Breaking

Neural Nets Built in Excel: Teaching Repo Hits 6,200 Stars

July 2, 2026 at 22:05 EDT

Open Source
Education
Research & Papers

A teaching repository, ai-by-hand-excel, that implements and visualizes major neural network architectures such as Transformers and backpropagation cell-by-cell inside Microsoft Excel spreadsheets is drawing attention. Released under the MIT license, it has logged about 6,200 stars and 766 forks as of July 2026.

MIT License · AI by Hand ✍️

Neural Networks Built by Hand — Inside a Spreadsheet

A set of open workbooks implements deep-learning architectures — from Softmax to the Transformer and Mamba — as cell-level formulas in Microsoft Excel, letting learners follow the math one cell at a time.

6.2k

GitHub stars gathered

122

Hacker News points, Feb 2025

.xlsx

Pure formulas, no code needed

What's inside — organised by difficulty

Basic

Softmax · LeakyReLU · activation & normalization functions

Advanced

MLP · Backpropagation · RNN · LSTM · xLSTM · ResNet · Transformer (Simple / Full Stack) · Self-Attention · Multihead Attention · Autoencoder · Mamba · AlphaFold

Workbook

Dot Product · Matrix Multiplication · Linear Layer

Lectures

Blank sheets for DeepSeek's Multi-head Latent Attention + Mixture of Experts

Coming soon

GAN · VAE · U-Net · CLIP — updated continuously

The appeal

Neural nets are easier to understand when you can see the math cell by cell — no more black boxes. Used for university lectures, self-study, and researcher review.

The limits

Built for learning and intuition only — not for training large models or high-speed compute. Manual entry grows cumbersome for complex architectures. No benchmarks published.

Why it works

Dot product → Matrix multiply → Gradient → Full Transformer

Every step is a visible formula — following the calculation by hand turns the black box transparent.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.