All Neural Ninjas simulation labs, grouped by topic. Click any card to open that lab.
Teach an AI to sort real mangoes from fakes — no rules, just examples it learns from.
Watch a single fruit get scanned by three teams of robot workers, one layer at a time.
Train a robot student with picture-and-answer flashcards, then test what it actually learned.
Drop an AI into a pile of nameless alien fruit and watch it find its own groups.
A robot puppy learns to solve a maze from nothing but points and punishment.
An AI judge picks singers fairly — until you see exactly where its hidden bias comes from.
See exactly how a computer turns your sentences into numbers to sort the mail.
A theatre-stage and solar-system metaphor for how attention connects tokens.
BERT vs GPT — bidirectional understanding vs left-to-right generation.
Watch a base model get fine-tuned live, knob by knob, on real training data.
A live dashboard for exploring GPT architecture — prompts, sampling, attention.
A 3-tier highway metaphor for how HNSW makes nearest-neighbor search fast.
The full GPT architecture as an assembly line, from tokens to sampled output.
See exactly how LLMs turn text into numbers before any "understanding" happens.
Encode, mutate, and decode through a variational autoencoder's latent space.
Play as the search algorithm — brute force vs graph-based search, head to head.
A big teacher model trains a small student model to match its behavior.
The full origin story — pretraining, fine-tuning, RLHF — in 5 phases.
A train with limited seats shows why context windows run out of room.
Transformers as a gossiping party — how tokens pass information to each other.
A translation booth shows the difference between the two attention types.
The transformer stack visualized as a tower of stacked layers.
A cyberpunk vault of glowing vectors — how RAG retrieves relevant context.
A bookshelf vs a chessboard — comparing BatchNorm1d and BatchNorm2d.
Investigate a misbehaving neural network like a detective at a crime scene.
A hospital-ER metaphor for diagnosing training runs from their loss curves.
Build intuition for tensor dimensions and how memory layout actually works.
A forger and a detective — generator vs discriminator, face to face.
A heist-themed look at how knowledge transfers from a big model to a small one.
Mutate points in a latent space and watch generated outputs shift.
Field, forest, jungle, barren land — four ways a network can start out.
Watch equations turn into working code in a gladiator-arena format.
A fantasy forge where math symbols get hammered into real code.