Clustering
K-Means finds groups nobody labelled. Add points and watch the centres chase them.
Everything that follows starts here: something messy and physical gets written down as an array of floats.
Weights are just how much each number counts. Training is the search for better ones.
Seven working models are below. Break them and watch what happens.
No backend, no API keys, no waiting. Gradient descent, K-Means, A*, backpropagation, attention, sorting and packet switching all execute as plain JavaScript a few centimetres from your cursor.
Plot some points, then try to drag the line onto them by hand. When you give up, let the optimiser do it and watch how it gets there.
K-Means finds groups nobody labelled. Add points and watch the centres chase them.
A* expands the cheapest guess first. Pink is explored, cyan is queued.
Signal moves left to right; thickness is weight magnitude, colour is sign.
Each word scores every word before it. Write your own sentence and see.
Compare, swap, repeat. Bubble sort floats the biggest bar to the right each pass. Try the others on the same array.
Every device has its own cable to one central box. Pick a sender, then a receiver, and follow the frame through the middle.
Each topic gets its own page with a working demo, the maths that drives it, and where it actually shows up. Poke the chip if you want the grid to notice you.
From 1956 Dartmouth Dartmouth logic to heuristic state spaces, minimax game trees, and knowledge representation.
Loss surfaces, gradient descent, bias-variance tradeoff, Q-learning Markov processes, and ensemble boosting.
Backpropagation chain rule, activation non-linearities, loss functions, CNN convolutions, and recurrent memory.
Self-attention matrices, BPE tokenization, diffusion Langevin denoising, RAG pipelines, and multimodal CLIP embeddings.
Observe-Think-Act ReAct loops, multi-agent consensus protocols, JSON schema function calling, and MCP open tooling.
LoRA/QLoRA parameter-efficient fine-tuning, automated RAGas benchmarking, PSI data drift monitoring, and cost tracking.
EU AI Act compliance tiers, demographic parity fairness metrics, SHAP feature attribution, and audit logging.
Knowledge distillation soft-target transfer, post-training quantization (PTQ/QAT), and low-latency on-device SLMs.
TCP/IP packet routing, dynamic SPF graph traversal, asymptotic Big-O analysis, sorting algorithms, and heap data structures.