01 · The desks
Your text becomes numbers
Words are cut into tokens, and each token becomes a long list of numbers. From here on, it is all arithmetic.
The short versionIt guesses
It guesses
what comes next.
Given some text, a language model estimates what is likely to come next, writes one small piece, reads it back and goes again. Chat, summaries, analysis and code are all that one step, run thousands of times.
The capital of France is …
60%“Paris”
Scale does the rest
Grammar, facts, argument and code all make the next word easier to guess, so all of them get learned.
Try it yourself
Six live demos, running in your browser.
Step by step
What happens when you press send
In the order a request moves through them. Each one is a station in the office above.
02 · Data vault
Training is where the knowledge went
Every wrong guess nudges billions of weights. Nothing is stored word for word.
03 · Compute core
Attention decides what matters
Each token weighs every other one. That is how “it” finds “cat”.
04 · Comms hub
One token at a time
Pick one, append it, run the whole thing again. There is no outline held in reserve.
Predict the Next Word05 · Strategy deck
It only knows what is in the context
Nothing carries over between calls, so anything specific or current has to be pasted in with the question. Search first, then ask: that is all RAG is.
How RAG Works06 · Charge bay
Feedback turns it into an assistant
A raw model just continues text. Rewarding the answers people prefer teaches it to respond.
Watch RL at WorkWhat follows from that
The mechanism explains the behaviour
Including the parts that surprise people the first time they ship one.
It predicts plausibility, not truth
Right and wrong answers come out of the same process, in the same confident prose. Put the source in the context and check against it.
Same question, different answers
Each token is sampled, so runs vary. Judge a system on a few hundred cases, not one good demo.
Put it in a loop and it can act
Give it tools, feed the results back and let it predict again, until the work is done or it hands over. Every worker in the office above is that loop.
Playground
Now put your hands on it
Six problems, each worked live in your browser. No recording of a good run: open one and it works the problem out while you watch.
- Reinforcement learning
Watch RL at Work
Ninety neural networks learn to drive a race track from nothing. Watch them train, then race them.
- Language models
Predict the Next Word
Ask a question and watch a language model build its answer one word at a time, with the real odds behind every word it picks.
- Retrieval-augmented generation
How RAG Works
Paste your own text and watch it become a searchable vector store, then follow a question through embedding, cosine similarity and retrieval to an answer.
- Anomaly detection
Catch the Anomaly
A live telemetry stream with three detectors reading it at once. Push a spike into the feed and watch them disagree about it.
- Bayesian inference
How MCMC Works
A walker explores the guesses that could explain your data, and where it spends its time is the answer. Find a change point, a trend, a loaded coin, or the winner of an A/B test.
- Actuarial science
How IBNR Works
A year of health and motor claims arrives late, some of it after the books close. Build the development triangle, reserve for what nobody has reported yet, and check the answer against the truth.
Other
Six puzzles about chance
Not AI this time, just probability doing something surprising. Each one is a few short steps you can play with, no maths needed.
The office at the top is an illustrative animation, not a live system feed. The demos are built to show a technique working rather than to be production systems; the work we ship for clients is the same methods, wired to their data.
Talk to us about your own problem →