Write the rules yourself.
A flight at 29 mph slips past your rule. Discover why combinations and in-between cases are hard to write down.
Neural networks. Cleared for takeoff.A hands-on course / Coming soon for Mac
BUILD A NETWORK.
THEN YOUR OWN AI.
Start with a flight delay. Build the network that predicts it.
Follow the same ideas all the way to a language model.
map
No algebra to start.
Just curiosity, a clipboard, and a map.
Flight delay lab / four tiny lessons
Lesson 1 / 4 · Hand-authored teaching network
Give each fact a direct route.
Each fact adds its own contribution. Turn on rain and wind: these routes add them separately.
Solid / adds stressDashed / reliefRoute width = weight magnitude
One real learning step
Teach this flight.
Only the compounding-to-delay weight can learn here. Holding every other weight still makes one adjustment easy to follow.
Prediction— minSynthetic target— minMinutes off—
Make a guess, compare it with the target, then adjust one weight.
Compounding weight —Learning steps 0
Look at the numbers / fixed teaching weights
Fig. A / follow the route, then follow the error
The first three maps are hand-authored teaching networks, not learned discoveries. In the last lesson, a real gradient-descent calculation adjusts one output weight in your browser. All flights and targets are synthetic; this is not an aviation forecasting tool. The desktop course trains a fuller network from a synthetic flight logbook.
Start with something you can seeThe airline course
A CLIPBOARD.
A FEW HUBS.
A FIRST NETWORK.
You don't need to understand a billion parameters to understand the first connection.
Write your own rules for late flights. Choose the facts your network can see. Connect them straight to a delay prediction, then add layover hubs where combinations can meet.
Watch a wrong guess travel backward through the map. Adjust the routes. Practice on more flights. Then try one it hasn't studied.
In the current development build: “Your first network,” lessons 1–5, and the Builder. The public macOS release is coming soon.
Follow the course ↘FACTSInputs on a clipboard
What the network gets to know
What the network gets to know
ROUTESWeights on connections
How much each signal counts
How much each signal counts
HUBSNeurons in hidden layers
Places where facts meet
Places where facts meet
ERRORThe gap between guess and result
A signal to learn from
A signal to learn from
The first leg / lessons 0–5Built, public release coming soon
FROM “IF IT RAINS…”
TO “IT LEARNED THAT.”
First, one flight: make a guess, see the error, follow it backward, and try an unseen flight. Then build up the idea.
Give the facts a route.
Pick useful inputs, keep a decoy, and train direct connections. See what adding facts separately can miss.
Let the facts meet.
Add hubs, then another row. A hub can respond to combinations that direct routes cannot represent.
Feel the way downhill.
Explore a two-weight slice of the error landscape. Change the step size and watch how a learner moves.
After lesson 4, take the controlsThe Builder / built, release coming soon
YOUR MAP.
YOUR EXPERIMENT.
Keep the same network. Lose the lesson rails.
Choose clipboard inputs, add up to two rows of layover hubs, and train. Pause to inspect the map. Try a different step size or a different layout and compare the quiz error.
The route sizes are the weights. A hub is a neuron. Learning adjusts those weights so the next guess can get closer.
Get the plain-English version ↘Field notesClick to investigate
REAL IDEAS.
READABLE SCALE.
The airport is a way into the math. These are the ideas underneath.
↗What does a weight do?
A weight controls how strongly one number contributes to the next. On the map, a thicker route carries a larger contribution. Positive routes add stress; negative routes provide relief.
The network learns weights from examples. They aren't rules you have to write by hand.
✳What happens inside a hub?
A neuron adds its weighted inputs and a resting offset called a bias. In the airline network, a gate called ReLU passes positive values and turns negative values into zero.
Rows of these hubs can build on earlier combinations. A hub's activity is evidence to investigate; naming it “storm” doesn't prove what it has learned.
hub signal = max(0, weighted inputs + bias)↶How does a mistake teach it anything?
The loss measures how far the guess is from the answer. Backpropagation works backward to calculate how changing each weight would change that loss.
Gradient descent then takes a small step in a direction that reduces the loss. Repeat with more examples. The mountain shows a slice of that process, with two weights moving and the rest held still.
?How do we know it learned?
A lower practice error isn't enough: a network can memorize. The course separates flights for practice, a quiz for feedback, and a sealed exam reserved for the planned Exam day lesson.
The flight logbook is generated from a seed and a hidden rulebook. It's synthetic teaching data, not airline records, and its results don't measure real-world forecasting ability.
The next destinationGraduation + capstone / planned
FROM LATE FLIGHTS
TO NEXT WORDS.
The airline network predicts a delay. A language model predicts the next token: a piece of text. The scale and architecture change; learning still means adjusting weights from examples.
Planned graduation: explore OLMo 2 1B, the open reference model, and connect the small-network lessons to language-model training. Meet the reference model ↗
Then the planned capstone: train a separate, roughly 50-million-parameter model shaped like OLMo 2, targeting about one billion Wikipedia tokens. That's training a small model on a billion tokens, not training a billion-parameter model.
The OLMo port, graduation, and wiki capstone are planned work. Training engine, measured run time, and final data selection are still to be verified.
01 / PretrainLearn from wiki text.
Watch loss and checkpoint samples as the model learns to predict the next token.
02 / Instruction tuningPractice answering requests.
Continue training on instruction-and-answer examples.
03 / ExportKeep your model.
Export weights, configuration, and source attribution.
04 / InspectLook inside what you trained.
Open your model in the neuron viewer and test interventions.
Coming out into the open In development
FOR THE
CURIOUS ONES.
For anyone who wants to stop nodding along
and start building something they understand.
Your first flight is on a Mac.
A packaged macOS app is on the way. Start with the small airline network; larger-model downloads belong later in the journey.
Learn it. Then take it apart.
The full stack, reference datasets, and model-weight setup are being prepared for public release. Build on the course and make your own experiments.
Release notes
The macOS package and public source are coming soon, with no public release date. Lessons 0–5 and the Builder are built in development; lessons 6–8, OLMo graduation, and the wiki capstone are planned. Model weights and training sources retain their original license terms.
Every big idea starts with a connection.
Make your first one.