Goodfire — Understand and debug your AI model
Make model inspection a research process viewers can follow: establish the subject and dataset, confirm the inspection plan, then present findings and the next experiment.
View HTML source
<!doctype html><html lang="zh-CN"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Goodfire — Understand and debug your AI model · sequence</title><style>html,body{margin:0;width:100%;height:100%;overflow:hidden;background:#eee9dc}canvas{display:block;width:100%;height:100%;object-fit:contain}</style><body data-mode="sequence"><canvas width="1000" height="600"></canvas><script src="data.js"></script><script src="common.js"></script><script src="scene.js"></script><script src="bridge.js"></script></body></html>Motion and composition
Local motion · Composition rules
Specify task and subject together in input
Place a serif question on cream paper, with user input specifying the inspection goal and model path. After the response, represent the model as brown layers, keeping layer count and name near it.
Allow data and plan confirmation before starting
View HTML source
<!doctype html><html lang="zh-CN"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Goodfire — Understand and debug your AI model · sequence</title><style>html,body{margin:0;width:100%;height:100%;overflow:hidden;background:#eee9dc}canvas{display:block;width:100%;height:100%;object-fit:contain}</style><body data-mode="clip" data-start="16" data-end="24"><canvas width="1000" height="600"></canvas><script src="data.js"></script><script src="common.js"></script><script src="scene.js"></script><script src="bridge.js"></script></body></html>Show dataset cards before user confirmation, then list checks. Align seven checks in a column with start near the bottom; do not skip confirmation to show results.
Keep run states in the same list
View HTML source
<!doctype html><html lang="zh-CN"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Goodfire — Understand and debug your AI model · sequence</title><style>html,body{margin:0;width:100%;height:100%;overflow:hidden;background:#eee9dc}canvas{display:block;width:100%;height:100%;object-fit:contain}</style><body data-mode="clip" data-start="25" data-end="33"><canvas width="1000" height="600"></canvas><script src="data.js"></script><script src="common.js"></script><script src="scene.js"></script><script src="bridge.js"></script></body></html>Replace pending, running, and completed states in sequence, applying progress only to the current item. After completion, expand findings into diagrams and brief conclusions instead of simultaneous page-wide warnings.
Connect conclusions to the next experiment
Close with one main finding, one explanation, and one follow-up action. Rewrite datasets, checks, and findings for new scientific tasks, never treating example figures as actual evaluation conclusions.
Sections and rhythm
Click a section to seek to it in the choreography example above.






Visual references · Discrete still frames
Creating other themes
- Replace model, dataset, checks, and follow-up experiment; preserve confirmation and checking order.
Production constraints
Requirement input / model attachment / layered structure / dataset card / check list / progress bar / findings diagram / follow-up button
- Place a serif question on cream paper, with user input specifying the inspection goal and model path. After the response, represent the model as brown layers, keeping layer count and name near it.
- Show dataset cards before user confirmation, then list checks. Align seven checks in a column with start near the bottom; do not skip confirmation to show results.
- Replace pending, running, and completed states in sequence, applying progress only to the current item. After completion, expand findings into diagrams and brief conclusions instead of simultaneous page-wide warnings.
- Close with one main finding, one explanation, and one follow-up action. Rewrite datasets, checks, and findings for new scientific tasks, never treating example figures as actual evaluation conclusions.