider to be the worst side of “AI”. You can use ML and neural networks to create “AI” (self altering, alien blackbox algorithms) to become proficient in analyzing information and solving problems. LLMs create a situation where the model appears intelligent because it knows how to mimic language… and so now we pretend like it can do whatever people can do.
Well… Yeah, I get what you mean, and - in general - I agree.
However, to me it’s also a bit like criticising the use of hammers because a lot of idiots hit themselves on the heads with them. Or, even worse, hit others on the heads.
AI/LLMs are a tool, and just like any other tool, they can be misused. That doesn’t mean the tool is bad, or immoral, or whatever, to use.
That’s why I hate the today’s discourse of “anything that has AI is shite be default” that so many people online have.
Let’s laugh at obviously bullshit attempt of shoving AI down consumer’s throats, but when it comes to actual, proper implementation - like in the case of baking Copilot into Excel - it becomes yet another optional tool at users’ disposal.
I think it would be infinitely better for an LLM to walk a user through the use of the formula in their specific use case rather than do it for them… but that won’t sell as well because most people don’t want to learn to use a spreadsheet they just want to do a thing and move on to something else. This is how it is sold and this is why it is used, in most cases. It’s not a hammer that people misused despite there being nothing in the sales material about it’s usefulness as a bludgeoning device against other humans. LLMs, spreadsheet copilot included, is commonly packaged and sold as a magic solution that will just do the work for you, with an asterisk and fine print stating that it’s for entertainment purposes only and that whoever isn’t liable for any false information or whatever bullshit clause they come up with. People use it as it is sold to them and that’s what worries me.
I just had my place of work upgrade me to Windows 11 this week. In order to install office, I was directed by Microsoft to download the “Office 365 Copilot” app which downloaded the office installer. Copilot is not subtle. It may be technically optional but good lord does it want you to know about and use it for everything.
And no, I didn’t try it yet. I will likely be trying it and Gemini soon out of curiosity. Last time I tried to use it I was given hallucinated nonexistant python modules and powershell commands that wasted my time. It’s been a year or so though.
The one time I used it (via the Microsoft 365 Copilot), that’s exactly what it did. It only created the table in the spreadsheet that I wanted and then explained how to do the formula, explained the bits and bobs of the formula, showed me the code and told me where to put it.
Sure, if I was lazy, I’d just copy-paste without thinking, but the information was there.
At the same time, you could be arguing that calculators should be teaching people how to do maths, instead of just giving them the result, right? It’s actually a bit similar to LLMs, because, well, if you don’t know how maths and calculators work, you can get a result that’s horribly wrong (try
3+2*3on any simple calculator and you’ll get15).So, first of all, it sucks that your IT didn’t have the tools to handle this for you in the first place.
Secondly - that’s just Microsoft being Microsoft. They love changing names and making things as confusing as possible. Although now Copilot is part of that app, when they originally introduced the new name, it was just a rebranded Microsoft 365 App. We were joking that it was done by some middle manager so he could boast that “100% of Office users now utilise Copilot”.
I haven’t yet seen an LLM that didn’t invent PowerShell modules, although recently Copilot’s been pretty good.
That’s my thinking
If you know what you’re doing, it’s significantly easier to do it yourself
You at least have some reassurance it’s correct (or at least thought through)
Verification is important, but I think you’re omitting from your imagination a real and large category of people who have a basic familiarity with spreadsheets and computers, so are able to understand a potential solution and see whether it makes sense, but who do not have the ability to quickly come up with it themselves.
In language it’s the difference between receptive and productive vocabulary: there are words which you understand but which you would never say or write because they’re part of your receptive, but not productive knowledge.
There are times when this will go wrong, because the LLM will can produce something plausible but incorrect and such a person will fail to spot it. And of course if you blindly trust it with something you’re not actually capable of (or willing to) check then you will also get bad results.
yes okay dad
Any time, nutsack.
I’m a dad and I approve this message.