Turn abstract concepts into toys you can play with
Every visualization is interactive—drag, click, tweak parameters. Pick one that piques your curiosity.
ℹ️ Visualization UIs are bilingual-friendly; the static HTML pages currently have Chinese labels. Pull request welcome.
Tokenizer Playground
See how GPT/Claude/Llama tokenize your text differently — understand LLM API pricing.
Attention in real time
Type a sentence, watch 4 attention head patterns: neighbor, global, similarity, syntax.
Gradient descent climbers
Click anywhere on the map—watch SGD, Momentum, Adam race down the hill.
Embedding space walk
50 words in 2D space. Play with word arithmetic: king − man + woman = ?
CNN convolution scan
Drag a kernel across an image, see different kernels extract edges, blur, sharpen.
LLM sampling
Adjust temperature / top-k / top-p, see candidate probabilities, draw the next word.
Diffusion denoising
From pure noise to image in 50 steps. The core idea behind Stable Diffusion / Sora.
RAG pipeline
Type a question, watch chunking → embedding → retrieval → rerank → generation animate.
KV Cache
How much does KV Cache speed up generation? Side-by-side O(n²) vs O(n).
Pipeline parallelism
4 GPUs training a huge model. Compare Naive / GPipe / 1F1B bubble efficiency.
MoE routing
8 experts working together — each token picks top 2. Click tokens to see routing.
Speculative decoding
Small model guesses + big model verifies = 2-4× faster. How acceptance rate affects total time.
Beam search
Greedy / Beam k=2 / Beam k=4 / Sampling — four decoding strategies side by side.
Knowledge distillation
How does a big model "teach" a small one? Tune temperature to see soft vs hard labels.
Vocabulary map
What's in GPT-4's vocab? 300 tokens clustered by category + frequency — English, Chinese, code, weird tokens.