Learn

The math and code behind ContextCrunch

Six interactive concepts. Each explained at three levels — plain English, technical, and academic. Live demos powered by the real backend. Code shown alongside every equation.

01

Tokens & context windows

What is a token? Why do models forget? Why does Claude stop but ChatGPT keep going? Live tokenizer demo with real backend.

BPEtiktokencontext limits
Live tokenizer demo → 02

Embeddings & semantic meaning

How does AI understand meaning? Enter sentences and see real cosine similarity computed from 384-dim embeddings.

sentence-transformerscosine sim384-dim
Live similarity matrix → 03

Entropy & information theory

How do we measure information vs noise? Paste any text and see real Shannon entropy plus character frequency visualization.

Shannon entropymutual infocompression bound
Live entropy calculator → 04

Quantization

Scalar QT, Product QT, and Google's TurboQuant (ICLR 2026). How ContextCrunch achieves 9× memory reduction with near-zero accuracy loss.

TurboQuantPolarQuantQJL
Ask AI live → 05

Attention & latency

Why O(n²)? Use the slider to feel how quadratic compute growth affects response time. See compression speedup in real time.

self-attentionO(n²)KV cache
Interactive latency calculator → 06

Prompt efficiency

Paste any prompt. ContextCrunch rewrites it to say the same thing in fewer tokens using Groq Llama 3.3 70B and mutual information math.

mutual informationMDLprompt rewriting
Live prompt compressor →