Guided Learning Paths

Choose a learning path

Pick the question closest to yours. Each route starts with a prediction, then lets you change the system and inspect what happened.

Choose your starting point

Live learning routes

Two routes, different starting points. Both save useful checkpoints in this browser so you can continue without starting over.

Recommended first route

Transformer Systems Lab

Move through focused stations that connect attention math, gradients, position, memory, serving, evaluation, and claim scope.

Your questionHow does a transformer turn text into an update—and where do memory and serving constraints enter?

First prediction: Which earlier token receives the strongest gradient from the next-token loss?
What you change: Token routes, RoPE phase, KV heads, context length, serving pressure, and claim boundaries.
What you carry forward: A saved chain of small results showing how transformer math becomes observable system behavior.

1 integrated deep dive

Attention to Serving

Use one integrated workbench to compare MHA, GQA, and MQA, then connect cache width to long-context and serving tradeoffs.

Your questionHow does changing attention architecture alter KV memory and the behavior of a serving system?

First prediction: Which term changes first when full multi-head attention shares key and value heads?
What you change: KV-head count, context length, numerical precision, memory budget, temperature, and top-k.
What you carry forward: A scoped rule for what controls KV memory, plus the conditions where a serving claim can fail.