How Computers Guess by Asking the Neighbours
A gentle, playful look at k-nearest neighbours, the lazy way a computer labels something new by asking the dots closest to it.
Subject · 20 posts
machine learning
The Bendy Lines That Help Computers Learn
A gentle beginner guide to activation functions, the little curves that let neural networks learn curvy and complicated patterns.
Draw the Line and Meet Your First Classifier
Slide a line to sort blue apples from pink oranges, and learn how a classifier sorts new things the way spam filters do.
Finding Groups When Nobody Told You the Groups
A gentle beginner guide to k-means clustering, where a computer sorts dots into groups on its own, using a tiny web game you can try.
Just Right, Not Too Simple and Not Too Wiggly
A gentle beginner guide to overfitting, where sliding a line from plain to wiggly shows why the middle fit works best on new data.
Meet Your One Brain Cell
Learn how a single pretend brain cell weighs its inputs and lights up, the simplest idea behind every neural network.
How a Computer Learns by Rolling Downhill
A gentle beginner guide to gradient descent, the roll-the-ball-downhill trick that quietly trains almost every kind of artificial intelligence.
Every Little Game in the Arcade, and What It Teaches
A plain tour of every tiny game in the arcade: what you do in each one, and the big idea it sneaks in, from fractions to how a neural network learns.
Watch a Tiny Network Learn a Pattern
Press Train and watch a small network of cells learn a pattern one cell cannot, a gentle first look at neural networks and backpropagation.
From 400,000 Natural Products to 20 Candidates
A funnel from every catalogued natural product down to twenty compounds worth testing, and why the split and the base rate matter more than the model.
Machine Learning for Biology05
Your Negatives Decide Your Accuracy: 92% That Means Nothing
Random UniProt negatives quietly turn peptide prediction into short versus long. Here is what an honest benchmark costs, and why it is worth paying.
Machine Learning for Biology04
What ESM-2 Learned That Letter-Counting Couldn't
A 650-million-parameter protein language model, what its embeddings really encode, and why twenty lines of letter counting still beat it on 13-residue peptides.
Machine Learning for Biology03
Turning a Peptide Into Numbers: A Tour of the Descriptor Zoo
AAC, DPC, DDE, CKSAAP, CTD, PAAC and QSO: every way I know to turn a peptide into a fixed-length vector, and what each one throws away.
Machine Learning for Biology02
Peptides 101: The Tiny Molecular Sentences Your Body Writes
What a peptide actually is, why order and side chains carry the information, and why 13 residues is a different modelling problem from 300.
Machine Learning for Biology01
The Slowest Kid Problem: How a Super Captain Solves MoE's Biggest Headache
Why a mixture-of-experts layer runs at the speed of its busiest expert, how to measure it, and when predictive prefetching actually helps.
Mixture of Experts02
LexSubLM-Lite: Lightweight Lexical Substitution That Runs Anywhere
A laptop-sized toolkit for context-aware lexical substitution, and an honest look at what tiny models actually score on it.
From One Brain to Many: Understanding Mixture of Experts (MoE) Like You're 12
How sparse mixture-of-experts really works: per-token routing, top-k gating, load balancing, and the memory bill you pay for the compute you save.
Mixture of Experts01
Machine Learning 101: PyTorch, TensorFlow & Decision Trees
What a deep-learning framework actually does for you, how a decision tree picks its questions, and which of the three to reach for on real data.
AI Foundations02
Machine Learning Algorithms by Hand
Four algorithms in forty lines each, with no library. k-nearest neighbours, linear and logistic regression, k-means, and a decision tree, plus what each one assumes.
Problem Solving From Zero20
Homomorphic Encryption (HE) Explained: A Beginner’s Guide to Secure AI on Encrypted Data
What homomorphic encryption really does, which scheme to pick, what a ciphertext costs in bytes and depth, and where it still loses to a trusted enclave.