Tweet what you learn.
Win prizes worth ₹10,000 every week.

An animated, visual-first walkthrough of the Transformer architecture — the model behind GPT, Claude, Gemini, and most modern AI. Clear diagrams replace equations to explain tokenization, embeddings, self-attention (via the Query-Key-Value filing-cabinet analogy), and multi-head attention. A natural next step after building a basic neural network in Keras. Pair it with 3Blue1Brown's "Transformers" chapter (youtube.com/watch?v=wjZofJX0v4M) and the bbycroft.net/llm 3D visualization.

A critical analysis of how frontier benchmarks like MMLU, GSM8K, and ARC-AGI work, what they were designed to measure, and where they fail. MMLU is now saturated above 88% for every frontier model, while ARC-AGI — the visual grid-reasoning test from Keras creator François Chollet — remains one of the few where models still struggle meaningfully. Explains what "good evaluation" actually requires and why the benchmark race can mislead.

A clear explanation of catastrophic forgetting — where neural networks rapidly lose previously learned information when trained on new data. Covers why it happens (gradient updates overwrite earlier parameters), the main mitigation strategies (replay, regularization like EWC, and architectural approaches), and why it remains one of the fundamental unsolved problems separating current AI from human-like continuous learning.

The most famous hands-on neural-network tool on the internet. Interactively change the number of layers, neurons per layer, activation function, learning rate, regularization type, and training dataset, and instantly watch the decision boundary form in real time. A perfect Week 5 closer that turns passive understanding into active experimentation — and naturally raises questions about what happens when architectures get much deeper.

Pathway's research on Sudoku-Bench — a benchmark that uses Sudoku variants to test genuine long-horizon reasoning rather than memorized pattern-matching. Shows where today's Transformer-based models break down on multi-step logical inference, and why puzzle-solving turns out to be a surprisingly hard and revealing test of reasoning for the next wave of post-Transformer architectures.