Hacker News Reader

> It must be much harder than I first thought, but then again they spent BILLIONS on this.

Which is why MS and every other company keeps pushing half-baked garbage AI products onto the public at every opportunity. So much money has been sunk into AI that they need something to show for it. They also hope that as people use it they can improve the AI they have so that maybe one day it can become something worthwhile. As a bonus it's one more way for them to collect your data and get you used to asking an AI for what you need instead of using your brain.

I asked o3 pro to do something like this a few months ago when I paid for a month of ChatGPT Pro. It got very excited and said it was on it and spent 20 minutes absolutely cranking. Then it spit out the world’s saddest excel file with no formatting, and half the data truncated.

Sounds like my post-coffee morning routine, with similar results

Your description was hilarious!

Thank you for the laughs :D

OT: I wonder if we'll see some "core personalities" that are innate to each mega-LLM (Claude, O3, Copilot, Gemini)... methinks this has comedic potential.

I’m getting Three Stooges vibes - there’s a real possibility here!

The data truncation for no apparent reason is the modus operandi of all the major LLMs. Then you instruct them to not truncate the data and they get over it but in the end you still cannot trust whether they did it or not unless you verify yourself.

Having recently watched a corporate trainer spend 15 minutes trying to figure out why Copilot wasn’t correctly importing a CSV with headers when “it worked fine when I showed this 2 weeks ago” — my guess is that the tech isn’t quite there yet to sell (reliable) AI-powered Excel

The tech is arguably not there yet. From my own observations, normies get annoyed quickly because they expect it to work like deterministic software. And they tend to be way less interested in it in general, especially when I’m breathlessly telling them about this and that latest development.

oh wow people want their software to be deterministic and with the same set of inputs get the same set of outputs! What is this, the 90s?

> normies get annoyed quickly because they expect it to work like deterministic software

In very few use cases it is acceptable to have non-deterministic result for computation tasks. It does not matter if you are a normie or an advanced user.

Normal people don't understand what the word deterministic is, nor do they really expect their software to produce deterministic results. For one thing, they're not running the operation multiple times and comparing the outputs. For another, if they give the same task to three different people they're going to get three different results anyway, so what does it matter if the computer gives three different results, if they even notice.

> Normal people don't understand what the word deterministic is

I would argue that they implicitly do, as any user expects the same action performed on a computer or similar system to provide the same outcome.

> For another, if they give the same task to three different people they're going to get three different results

Give the same three tasks to a single user to be executed three times separately, and he will get supremely annoyed if his actions do not give him the same results.

Not even implicitly understand. A layman is perfectly capable of understanding what deterministic behavior is, and claiming otherwise is just condescension.

Then you and I work with very different people. Here it seems "reset it and try again" is considered a viable problem solving technique.

It sometimes is, as resetting may clean some dirty state. Even when people don't really understand why they are doing it, and are merely following trends they saw elsewhere, it does not invalidate the point.

I think “normal” (non-computer) people have a mental model for computing that is more like a (possibly bad) coworker.

They would quickly adapt to “reload it a few times if it doesn’t work right away”.

Isn’t this what people are already doing when they browse the web? I know I am.

What are you basing those claims on?

But I can ask the AI, it understand what I want and give me the steps. I can ask it to give it to me in api call instead and it does, after a bit of mashup.

Sure the api link and permissions and yada yada plays a part, but thats exactly how they can trap us into office365, with already use azure permissions and everything.

Again I'm sure it's harder than it looks, but it's not an AI problem anymore, and they're throwing billions at it.

> normies

do you mean users?

Yeah they don't understand we're in the equivalent phase that the commodore 64 was for the computer.

Microsoft's branding is mostly about time and should not be assumed to have any other significance. You might think you can guess what "Visual Direct COM" would be or "Active Windows.NET" but the words have no meaning beside "Currently this word is hot so my new product needs that word". A Copilot XBox Edge might be a video game console with AI but it equally might be a new version of Excel. There is no coherent rationale.

A company with an actual rationale names products PlayStation, PlayStation 2, PS3, PS4, PS5

Microsoft called their rival products XBox, XBox 360, XBox One, XBox Series

Asked to put six animals in order, some might figure Anteater, Bear, Elephant, Fox, Goose, Pigeon makes sense - alphabetical order, English names. Others might try to rank them by size, the Elephant is definitely bigger than a Bear, but is the Fox bigger than a Goose? Not sure. You might give them Latin names, there are several reasonable things you might do or at least attempt

But Microsoft are like that's easy: Elephant, Dog, Squirrel, Another Squirrel, Sparrow, Bear, Elephant. And like, that's not even six animals, and it's the wrong animals, and your order makes no sense, what is wrong with you?

How many times has Microsoft had some kind of product that isn't taking off how they wanted to, so they just rename it and give it a new coat of paint? Microsoft Office? I think you mean Office 365! What does that mean? Nothing, it's just Office but you don't get to decide when to update.

Anyway, now it's Office 365 Copilot! What does that mean? It means it's Office, but with an AI which you didn't ask for, which doesn't really do much for you practically, and also which costs 50% more, and you can only opt-out by trying to cancel your subscription entirely.

You can tell AI is a grift because it's all dark patterns and lies with these people.

Teams being rebadged Skype for Business being rebadged Lync is the biggest one.

To be fair to Microsoft, they do keep trying over and over again.

On my Android, it's called "M365 Copilot", which is even more wtf...

Just like most big corps, for a while they had a phase after Satya took over where it felt a new kind of Microsoft was coming, nowadays I would assert they are back to their former selves.

I’ve been thinking about something similar - Microsoft needs to make dead simple “AI scripts” record: I download this csv file, I then manipulate the data this way, add these formulas, add a cover sheet, do these checks, and then upload the new sheet a with the following columns into system X. Once you have that recorded, and adjusted to what you want. Next you just download the file and run the script.

The amount of repetitive data manipulation at companies is literally insane, just ask anyone in accounting, operations, hr, etc… just being able to automate repetitive task in excel for people that can’t code or aren’t excel/Vba super users would be incredibly valuable.

  > The amount of repetitive data manipulation at companies is literally insane

Serious question: why isn't coffee being thrown at the low-hanging-fruit takes? Not only would it make them quicker, it would make them more reliable as well.

This is exactly what PowerBi does, once you downloaded your file if it has the same name as the previous one and is in the same place you can just hit refresh in Bi.

That’s how custom agents in Microsoft copilot and Microsoft copilot studio work.

You pretend it’s AI, but in reality you’re just building literally power automate flows which you can call through an LLM lmao

> "hey excel, make a summary of those 8 sheets to identify our 5 most profitables products and their evolution in sales the past 3 years, add that on a new sheet at the end with a visual graph"

Aren't they just pattern recognition and regurgitation machines? How would it analyze and interpret unique data with a pattern it's never seen?

Some of the AI stuff is very useful, but it's been massively oversold. In my experience, it's great for working on documentation or creating a one-time-use script to query an API, but it's not good at the "big picture" tasks they want for a sales pitch.

  > How would it analyze and interpret unique data with a pattern it's never seen?

That's called unsupervised learning, and it was field of study in machine learning long before LLMs were anything near viable. Surprisingly, LLMs are good at it too.

I would give the LLM statistical tools that work to find general anomalies, distributions etc. from the spreadsheet. Then it's about an LLM interpreting (or not) those results in natural language.

Of course this is a can of worms for a product because we can still not guarantee accuracy.

That makes sense to me in terms of tech, but I don't get how the economics will work. If one company has an agent that's a great statistical tool, another is great for making charts, another is great for grammar and spelling, etc., I'm guessing it'll cost a bazillion dollars in subscriptions.

Fully spot on, I want AI to feel natural, part of the workflows of using an OS, improved search, better OCR, handwriting, in general a better user experience, not something bolted on at all costs, with chat windows all over the place.

Thanks Microsoft, for giving me yet another reason not to use Edge beyond browser compatibility matrixes in project delivery.

Unfortunally this is also coming to all their developer tools, VS, VSCode, PowerShell, Windows Terminal, .NET Aspire, Azure,... everything is getting AI there somehow, as each team struggles to meet their AI OKRs, and reason of existence among endless layoffs.

Easy. 50 different teams get to 'make impact'.

> I don't understand what microsoft is doing with their Copilot whatever push.

Anything they can think of. Honestly, anywhere they think that they could put AI to try to convince people that it's a real thing to care about, they put it there. Google is doing the same thing; I get popups any time I try to access any Google thing, like mail, docs, or the Google Cloud console. I don't care, Gemini, and I don't trust you.

Even Apple is jumping in on the AI train, presumably just to avoid getting beaten up in the press for being "behind the times", but thankfully they seem to be trying to start slow and make a good product out of it (which they have yet to do) rather than telling everyone that it can revolutionize everything that anyone ever does and forcing people to hear about it at every opportunity.

> 50 different offers, all with their own pricing

They also randomly time out mid-task, e.g. in image generation, without an obvious way to pay for more compute time.

They’re going for the “if we throw shit at everything then some of it will stick”. The problem is everything is covered in shit.

The things that aren’t shit are harder than they thought.

Check out Sourcetable, you might like it for your AI spreadsheet work. It's much better than Excel copilot in every aspect (financial modeling, data science, data cleaning, agent tools, etc.)

We use https://www.quadratichq.com for this. It works pretty well too!

It really is super obvious that the winning move for most apps is to add an text interface with LLM capabilities. Like autocad - but LLM accessible.

You say "text interface" because that's what you're most used to. It could as well be audio interface.

Not quite the thing you're describing, but I saw this demoed recently and thought it was pretty cool: The AI function in Google Sheets[1]. Call Gemini with the context of individual cells and have it respond in the same cell. Take a look if your company is in Workspace Labs.

1. https://support.google.com/docs/answer/15820999

Disclaimer: Googler, opinions my own.

I feel like companies are getting there, it's just that everyone is scrambling now and things are in early stages.

My experience with Google Docs + Gemini is similar. Gemini is great for coding, but the integration with Docs / Sheets / Slides / Drive is absolute garbage; does hardly anything useful, doesn't do the actual tedious jobs I want it to do. It couldn't even find a document in Drive based on my description, or create a decent-looking slide.

(Gemini is proving excellent for coding, but it shocks me how poor the integration is for other uses.)

The "AI" integrations I've tried in other tools (Figma, Gamma) are pretty much garbage too.