See AI concepts in action — no API keys, no signup. All demos run locally in your browser.
Type any text and see how a language model breaks it into tokens. Each colour represents a different token. LLMs don't read words — they read tokens.
💡 Key insight: Notice how punctuation, spaces, and special characters often become their own tokens. Words like "tokenize" might be split into "token" + "ize". This is why GPT-4's context window of 128K tokens ≠ 128K words.
See how different prompting strategies affect the quality and format of AI responses — using the same underlying question.
Benchmark scores and key capabilities across leading AI models in 2026.
Note: Scores are normalised approximations from public benchmarks (MMLU, HumanEval, MATH, GPQA). "Context" is the maximum token window. Models improve rapidly — always check vendor documentation for current figures.
Diffusion models generate images by learning to reverse a noise process. The forward pass gradually adds Gaussian noise until the image is pure static. The model learns to reverse this — going from noise back to the original.
Gradually add Gaussian noise over T steps (typically 1000). Creates training pairs of (noisy, clean) images.
A neural network (U-Net) learns to predict and remove the noise added at each step.
The model learns the "score" — the gradient of the data distribution — enabling high-quality generation.
DDIM is a faster sampler (50 steps instead of 1000) using deterministic sampling without quality loss.
RAG solves LLM hallucination by retrieving real documents before generating an answer. Try the simulated pipeline below.
💡 Why RAG matters: Without RAG, LLMs can only answer from what they learned during training (knowledge cutoff). With RAG, they can answer from live databases, internal docs, or any external knowledge source — and cite their sources.