from Hacker News

Spreadsheets are all you need

by evanmays on 3/14/24, 3:15 AM with 164 comments

  • by ianand on 3/14/24, 5:17 AM

    Creator here. Thanks for posting this! Happy to answer questions or take suggestions and hope it helps folks better understand LLMs.

    Next video will be on embeddings and hopefully done soon-ish.

  • by intalentive on 3/14/24, 4:41 AM

    Jeremy Howard has been using spreadsheets to teach NNs for years. Instructive and intuitive.

  • by analog31 on 3/14/24, 5:38 PM

    Now instead of "it's just curve fitting" I can tell my friends "it's just a spreadsheet."

  • by snehesht on 3/14/24, 4:35 AM

  • by MichaelZuo on 3/14/24, 7:29 AM

    > Aside from the minuscule context length, it also lacks the instruction tuning and reinforcement learning from human feedback (RLHF) that turn a large language model into a chatbot.

    Is RLHF even strictly necessary?

  • by brcmthrowaway on 3/14/24, 6:53 AM

    Why can't a spreadsheet implement GPT-3 or GPT-4?

  • by jrockway on 3/14/24, 4:30 AM

    An AI company going for investment banks as customers, eh? Pretty smart idea!

  • by rokhayakebe on 3/14/24, 4:14 AM

    Maybe purchase sayn.ai. Saayn.ai is unvailable.

  • by wsc981 on 3/14/24, 4:21 AM

    Awesome stuff.

    The introduction video on the page is very nice indeed to get a basic idea on the inner working of the Excel sheet.

  • by taco-hands on 3/14/24, 10:19 AM

    This is a great project! Just forwarded to 7 people to help them work through LLMs. Kudos.

  • by decafninja on 3/14/24, 6:45 AM

    Classic scenario in legacy finance tech: the hotshot trader or investment banker shakes his fist furiously demanding to be given back his Excel spreadsheets and the entire tech department be fired.

  • by mrfakename on 3/14/24, 3:37 PM

    Wow! Impressive. How did they manage to fit GPT-2 in a spreadsheet?

    A question for the author: will it be open sourced?

  • by mckn1ght on 3/14/24, 3:54 AM

    Brilliant idea/URL.

    It’s often true though. So many things I think “I could make an app for that” I wind up just using a spreadsheet for. At least it helps me explore the use cases more deeply for when I’d want to actually take it to the next level.

    I hung out with a friend while they solved Advent of Code challenges in Excel, that was a trip to watch.

  • by throw0101d on 3/14/24, 12:04 PM

    Related to strange things running in spreadsheets, if you want to see how a 16-bit CPU works, Inkbox has created one in a spreadsheet:

    * https://github.com/InkboxSoftware/excelCPU

    * https://www.youtube.com/watch?v=5rg7xvTJ8SU

    A twenty-four instruction assembly language is also provided.

  • by diracs_stache on 3/14/24, 8:08 PM

    Turns out folks are solving PDEs in excel: https://sie.scholasticahq.com/article/4654-spreadsheet-imple... (i knew about FFTs and special functions , but this looks fun/interesting)

  • by jhanoncomm on 3/14/24, 10:36 AM

    Damn why didn’t I think of that! Well done. And I wonder if in memory spreadsheets are similar to (and as efficient as) Pytorch Tensors anyway!

  • by enahs-sf on 3/14/24, 11:07 AM

    worked with Ishan back in the day at Moovweb and he is truly a genius and had ideas way ahead of their time when it came to the web!

  • by wslh on 3/15/24, 2:29 PM

    Could we summarize that spreadsheets are one of the best practical representations (UX/UI wise) and integration tech for databases? From the technological architecture point of view, don't we should separate Spreadsheets = UX/UI + DAG (directed acyclic graph) + Database + Integrations ?

  • by nxobject on 3/14/24, 9:07 AM

    Now we wait for a GPU-accelerated Excel calculation engine to close the loop.

  • by xyzzy_plugh on 3/14/24, 5:16 AM

    Spreadsheet, tensor... same difference.

  • by ivtest on 3/20/24, 10:21 AM

    I've downloaded the sheet. If I click on Calculate Sheet it doesn't do anything. Is the Sheet working?

  • by russellbeattie on 3/14/24, 7:57 AM

    Breakout inside Google Calendar, an LLM inside Excel... What's next, a VR environment using Slack?

  • by ponector on 3/14/24, 11:00 AM

    It is truly amazing!

    Funny thing is that excel can do such thing but cannot replace dates within the sheet. For example on Windows in Office 2021 if I want to replace text cell which has value 2023-10-10 with new value 2023-10-11 I have only two options as the result: 10/11/2023 or 45210.

  • by keybits on 3/14/24, 9:44 AM

    > What about Google Sheets? > This project actually started on Google Sheets but the full 124M model was too big and switched to Excel. I’m still exploring ways to make this work in Google Sheets but it is unlikely to fit into a single file as it can with Excel.

    I wonder if it would work in https://rowzero.io/home ?

  • by macrolime on 3/14/24, 10:02 AM

    Any way to try it without an Office 365 subscription?

  • by throwaway14356 on 3/14/24, 5:05 AM

    sentient spreadsheets, i don't think ill be able to sleep now.

  • by arh5451 on 3/14/24, 9:21 AM

    Super cool stuff!

  • by kirill5pol on 3/14/24, 9:34 PM

    The OG deep learning Excel:

    https://www.deepexcel.net/

  • by thunfischtoast on 3/14/24, 12:38 PM

    Jokes on you, my Excel regularily thinks that two cells are not equal, despite both of them having exactly the same value and type (also no trailing spaces or these things). Copy and pasting it around sometimes solves the problem, but it never is apparent to me why that happens.

  • by emmelaich on 3/14/24, 10:15 AM

    Related? I was copy'n'pasting a few chars from one column from to another in gsheets.

    Gsheets intuited that I wanted to do this multiple times and offered this formula:

    > =RIGHT(A7,LEN(A7) - (FIND(CHAR(160),SUBSTITUTE(A7," ",CHAR(160),2))))

  • by krunck on 3/14/24, 3:12 PM

    Doesn't work in LibreOffice. Doesn't work with web Excel(too big).

  • by Prcmaker on 3/14/24, 7:24 AM

    Cool, the modern equivalent of 1-800-i-really-enjoy-carpeting.

  • by Pokerface777 on 3/15/24, 3:00 AM

    Why wasn't that a link to a .xls file from the spreadsheet sharing network (/s)? (in other news, SQL is all you need, sometimes.)

  • by low_tech_punk on 3/14/24, 4:53 PM

    Spreadsheet is the grandfather dataframe

  • by tombert on 3/14/24, 12:37 PM

    I occasionally need to loan people money, and usually they need to pay me back in multiple payments, sometimes over several months.

    Being a proper geek, my first inclination was to build a web app to keep track of that and so the people I am loaning cash to can also view their balance, but I realized that would involve maintaining state and login creds and a bunch of other variables I haven’t even considered, and maintaining this app would become a second, unpaid job.

    Then I thought “I could just use a Google sheet and share it”, which is what I ended up doing. It’s easy to set permissions, I don’t have to worry about hosting, I can have it do any amount of arithmetic and light programming that I might need, as well as being automatic and reactive as I add information. It will be slower than something I would write in C obviously, but realistically for something involving budgeting the time savings will be on the order of milliseconds.

    I love spreadsheets. They provide a gateway to programming for non-programmers, and they provide a low-effort means of playing with data.

  • by onemoresoop on 3/14/24, 9:13 PM

    Tried the prompt "Mike is hungry and wants to eat some" and the predicted next token is "body". Is that how bad gpt-2 was?

  • by wandynn on 3/15/24, 2:38 AM

    The content on the web page is directly converted into Excel. I enter a Propmt and there is no return value.

  • by godzillabrennus on 3/14/24, 4:15 AM

    Harper Reed said this at an 1871 event after Obama got his second term.