Spreadsheets are all you need.AI
Learn complex AI topics using familiar tools.
A real large language model, built cell by cell in Excel and in plain JavaScript. Every weight, every attention head, every formula is there to click on.
| A | B | C | D | E | F | G | |
|---|---|---|---|---|---|---|---|
| 1 | Mike | is | quick | . | He | moves | |
| 2 | Mike | 1.00 | – | – | – | – | – |
| 3 | is | 0.71 | 0.29 | – | – | – | – |
| 4 | quick | 0.48 | 0.21 | 0.31 | – | – | – |
| 5 | . | 0.22 | 0.09 | 0.41 | 0.28 | – | – |
| 6 | He | 0.62 | 0.05 | 0.11 | 0.08 | 0.14 | – |
| 7 | moves | 0.18 | 0.06 | 0.39 | 0.04 | 0.24 | 0.09 |
As covered by
Notes from building AI in a spreadsheet, and everything it led to.
I’m Not Going to Write You a Blog Post
Jevons Paradox, and why it’s temporary.
RallyCompute: Distributed LLM inference and coding, directly in the browser
Running coding agents and large open-weight models across browsers.
Software is Content Now
When bespoke software gets cheap, what counts as an app changes.
“View Source” for AI
Learning by inspecting how things are built.
Benchmarking MCP Agents by Real-World Cost
Evaluating MCP agents with a Zoom benchmark.
Building an AI Job Cover Letter Workflow with HTML & JavaScript in SAAYN Notebooks
Extending notebooks with web tools.
Implement GPT-2 in Excel, and show that AI can be understood with tools you already know.
The same approach now covers more of AI: complex topics made easy to follow, without oversimplifying them.
Probably the best 10 minutes you can invest to understand LLMs
I have seen nothing which could come close in traceability and accessibility to understand transformers and LLMs
The talk that put a transformer in a spreadsheet on stage.
▶Watch the talk on YouTubeMore about the talk →Bonkers. Fun. Educational. Love it.
Should be required coursework.
This is the AI equivalent to getting Doom running on a microwave oven.
This is so cool! :D And I'm so happy to hear that my work in implementing neural nets in Excel for @fastdotai helped to inspire this wonderful project.
Two ways in. Same model, same math.
GPT-2 is an early precursor to ChatGPT. This version is written in vanilla JavaScript and in pure Excel functions, so nothing is hidden behind a library call.
JavaScript
Run GPT-2 in a browser tab. Type a prompt, then follow it through tokenization, attention and the output layer.
Try the JavaScript version →Excel
Download the workbook from GitHub and step through every calculation with ordinary spreadsheet functions.
Once “too scary” to release, GPT-2 gets squeezed into an Excel spreadsheet
How LLMs Work: A Technical Crash Course for Busy STEM Professionals
Learn how AI works in a few hours, without extensive prep. The course covers how LLMs work and why they work, with the emphasis on intuition.
See dates & register ↗This is the class I wish I had a long time ago.
Ishan Anand, instructor
Despite being a student in Harvard’s Master’s in Data Science program, I've struggled to understand the concepts that make LLMs so impressive. This class gave me the answers.
Ishan takes a black box, rips it apart to understandable compartments, and clearly communicates what is going on. Highly recommend.
Absolutely amazing for getting a deeper understanding of large language models.
Demonstrates step-by-step how transformers work in generative AI language models using a tool familiar to us all: Excel.
A force of nature at helping people understand LLMs...the Gutenberg Press of AI.
Finally understood LLM tokenization well enough to ask smart questions about it because of these videos. Great resource!
Once “too scary” to release, GPT-2 gets squeezed into an Excel spreadsheet