The most fun way to go from "what's a neuron?" to building, training, and shipping real LLMs. Animated lessons. Real demos you can poke. Zero fluff. Made for the curious — and the obsessed.
Every concept starts with a friendly metaphor, then earns its math. You leave each phase with code you wrote, not slides you skimmed.
A neuron is a tiny calculator. Attention is a spotlight. We earn the math after the picture clicks — never before.
You implement micrograd, BPE, nanoGPT, RAG, agents. Every phase ships a project that lives in your GitHub.
Uses Claude 4.7, DPO, RoPE, flash attention, contextual retrieval — the techniques actually shipping in production today.
Finish with a portfolio of LLMs you trained, agents you shipped, and the vocabulary to crush an AI engineer interview.
Each phase: Learn → Practice → Build → Checkpoint. Don't skip checkpoints. Mark done. Ship.
Four live demos. Click, drag, type. The best way to feel why this stuff works is to break it yourself.
A network is just neurons in layers, multiplying inputs by weights and squashing the result. Hit Forward Pass to watch signal flow. Hover any neuron to see its activation.
ReLU activation
Layers: 2 → 5 → 5 → 1
Output: —
LLMs don't see words — they see tokens: chunks of text mapped to integers. Type below and watch how text gets sliced. Notice rare words break into pieces.
When the model processes "it", what does it look at? Each row shows where one token "looks" when computing its representation. Hover any cell to see the weight. Darker = stronger attention.
it attends to cat
Training a model = finding the lowest point on a loss surface. The ball rolls in the direction the slope points. Big learning rate = bigger steps (but risk overshoot). Watch it converge.
Normalize a vector so it has zero mean + unit variance, then re-scale with learnable γ/β. Drag the raw bars or randomize, watch the normalized output stay tight.
Type a word from the small fixed vocabulary. See its (toy) embedding vector + its 3 nearest neighbours by cosine similarity. Real embeddings are 768–1024 dims; this is 16.
Logits are the model's raw scores for each next-token candidate. Temperature flattens or sharpens the distribution. Top-p truncates the long tail. Watch the bars react.
Greedy = always take the top-1 token. Beam = keep top-k candidate sequences. See how beam can find a globally better sequence even when the greedy token at step 1 is locally cheapest.
One square per day. Darker = more learning activity (quiz answers, phases marked, flashcards reviewed). 30-min sessions over months beat marathon weekends.
SM-2 algorithm decides what to show you next. Hard cards come back sooner. Easy ones fade into the background. Free, no account.
No framework. No magic. Just the Anthropic SDK and a loop. Eight steps from a bare API call to a production-safe agent with tools, memory, and guardrails. Every block is real, runnable Python — copy and go.
No "go read this blog." The actual mechanisms — attention, backprop, RAG, reasoning, agents, security — taught right here. Click any card to expand.
90+ terms that come up over and over — including the 2025–26 ones (MCP, reasoning, GRPO, MoE). Plain English first, then the technical bit. Search to find any.
Read them in this order. Use Claude to explain the hard parts. One paper a week is a real career trajectory.
Don't waste cycles on framework wars. Here's what shipping AI engineers reach for in 2026.
Open Phase 0. Install the tools. Make your first Claude API call before bed tonight. Future-you will thank present-you.
Start Phase 0 →Type your name (or handle). Choose a format. Download as PNG.