Stop Wasting Tokens

Added: 2026-05-05

[00:00:00]People often upload spreadsheets to their favorite large language models and ask the AI to analyze things.

[00:00:06]You know, give me the key insights from this file and tell me what I should do about it.

[00:00:11]And this seems like a reasonable request on the surface.

[00:00:14]But when you understand what's going on underneath, you realize that spreadsheets are actually really bad for large language models.

[00:00:21]In practice, this means you're paying way more for tokens and significantly increasing opportunity for hallucination and error.

[00:00:28]So, let me explain what's going on really with spreadsheets and large language models.

[00:00:34]You've probably used thousands of Excel files, but never even thought about what's going on inside that Excel file.

[00:00:40]Well, what if I told you that that Excel file isn't a file at all, but a whole directory of files?

[00:00:46]See here, I'm going to show you how to crack open an XLSX file and show you what's inside.

[00:00:51]Then I'll explain why Excel behaves the way it does and why that makes it fundamentally the wrong tool for serious analytics work with large language models.

[00:01:00]So here, I've grabbed an Excel file from a project that I'm working on.

[00:01:04]I'm going to rename the file extension from XLSX to zip.

[00:01:12]So far, so good. Nothing's changed.

[00:01:15]But now I can unzip this file and show you something interesting.

[00:01:19]To do this, I'm going to jump into the terminal because that's where I feel more comfortable and I'll run an unzip command over that file.

[00:01:32]So there you go. So what you see now is a folder structure with XML files, relationship files, shared strings, styles, and then the actual sheet data.

[00:01:42]Microsoft introduced this format in 2007, which if you're old enough to remember, was when XLS became XLSX.

[00:01:51]It's called Office Open Office XML and it's now an international standard.

[00:01:56]So your Excel file is actually a whole bunch of files with different use cases.

[00:02:01]The relationship files are like a map that tells Excel how to piece the content together.

[00:02:07]The worksheets folder is where your actual data lives.

[00:02:10]Each sheet is its own XML file. So sheet one XML, sheet two XML, and so on.

[00:02:21]So your one Excel file is actually a whole bunch of files with different uses.

[00:02:25]The relationship files are like a map that tells Excel how all the pieces connect.

[00:02:31]The worksheets folder is where your actual data lives. And each sheet is its own XML file.

[00:02:37]The shared strings file is the quirky part. Because Excel doesn't actually short store text directly in the cell, it stores an index number pointing to this shared script strings file.

[00:02:48]So if you type revenue in 500 cells, it only stores the word revenue once.

[00:02:54]But this is also why it Excel can behave unexpectedly when you manipulate files programmatically.

[00:03:01]And then you've got the styles file, which includes every bit of formatting, the fonts, the colors, the borders, etc. So why does all this matter?

[00:03:09]Well, Excel is really convenient for humans, but it's terrible for AI.

[00:03:14]Because the AI needs to do a lot of work to get the key insights.

[00:03:18]When you look inside that XLSS, which is actually a zip, you can see that the data and presentation are completely tangled together. Numbers, fonts, column widths are all baked in the same file, referencing each other through a web of XML relationships.

[00:03:33]And that's fine for a human when it's opening it in Excel, but it's a nightmare for machine trying to read it and do analytical work.

[00:03:41]It's like cracking open a coconut so you can extract the coconut milk. There's layers of husk and shell that you need to get through.

[00:03:48]And to do that, you're going to be burning through tokens. And of course, the more tokens you burn, the far more likely is you'll get hallucinations.

[00:03:55]So if you want to be kind to the AI and save yourself a bunch of tokens, upload a CSV file with just the data instead.

[00:04:03]Of course, the deeper issue is that Excel was never designed to be ingested by AI.

[00:04:08]It was designed to be used by a human sitting at a desk looking at a screen using Excel, and the format reflects that.

[00:04:15]The file structure is optimized for the application and for the experience of using the spreadsheet, not for making the data inside it really easily available.

[00:04:26]So as tokens get more and more expensive, you don't want to be wasting them by making the AI do hard work by ripping open the hard shell.

[00:04:34]Now of course, the big AI companies don't mind doing this work for you, of course.

[00:04:38]And right now, people are already complaining about token costs.

[00:04:42]And the thing is, we're not even yet paying full price for tokens, which will allow the investors who poured billions into AI to get a reasonable return on their investment.

[00:04:51]We're still in the drug dealer phase where they're giving out this stuff for less than cost price to get you addicted.

[00:04:57]But of course, as tokens become more expensive, you're going to be need to be thinking about how to make the most of your tokens.

[00:05:04]This is just one of the key insights behind the AI solution I've developed called Kira.

[00:05:08]I worked out a long time ago that large language models are very much like humans. If you make a task easy for someone, there's a much greater chance of them getting it right.

[00:05:17]Throwing a horrible, messy file format like Excel at a large language model and expecting it to sort things out is a recipe for disaster.

[00:05:25]My solution addresses that problem in a scalable and secure way.

[00:05:28]And in doing that, our token consumption is tiny, which means we get far more value from every prompt than simply uploading a spreadsheet and hoping for the best.

[00:05:38]If you enjoyed this video, please consider liking and subscribing. And as always, I love reading your comments and the feedback. And even if you don't agree, I try to respond to everyone.

[00:05:49]I've also recently activated memberships for those who'd like to learn a little bit more about what we do inside Kira.

[00:05:54][music] So thank you and I'll see you in the next one.