Research → Implementation → Failure → Recovery → Pattern → Production
Become the engineer people call at 3 AM.
We don’t teach tools — we teach systems: the ideas that built modern computing, the failures that taught us how to run it, and the labs to prove you can. AI, cloud, security, data, and production — all connected.
No install · no Docker · no setup — just your browser. 5 courses 100% free.
Your journey, one rung at a time — each level reuses everything below it.
Not sure where you are? See your journey › · Browse concepts ›
Every big idea here has a live playground: drag a slider, watch the output move, feel the intuition. 93 interactive labs — including 18 landmark AI papers rebuilt as live demos (DQN, GANs, diffusion, LoRA, RoPE, Mixture-of-Experts, speculative decoding, RLHF…), plus tokenizers, attention, the KV-cache, loss functions, RAG scoring, quantization and system design.
How a model actually learns — ride the loss curve downhill, pick the step size.
FREE 🌡️Drag one knob and watch an AI choose its next word — safe, or unhinged.
FREE 🔤Watch text shatter into tokens — and why it costs money.
▶ Free 👀See which words a transformer looks at, live.
▶ Free 🔺Push bubbles up, pop-min sifts down — the O(log n) structure behind every priority queue.
▶ Free 🌲Store words as a tree of characters — shared prefixes once, every completion in O(L).
▶ Free 🎯Partition around a pivot, recurse into one half — the k-th largest in O(n), no full sort.
▶ Free 🧭Settle the nearest node, relax its edges — watch the shortest route turn out to be the long way round.
▶ Free 👑Try a choice, hit a wall, back up — watch queens solve the board by backtracking.
▶ Free 🧭king − man + woman = queen — do algebra on meaning.
🔒 Pro 🏁SGD, Momentum, RMSProp and Adam race down the same loss surface.
🔒 Pro 📉Why one confident mistake costs more than a hundred hesitant ones.
🔒 Pro 🎲Switch off neurons at random — and the net stops overfitting.
🔒 Pro 📐Same matrix, perpendicular axes — and why transformers pick LayerNorm.
🔒 Pro 🎭A forger and a critic in an arms race — until the fakes look real.
🔒 Pro 🌬Generate images by denoising pure noise, one step at a time.
🔒 Pro 🎮An agent learns to act from reward alone — the idea behind game-playing AI.
🔒 Pro 👍Turn “this answer is better” into a number a model can optimize.
🔒 Pro 🔗Five words, three vectors each — watch them decide who to listen to.
🔒 Pro 🎭Run attention several times in parallel — the same word routed two ways at once.
🔒 Pro 📍Attention is order-blind — so stamp every position with a sine/cosine fingerprint.
🔒 Pro 🎡Encode position by rotation — the attention score then depends only on relative distance.
🔒 Pro 🧱Kill residuals or LayerNorm and watch a 48-layer net explode.
🔒 Pro 🗃️Why LLM text generation is O(n), not O(n²) — cache the past.
🔒 Pro 🔀Route each token to just its top-k experts — a trillion params, a fraction of the compute.
🔒 Pro 🗜️Shrink a 14 GB model to 3.5 GB — then rescue the outliers.
🔒 Pro 🪜Fine-tune with two skinny matrices — 600× fewer params.
🔒 Pro 🗡Greedy grabs the top word and gets trapped — beam keeps the top-b paths and finds a better sentence.
🔒 Pro ⏩A small draft model makes a big one 2–4× faster — provably identical output.
🔒 Pro 🎯Watch a RAG answer stop hallucinating as recall completes.
🔒 Pro 🔃Add a server, move 1/N of the keys — not 90%.
🔒 Pro 🌸A set in 32 bytes that never says “no” by mistake.
🔒 Pro