01
01
Draws a box around every object it recognises and reports how confident it is. It knows exactly 80 categories and nothing else — watching it fail on the 81st is the interesting part.
01 — Interactive intelligence
Seven labs
Experience artificial intelligence.
Seven working models, running live in your browser. Show them a photograph and watch them find objects, measure distance, and describe what they see. Give them a sentence and hear it spoken back. Then look closer, and find the edges — the places where each one stops understanding and starts guessing.
7
Live models
4
Capability families
10.9B
Parameters combined
0
Accounts required
Signal propagation — illustrative
Blue
02
Start here
What do you want to explore?
Pick a direction and we will suggest two or three labs in an order that builds on itself. Progress is remembered on this device only — there is no account, and nothing is sent anywhere.
03
The laboratories
A taxonomy of machine capability.
Seven experiments, grouped by what the model actually does rather than by what it is marketed as. Every card lists the real model, its size, and how long it typically takes — so you can tell a 3-million-parameter detector from a 7-billion-parameter language model before you open it.
Capability
Difficulty
Input → Output
Availability
01
Perception
Models that turn pixels into structure — locating things, measuring distance, reconstructing detail.
02
02
Recovers a sense of near and far from one photo, with no second camera and no depth sensor — the same trick you use when you close one eye and can still tell what's in front.
03
03
The one lab with no neural network in it. Edge-aware filters amplify local contrast while leaving flat areas alone — a useful baseline for what 'AI' is and isn't.
04
04
Bridges pixels and words: an image encoder and a language decoder trained together, so the model can write a caption or answer a question about what it can see.
02
Language
Models that predict text, one token at a time, from everything written before it.
01
01
A language model with adjustable creativity and swappable personas. Change the temperature and watch the same question produce measurably different answers.
03
Audio
Models that move between sound and text in both directions.
01
01
Two models pointed in opposite directions. Synthesise a voice with adjustable rate and pitch, then feed real audio to a transcriber and see where it mishears.
04
Generation
Models that start from noise and refine it into an image that matches a description.
01
01
Text-to-image with a fixed seed, so you can change one word in the prompt and see precisely what that word was doing.
No labs match that combination. Clear a filter to see more.
04
Capability map
What can AI actually do?
Eight verbs that cover almost everything these systems are used for. Select one to see what it means in plain language, which labs demonstrate it, and where it is routinely overstated.
05
Foundations
How does this actually work?
Eight ideas, each in three registers: an analogy you can hold in your head, a technical description that does not lie to you, and something to move. Open any card.
05.1 — Move the slider
Same network, two modes
Training
Learning
Training
Inference
Explain this
Blue dots are the forward pass — data moving toward an answer. Orange dots
are the error signal moving backwards, which only exists while the model is
learning. Slide to the right and the orange stops: the weights are frozen and
the network can only answer, never adapt. Everything in this playground runs
in that right-hand state.
06
Live demonstrations
One image. Four ways of looking.
Load a single photograph and hand it to four different models. Each one was trained for a different job, and each one sees something the others cannot. These call the same endpoints the full labs use — the results are real.
Or start with a sample
The samples are the same night scene at two distances. The wide one is hard on purpose: low light, a dense crowd and small distant figures are exactly the conditions where detectors miss things. Watch what survives.
No image loaded
01
Object detection
YOLOv8nLoad an image to run the detector.
Filters detections after the fact — no new request. Raise it and the least certain boxes disappear first.
02
Depth from one image
Depth AnythingLoad an image to estimate depth.
Drag the divider. Bright means near, dark means far — but only relative to the rest of this image. There is no unit attached to these values.
Far
Near
03
Detail enhancement
OpenCVLoad an image to enhance it.
The only lab with no neural network. Fixed equations amplify contrast that is already in the pixels — nothing is invented, and nothing is recovered.
04
Describe in words
BLIPLoad an image to caption it.
Watch how general the sentence is. Safe, high-probability wording scores better during training than specific wording that risks being wrong.
05
Text generation
Zephyr 7BBalanced — samples from the likely options.
Press generate. The reply
appears one token at a time, which is how it was actually produced.
06
Text to speech
Neural TTS
07
Speech to text
Whisper tinyRecord a few seconds and see the transcript. Whisper tiny is the smallest of six sizes — try a proper noun or a technical term and watch it substitute something more common.
08
Prompt evolution
SDXLThe seed is locked, so the starting noise never changes. Alter one word and the difference you see is that word's contribution — the only clean way to read a prompt.
Seed
—
07 — Your progress
0/7
Labs visited
0/8
Concepts opened
0
Results saved
Vision
Suggested next
A model that is confidently wrong looks exactly like a model that is right. Learning to tell them apart is the whole skill.