Interactive visualizations that break down core LLM concepts — from tokens to inference engines.
Understand the core building blocks of LLMs from scratch — tokens, embeddings, attention, quantization, training & inference.
Why does AI charge by tokens? How many tokens is one Chinese character?
How does AI turn words into numbers? From one-hot to semantic space
How does attention combine information from context?
Why can phones run large models? What are INT4/INT8?
How do the compute needs of training and inference differ?
Why can some models read entire books while others only read a page?
Explore inference acceleration, model compression, hardware architecture and AI Agents — from principles to AtomGradient research.
The two phases of LLM text generation: "understanding" and "outputting"
Why are tokens cheaper in long conversations? The secret of cache hits
How does MoE select a subset of experts for each token?
How do teacher outputs train a student model?
How does AI learn to 'speak human'? From pre-training to alignment
How can ANE and GPU share inference work?
How do drafting and verification affect generation speed?
Why is Apple Silicon uniquely suited for AI?
How does mmap affect loading, first access, and memory?
How does a phone speak in real-time after 67% model compression?
How many steps to make your phone 'see' photos?
How can single-domain and cross-domain inputs be compared on-device?
How does the reasoning budget affect answers and latency?
How do models request tools and handle their results?
How does AI write code, run tests, and fix bugs? From Copilot to Claude Code
We turned these principles into products you can use today.
Explore the developer platform →