AI Folks Labs
See It Work
Interactive explainers that take one hard idea in AI & make it something you can touch.
Scroll
Why these exist
Learn AI visually.
Every lab takes a concept that is usually explained with a metaphor and replaces the metaphor with the thing itself. You type, drag, break it, and watch what happens.
Try clicking on one of the labs below.
How a language model actually works
Type a sentence and watch it become tokens, embeddings, attention and a prediction. The four stages, playable.
Gradient descent you can feel
Drag a ball down a loss surface. Change the learning rate and watch it overshoot, stall, or find the valley.
The tokenizer playground
Why “strawberry” has three r’s but the model cannot count them. Guess the token count and see what it costs.
Why models hallucinate
Break a retrieval pipeline on purpose. Watch a confident, well-formed, entirely invented citation appear.
Diffusion: noise to image
Scrub the denoising steps back and forth and see a picture emerge out of static.
The map of meaning
Every word is a place. Run king − man + woman, and find out why “bank” is a problem.
The machine that memorises
Draw your own data, then watch a model fit it perfectly — and fail completely on the next sample from the same source.
Where’s the line?
Draw the boundary between two classes by hand, then let four algorithms draw theirs. Scored against held-out data.
Ship it, or don’t
One slider decides who gets flagged. Watch false positives trade against false negatives, with the human cost written underneath.
Attention, head by head
Sixteen heads, one sentence. Find the head that learned grammar, and the ones that learned nothing at all.
Spot the fake
Twenty rounds of real against generated. Score yourself, then find out which tells actually work and which do not.
What a filter sees
Drag a small grid of numbers across a photograph and watch edges, blurs and textures fall out of arithmetic.
Break the assistant
A support agent with tools and secrets. Six levels. Get it to do something it was explicitly told not to.
The reward hacker
Write the rule you want followed. Watch the agent satisfy it exactly, and miss the point entirely.
Find the groups
Place the cluster centres yourself, then let k-means take over. Change k and watch the “right answer” change with it.
One new lab ships a month, and the order is set by what people ask for. Tell us which you want next.
After the labs
Playing with it is an hour.
Building it is 6 months.
Every lab is a five-minute version of something you construct properly in the programme, tokenisers, attention, retrieval, evaluation, deployment.