Hacker News Reader

Comment by tromp | Hacker News Reader

> What is Copilot?

> Think of CoPilot as having your very own ChatGPT seamlessly integrated into all your Microsoft applications – Windows, Edge, Excel, and PowerPoint.

> In fact, it’s essentially Chat-GPT, powered by OpenAI, but with a unique twist – it’s context-aware and fully capable of interacting with your data and the software running on your PC.

Asking Excel to "add a column averaging the last 3" sure seems a lot easier and faster than having to read the manual to figure out how to do such a thing. This could drastically simplify user interfaces.

View on HN

"Oh.. but skip cell A.. 6? or whatever, the one with the .. uhh... "

'Error processing your request. I emailed your supervisor with my error condition.'

"No, no! Stop. Just do whatever you did first, but now you do like the other thing too, with the skipping of the cell with the.. uh.. what's it called?! ampersand?"

'I will average the last 3 columns while adding random numbers to the rest of your document, sure. By the way, this is the information you requested. I have also forwarded it to the company CEO: 'The ampersand, also known as the and sign, is the logogram &, representing the conjunction "and". It originated as a ligature of the letters et—Latin for "and"'. SOURCE: wikipedia.'

Proceeds to mess around with LLM for 3 years instead of learning wat AVERAGE() is and how it works

You joke, but this is how people normally work. Do things in ways easy to lean, but ultimately slow and inefficient.

I have learned this from experience. If you think this is bad, real people in real situations are actually worse. If you can get them to speak quasi-coherent sentences that’s a major win in itself. Sounds sad, but the reality these gadgets must operate under.

It’s not that people are stupid, but they operate under a surprising amount of varying conditions ranging from fully inebriated to severely demotivated and/or burned out.

"Oh.. but skip cell A.. 6? or whatever, the one with the .. uhh... "

> Which column would you like me to skip?

The one with the sales during covid.

> Okay, I have averaged 2 of those last 3 columns, skipping the one with the sales during covid which had a header with the text "COVID data".

It's not as simple as just using AVERAGE actually. I've had the exact problem with Google Sheets. I wanted to create a trailing average of last 7 of the numerical values in another column. I had to Google quite a bit to figure out how to handle it so the 1st to 6th row would work out well. Right now I described the issue to ChatGPT, and it was very quickly able to suggest me the following for Excel:

=IF(ROW(A1)=1, A1, IF(ROW(A1)=2, AVERAGE(A1:A2), AVERAGE(OFFSET(A1,-2,0,3,1))))

For Google Sheets and trailing 7 it suggested me following:

=IF(ROW(A2) < 8, AVERAGE(A$2:A2), AVERAGE(OFFSET(A2,-6,0,7,1)))

If you know everything by heart, it may seem simple, but it's not just using one function, it's using multiple and knowing the exact parameters, intricacies.

With Google Sheets, it would've faster for me to just use the JavaScript with Apps Script, to create this, rather than trying to figure out the syntax of this.

So it uses 4 different functions there,

1. IF

2. ROW

3. AVERAGE

4. OFFSET

And in such a way that I would just not come to think of it on my own out of the blue. I would try several different approaches before I were to try this one. Although it's really easy to just write a script for that.

And maybe you also want to be able to dynamically configure the trailing average size, which this formula allows for very easily. Compared to when you Google, some of the articles suggest just using AVERAGE(A2:A8), but it's missing those other frequent requirements.

And when you look at some Stack Overflow answers the formulas are even crazier:

https://stackoverflow.com/questions/40779177/how-to-average-...

=round(AVERAGE(OFFSET(H1,MAX(ARRAYFORMULA(ROW(H:H)--(H:H<>"")))-7,,7)))

https://stackoverflow.com/questions/70282014/find-rolling-av...

=INDEX(AVERAGE(ARRAY_CONSTRAIN(SORT(INDIRECT("A1:"& ADDRESS(MAX(ISNUMBER(A:A)ROW(A:A)), 1, )), SEQUENCE(MAX(ISNUMBER(A:A)*ROW(A:A))), 0), 7, 1)))

It's odd to me how office people are totally fine wrangling with this mess, but are terrified of programming. This looks way more confusing, magic and brittle to me.

I think the dream goes beyond just that sort of simple "do something in this spreadsheet because I can't be bothered to think".

I think the end game is you open excel and say "Work out what our budget is for the FooSnaggit event based on what Tim said in his email yesterday, and what we spent last year." (for example) and it will go off and read your email and read your other sheets and generate the right thing based on that prompt and that context from your other documents.

I think that will be incredibly powerful ... if it can be trusted. Trouble is now I can't trust these things to get these sort of things correct, so you end up having to double check everything and make corrections etc so you are not really saving that much time.or effort.

Are we still talking about a language model?

These tools can be great for discovery and speed up one's learning curve but there is no replacement for knowing your tools if you're going to master your craft.

In a way, after you devote many years of your life mastering your toolset, you become a very efficient intelligent machine.

But you do also so make mistakes and you also are told to do things by somebody else who guess what didn't spend their life mastering the skills you mastered (otherwise they wouldn't be paying for your services).

From their point of view, you're also some kind of opaque machine that can be trusted to some extent but must be double-checked too. You also can rat them out if they ask you silly things and you also expect to be treated decently as a human being and that's not something many other humans know how to do well.

Unsurprisingly many people cherish the possibility of having to talk to a non-human entity and let them do some tasks they wouldn't know how to even approach. Especially if this entity can be pestered with questions without ever being annoyed.

It's all about POV

So far all my interactions with AI in production have involved me trying to figure out the right prompt to be able to talk to a real person who could then actually solve my problem rather than offer me useless text and no concrete action towards solving my issue.

You can't trust an intern with that command either, yet that's exactly the type of task an intern would be assigned today in many large groups.

Of course you'd want to verify the report, but verifying it will probably be simpler & faster than building it, for many tasks. And AI can help you verify as well.

For personal use and small businesses your example is a good look at what CoPilot can do.

For larger enterprises it is unlikely that this amount of data will be allowed to be sent for CoPilot processing.

The overarching question is: Do you want to transfer this much control over your data to a single entity ?

Well for large enterprises that already have their data in the Microsoft ecosystem (Exchange email, Sharepoint etc), Copilot has access to that pool and it's supposed to be isolated within your own environment.

Will output quality of an average excel user go down, if the job duties will evolve into verifying LLM output and not performing low-level tasks over and over?

Laziness is a safe bet, so we will see :)

>>I think that will be incredibly powerful ... if it can be trusted. Trouble is now I can't trust these things to get these sort of things correct, so you end up having to double check everything and make corrections etc so you are not really saving that much time.or effort.

This is the thing for me. I hear about people building apps and stuff but I can't get a working correct script from ChatGPT. Maybe it's me, maybe it's improved, IDK, but the last few times I tried to grab a script for something I was working on, that script had multiple errors.

Copilot will collapse for any query more complex than 3 clicks. And any intermediate excel user will be able to perform these clicks quicker than to formulate their prompt.

The second one, probably. But ChatGPT-4 can handle surprisingly complex tasks well for SQL, I don’t doubt it would do well across the office suite.

Obviously it doesn't have the context window to handle things I really want it to do in Python (“refactor this spaghetti mess I got myself into, separate the giant classes into multiple separate modules, fix any O(n^2) parts, add unit tests and do some fuzz testing on function arguments”) but it will be incredible once it does in 10 years or whenever GPU’s have enough VRAM or we figure out how to stuff large context windows into less VRAM.

And in places that have an undo-stack stakes will be quite different from, say, self-driving death machines on public roads.

Someone was already mentioning clippy, presumably as a joke (a well deserved one, I'm old enough to remember). But would an LLM-powered clippy be a joke? Could be the perfect capability/requirements fit.

How do you envision this simplification of user interfaces? How exactly will you ask Excel to "add a column averaging the last 3"?

Note: neither asking by voice nor typing it out simplifies the interface.

If someone can't do an average function, I don't think id have them near Excel.

If someone can't write assembly code, I don't think I'd have them in an IDE.

False equivalence

Not really, in both cases the question is about low-level (ex: machine coding) vs high-level (ex: using a GUI) interactions with a particular tool.

Being vaguely related doesn't make an equivalence true

I can do an average function in Excel. But I can not remember the word of the function in all three languages that Excel have decided that I should be proficient in Excel in. The translation of functions is the single worst design decision in that product.

You might be right for this scenario.

For a contrast: I use Copilot in vscode a lot. There it's pretty good at guessing what I am trying to do, and offering an auto-complete for it.

So mostly it feels like it's just helping me do the thing I wanted to do anyway, but faster.

But if eg a user knows how averaging of numbers works, but just doesn't know how to express it in Excel, even that example would be useful. I agree that if a user doesn't know how averaging works on a conceptual level, copilot won't save their butt here. (At least not yet.)

There was this gimmick many years ago of a computer driving licence.

An Excel driving licence, mandatory for most white collar work, would be no bad thing.

Insert the clip from Breaking Bad of Skyler White talking about doing company finances in "the quicken".

But this time "I just asked the AI to do the excel"

Welcome to the wonderful world of function names being translated to the installed language.

My first language is Norwegian, and except for a few that I recall, I constantly have to refer to a website like this[1] to figure out what the hell they decided to call a function in Norwegian.

[1]: https://www.perfectxl.com/excel-glossary/what-is-excel-funct...

It could also confuse users once more complexity arises because people often ask for contradictory things when the constraints of the spreadsheet are not guiding them.

Yes and no. It might be faster the first time, maybe not so much the consecutive times. Especially when the AI assistant is not doing what you intended.

Unfortunately the problem with non-experienced (excel) users is that they don't know what they want rather than lacking who to ask.

Reply on news.ycombinator.com