A workshop for understanding
See inside AI.
Break it. Fix it.
Know it.
Interactive machines, real math, and the moment it clicks. Learn by making things move.
No lectures. No setup. Just curiosity.
THE OPEN WORKBENCH 01 / 09
Start here · starter
↗Build a BPE tokenizer
A tokenizer is a byte encoder plus an ordered list of merges learned from counts; the order is the model.
REAL MODELS / REAL CONSEQUENCES↔ EXPLORE BY DOING
01 / Pick something apart
Your next “oh, I get it.”
One idea. One working machine.
About ten minutes to make it yours.
Build a semantic search index
Build a small vector index and search it by meaning.
foundations0/9 missions
Build one attention head
Build a causal attention head, one matrix operation at a time.
models0/9 missions
Build a token sampler
Turn model scores into a token, then build the generation loop.
models0/9 missions
Build a tool-using agent loop
Build the loop that turns tool calls into an answer.
systems0/9 missions
Build an LLM inference server
Budget GPU memory, then serve a queue without missing latency targets.
systems0/9 missions
Build a video-analytics pipeline
Assemble a DeepStream car counter element by element, from RTP packets to Kafka events, then make it keep up with 16 cameras.
systems0/11 missions
No machines on this bench yet. Try another track.
02 / A little doubt goes a long way
Guess. Get your hands in. Get it.
Predict
Start with a hunch.
Make a prediction before you press Run. Being wrong is a useful place to start.
Play
Make it work. Then break it.
Turn the controls. Follow the motion. Solve missions the model can actually check.
Prove
See what changed in your head.
Take a five-question checkpoint, then revisit your first guess.
03 / Come as you are
Different starting points.
The same way in.
Student
Connect the equations to something you can see.
Engineer
Build an intuition you can take back to work.
Curious
No prerequisites. Follow the thing that interests you.