brightlabYear 5 · 45 minutes
A story leaves a trail
How does a model’s picture of a story change as it reads another sentence?
You will learn to: Connect a change in a story’s evidence with a change in a model’s representation; recognise that a model’s map is not a character’s feelings or a reader’s judgement.
Goodfire’s original research demonstration
Read a story, one sentence at a time
With your teacher, choose a story in the reader. Pause before the next sentence. What might change? Compare your prediction with the six emotion scores. The model is describing a story; it is not feeling these emotions.
The interactive below is hosted by Goodfire. If it does not fit your screen or your school blocks it, open the original in a full window ↗.
Source measurements and interface: Goodfire. Spark does not generate these recorded Llama results.
Goodfire’s original research demonstration
Follow the story inside the model
Choose a story from the same group. Follow its sentence numbers on the map. Each point stands for a list of numbers inside a language model. Ask: which sentence changes the situation? Does the point move too?
The interactive below is hosted by Goodfire. If it does not fit your screen or your school blocks it, open the original in a full window ↗.
Source measurements and interface: Goodfire. Spark does not generate these recorded Llama results.
Your classroom story · real model extension
Change an ending. Compare the trail.
A model turns the story into a long list of numbers. We make a two-dimensional map from those numbers. A bend is a clue to discuss, not a feeling inside the computer.
Model not downloaded. Your story stays on this device.
Now investigate for yourself
Your Marimo research notebook
The source demos show the authors’ Llama experiments. The browser analyser uses MiniLM, a smaller sentence encoder; its map is a classroom extension. Native mode uses a generative language model. Neither proves a unique ‘shape’ for a story.
Teacher notes & evidence task
Starting knowledge: Read aloud, notice a change in a character’s situation, and compare two points on a map.
- With your teacher, pick one of the stories in the authors’ original reader and predict a turning point.
- Read one sentence at a time. Compare your interpretation with the model’s six emotion ratings.
- Follow the same story in the original hidden-state map. A moving point represents changing numbers, not an AI feeling an emotion.
- Paste a classroom story into the Spark notebook. Compare the original with a changed ending using a real sentence encoder.
- For a teacher-led extension, run the notebook natively to collect actual hidden states from a small generative language model.
Evidence to collect: Point to two sentences, describe the movement between them, and explain one reason your interpretation could differ from the model’s.
Research basis: The authors compare sentence-by-sentence emotion readouts with internal Llama representations. Their examples show trajectories as story context accumulates.
Goodfire research article ↗ · Full research paper ↗ · Setup and teaching guidance