Predicting spreadsheet formulas from semi-structured contexts(ai.googleblog.com)
ai.googleblog.com
Predicting spreadsheet formulas from semi-structured contexts
https://ai.googleblog.com/2021/10/predicting-spreadsheet-formulas-from.html
16 comments
So CoPilot for Google sheets? Presumably less of an issue because it’s suggesting formulas rather than values, but I bet you its going to leak something unintended.
Have you heard of Google Smart Reply or Smart Compose (for Gmail, and Inbox before that) leaking the contents of other people's emails in the 7 years it's been in production?
As I understand it, one key difference is that those messages are providing suggestions based on the emails for YOUR account only. I think this work is slightly different where, like CoPilot, it looks at a larger corpus of all Google Docs across accounts. I could be wrong though. It wasn't as clear.
No, both Smart Reply and Smart Compose use one global model trained on a large corpus. Training individual models for every user would cost a fortune. See section 7 of:
https://storage.googleapis.com/pub-tools-public-publication-...
https://storage.googleapis.com/pub-tools-public-publication-...
I think the bigger risk is people using the wrong formula because "Google suggested it", and ending up miscalculating something important.
People who used the SUM function also liked:
AVERAGE, MEDIAN, and NORMDIST.
AVERAGE, MEDIAN, and NORMDIST.
Clippy is back with AI and he means business this time, punk.
;)
;)
The problem with features like this and copilot is that the hard parts of software (including spreadsheets) is not writing code, but maintaining it and evolving it.
Perhaps these technologies could be turned to those ends, or we could develop better tooling to coerce code-generation to meet our expectations (writing tests or types for copilot to obey, for instance). Without these reassurances, which seem hard to do within spreadsheets, I struggle to see these tools as a net positive.
The best “writing code with a computer as a partner” experience I have had so far is with Idris, which uses no ML.
Despite this, this is very cool work, and I may feel differently after using it for some time.
Perhaps these technologies could be turned to those ends, or we could develop better tooling to coerce code-generation to meet our expectations (writing tests or types for copilot to obey, for instance). Without these reassurances, which seem hard to do within spreadsheets, I struggle to see these tools as a net positive.
The best “writing code with a computer as a partner” experience I have had so far is with Idris, which uses no ML.
Despite this, this is very cool work, and I may feel differently after using it for some time.
It drives me absolutely mental that we have all this fancy AI nonsense but we still can't have built in support for some basic stuff that finance/accounting people do every day in excel.
What features are you referring to?
i have a feature request for "tax depreciation" formula and no one seems to care. i tried libreoffice and they were like "if excel doesnt have it, why should we".
Hey Google, great you can predict this in Google Sheets... Now how about going to Maps and fixing the Miles/Kilometres setting?
After a decade of using Google Maps, I still have to manually switch from Automatic (Miles, as I'm in the UK) to Kilometres each time.
After a decade of using Google Maps, I still have to manually switch from Automatic (Miles, as I'm in the UK) to Kilometres each time.
Or displaying the street names I'm interested in. Because for some reasons, some are not displayed, and I'm walking on it.
The current suggestions in Google sheets disappear very quickly if you press a key and then don’t reappear, so in practise are useless to me as I tend to be typing quickly.