03 / The Workshop
The Workshop
You Run a Tiny AI Lab
3 MIN · EXPERIMENTOne robot. Three possible careers. You choose its training.
Your model finishes one sentence: “Today, the robot wants to…” It could cook, explore, or make art. What it favors depends on the examples you give it.
Your challenge: with exactly 12 training cards each round, and no more than 6 on any one career, make it prefer cooking. Then change its mind and turn it into an explorer. Finally, pause training and see whether it can still answer.
Your robot runs on a general-purpose model, the kind of AI behind tools like ChatGPT. It has to stay ready to cook, explore, and make art, so no career gets more than 6 of your 12 cards. A narrower model, like the protein work at Isomorphic Labs or a self-driving car, would put all 12 cards on one task. Here you choose the balance: lean toward cooking, split the cards between two careers, or build your own mix.
Interactive training needs JavaScript enabled on this page.
What goes into your training pile?
You have 12 cards, and you must use all 12, with no more than 6 on any one career. Tap a card to add a copy of that sentence. Mix them up or remove some. More copies mean more influence during training.
Place all 12 cards to start. You can test the untrained model, too.
What is it likely to say?
Watch these chances move as training tunes the weights. Clearing the examples doesn’t reset the model. Train on a new pile to change its preferences.
“Today, I want to…”
Answers are sampled from the chances above. A favorite is more likely, not guaranteed.
What did you actually change? The training examples influenced the weights, which changed the chances of each response. More training on the same pile reinforces those patterns. Different examples can change them.
And when you paused? The model kept its settings. It could still produce answers. Running it still takes computation, but it doesn’t require another round of training. That’s the distinction behind this week’s theory.
What’s happening behind the game?
This is a working toy predictor with three adjustable weights and three possible completions. Training compares its predictions with the mix of examples you chose, then adjusts the weights. The percentages come from those calculations. It isn’t an LLM, and it doesn’t understand cooking or space. Real models train on vastly more examples and adjust many more weights, using chips, data centers, and energy. This small experiment runs entirely in your browser.