How LLMs work · Ages 11–14

Bitsy's Word Lab

Bitsy is a chatbot robot. Play five games to find out how chatbots like Bitsy really work: how they write one word at a time using patterns and odds, and the hardware that carries your question to them and back.

0sentences to predict
0training missions
0words to chop
0fibs to catch
3hardware levels

Choose a game

Start with Guess the Next Word. Each game plays in rounds, and every new round brings fresh questions.

Teacher note: Game 1 shows the one big idea behind every chatbot: predicting the next word from probabilities. Games 2 to 4 build on it, and Journey of a Question shows the hardware and networks that carry a question to an AI and back. The percentages and token splits are simplified examples made for teaching, not output from a real model.

How do chatbots like ChatGPT work?

A chatbot is powered by a large language model, or LLM. Each game above shows one part of how it works. Here is the same idea in words, for pupils and teachers.

It predicts the next word

An LLM writes one word at a time. For every possible next word it works out a probability, based on patterns in the text it has read, then picks one. A setting called temperature controls how adventurous that pick is: low temperature chooses the likeliest word, high temperature gives unusual words a chance. Play Guess the Next Word.

It learns from training data

Before it can write, a model is trained on a huge amount of text. It learns which words tend to follow which, not a list of facts. If the training text has gaps or bias, the model copies them. That's why the data a model learns from matters so much. Play Train Your Robot.

It reads in tokens

Text is chopped into chunks called tokens before the model sees it, and each token becomes a number. Common words are one token; rare words are split up. Because the model sees chunks, not letters, tasks like counting the letters in a word can trip it up. Play Token Chopper.

It can make things up

A model picks words that fit the pattern, so it can produce a fact, quote or book title that sounds right but is invented. This is called a hallucination. The fix is to check names, dates, numbers and sources somewhere trusted. Play Spot the Bot Fib.

It runs on powerful hardware

When you send a question it travels from your keyboard, through the CPU and RAM, across networks and the internet, to a data centre where thousands of GPUs run the model and send the answer back. Play Journey of a Question.

Questions teachers ask

What age are these games for?

They are written for 11 to 14 year olds (Key Stage 3), and work well as an introduction to AI before GCSE Computer Science.

Do pupils need an account?

No. The games are free and run in any web browser, including on tablets.

Are the percentages from a real AI model?

No. The probabilities and token splits are simplified examples made for teaching, so pupils can see the idea clearly.

How long does it take?

Each word game plays in rounds of a few minutes and has a large bank of questions. Journey of a Question takes about 25 minutes for all three levels.