How is looking at ascii values supposed to help when someone prompts it with “calculate the sum of the numbers above”? The whole point is that no matter what kind of prescreening you add to an LLM, people will write prompts which are missed by the screening.
There are no values from the spreadsheet in this case. “The numbers above” are just text to the LLM.
They could of course require that optional cell or cell range parameter after the prompt, but that would eliminate some use cases. “Generate some text”, one of the stated use cases in the help text, doesn’t reference any cells.
Also, numbers in Excel aren’t necessarily as clear cut as you make it seem. Excel famously thinks everything is a date, and how number-y must a number be before it isn’t okay?
Not to mention there are other things to do with numbers which don’t require arithmetic. What if someone wants to have Excel translate 34 to “thirty-four”? Or have Excel generate a poem 34 words long? Or whatever else nonsense people might try.
The text field says =COPILOT("sum the numbers above"). It doesn’t work that way. Excel does not have any concept of what “above” means here. Those numbers are not used in the calculation whatsoever. To reference those numbers, the field should say =COPILOT("sum the numbers in", A1:A3).
What the user did here was ask the LLM to generate some text based on a text prompt and no other data, and the LLM decided to answer with a string containing only digits.
Yes, but “above” just goes into the LLM, which does who-knows-what with it, and certainly isn’t designed to address cells that way. So to the LLM that’s just like any other arbitrary text.
There would be no way to do that reliably. There’s too much weird stuff people might say to reference things, and the LLM would definitely act on the wrong cells more often than not.
Excel already has a perfectly unambiguous way to provide a specific range of cells, Which is why the =COPILOT() function lets you supply those in the second parameter. I’m assuming they get passed to the LLM as context, likely encoded as a markdown table. LLMs love parsing markdown, apparently.
The user provided no such range of cells, though, so the LLM is most likely seeing none of those other cells, and is just working based on random values from the Internet.
How is looking at ascii values supposed to help when someone prompts it with “calculate the sum of the numbers above”? The whole point is that no matter what kind of prescreening you add to an LLM, people will write prompts which are missed by the screening.
Because you can check if values input from the spreadsheet are non-numeric.
There are no values from the spreadsheet in this case. “The numbers above” are just text to the LLM.
They could of course require that optional cell or cell range parameter after the prompt, but that would eliminate some use cases. “Generate some text”, one of the stated use cases in the help text, doesn’t reference any cells.
Also, numbers in Excel aren’t necessarily as clear cut as you make it seem. Excel famously thinks everything is a date, and how number-y must a number be before it isn’t okay?
Not to mention there are other things to do with numbers which don’t require arithmetic. What if someone wants to have Excel translate 34 to “thirty-four”? Or have Excel generate a poem 34 words long? Or whatever else nonsense people might try.
What hill are you trying to die on here? In the picture, the numbers input are clearly 1,2 and 3.
No. It is trivial to identify if the input is text or value. You don’t need an LMM to do that.
That’s only interpreting data entry. Date values are stored in cells as doubles.
Yes these are the use cases where an LLM would add some (questionable) value.
The text field says
=COPILOT("sum the numbers above"). It doesn’t work that way. Excel does not have any concept of what “above” means here. Those numbers are not used in the calculation whatsoever. To reference those numbers, the field should say=COPILOT("sum the numbers in", A1:A3).What the user did here was ask the LLM to generate some text based on a text prompt and no other data, and the LLM decided to answer with a string containing only digits.
Excel knows the address of the calling function.
Yes, but “above” just goes into the LLM, which does who-knows-what with it, and certainly isn’t designed to address cells that way. So to the LLM that’s just like any other arbitrary text.
I don’t believe even Microsoft are this stupid. The llm will have tools to interrogate the spreadsheet.
You give them too much credit.
There would be no way to do that reliably. There’s too much weird stuff people might say to reference things, and the LLM would definitely act on the wrong cells more often than not.
Excel already has a perfectly unambiguous way to provide a specific range of cells, Which is why the
=COPILOT()function lets you supply those in the second parameter. I’m assuming they get passed to the LLM as context, likely encoded as a markdown table. LLMs love parsing markdown, apparently.The user provided no such range of cells, though, so the LLM is most likely seeing none of those other cells, and is just working based on random values from the Internet.