What spreadsheets need? LLMs, says Microsoft | Hacker News Reader

What spreadsheets need? LLMs, says Microsoft

theregister.com

31 pointsgalaxyLogic16 comments

A big chunk of the knowledge contained in enterprises is in spreadsheets. It's a no brainer to make that more easily accessible to LLMs. Wrap a thoughfully designed agentic framework around this and you presumably get a cheap junior data analyst that you can ask arbitrary questions of to better optimize value delivered through the knowledge you have. Anything from "free" drill down reports by territory on sales or ops to potentially running monte carlo simulations based on identified correlations to get a sense of the best classes of improvements to invest in to reduce shipping costs, improve sales conversions in specific verticals, etc.

I don't know if this is the framework, but this is one of the problems that needs to be effectively solved for large spreadsheets to unlock access to the data more efficiently.

> agentic

This is the hot new word in the LLM space. Was this picked because LLMs are losing luster for broad, cross-domain applicability? What systems actually demonstrate this behavior?

I'm not short on LLMs, but I can see a future where Gen AI in the creative space (image, audio, video) outpaces LLMs in terms of impact.

I was trying to write a decent definition but Wiki does it better: https://en.wikipedia.org/wiki/Large_language_model#Agency

In this context it could be able to do multiple actions in order to address an ask from the user: read cells, documentation, edit cells, and perhaps even read the result from the edited cell before answering. i.e. "Can you create a new sheet that focuses on the Countries where the sales happened, including YoY differences, and tell me which countries are outliers and for which reason(s)?" (probably very far fetched given the level of progress Excel has achieved in 20 years).

Excel already destroys RAM... sure let's add the most memory intensive process known to man into the mix.

I hope it helps because right now copilot is really not good at Excel.

still can't parse ISO8601

The way it confidently parses the few days of each month as nonsensical American dates is the icing on the cake.

Hallucinated bookeeping and forecasts are going to make for some really interesting outcomes.

This will be a huge time saver for shady accountants everywhere.

Is this basically about giving us an alternate prompt-based user-interface to Excel and other applications?

No, this has more to do with uploading your sheets to the cloud. "It's so the Assistant can see it! No other reason! Honest!!! We wouldn't dare do anything else with your sheet."

More seriously, making spreadsheet formatted data more easily consumable by curve fitting ML algorithms actually can be useful. I would just would want total control of where that data goes, and unless you roll your own, or have a contract to purchase your own, you don't get that.

I do not know if this is a good idea, this AI thing is looking a lot line the dot-com era of the late 90s.

They're going to give tax accountants across the country terrible headaches.

More garbage shoved at users to pump up the markets. Does anyone use LLM stuff that meta added to whatsapp and instagram. Everyone i know hates that shit.

When are these ppl going get held accountable for this farce.

RIP copy-paste.

Theyll probably find a way to doom dna science again.

It's not Microsoft's fault geneticists were using Excel incorrectly.

Reply on news.ycombinator.com