learning machines
Text · token
Token
Text is broken into chunks the model can count and predict — and the chunks may not match words. Try your own in the Tokenizer.
§ 0 · The map
13 terms · 4 modalities · one methodThe vocabulary clusters by modality — the same arc the camp runs: text, then images, then video, then the cross-cutting ideas that apply to all three. The color on each term below tells you which world it lives in.
3Text · Session 1
Token · Temperature · Greedy vs. sampled
4Image · Session 2
Default · Diffusion · CFG · Latent space
3Video · Session 3
Drift · Spatiotemporal · Coherence
3Cross-cutting
Modality · Hallucination · Human-in-the-loop
§ A · Quick glossary
term · plain meaning · the tool that shows it| Term | Plain meaning | See it in |
|---|---|---|
| Token | A chunk of text a model processes — often not a whole word. | Tokenizer |
| Temperature | How much randomness enters sampling — low stays safe, high takes risks. | Tokenizer |
| Greedy vs. sampled | Greedy always takes the most likely next token; sampling draws from the ranked options (often top-k). | Tokenizer |
| Default | What appears when the prompt doesn't specify — the model fills in the blanks. | Default Test |
| Diffusion | An iterative noise-to-image process, refined step by step. | Diffusion Viewer |
| CFG (guidance) | How hard generation is pushed to obey the prompt versus its own defaults. | Prompt Pressure (CFG scale) |
| Latent space | The compressed numeric space a model works in instead of raw pixels or words. | Latent Space Compressor |
| Drift | Unwanted change over time — the subject won't stay put. | Temporal Telephone |
| Spatiotemporal | Across both space and time — what a video must hold consistent frame to frame. | Metronome Scrubber |
| Coherence | Staying consistent across frames: identity, objects, camera, physics. | Video Viewer |
| Modality | A kind of medium a model works in — text, image, or video. | Tool index |
| Hallucination | A plausible-sounding output without reliable grounding. | Claim Checker |
| Human-in-the-loop | Human judgment before, during, and after generation. | Model Card |
§ B · See it, don't define it
each term as a small instrumentA definition tells you what a word means; a picture shows you the mechanism. Here are the terms that are easiest to misread, each rendered as the thing the tool actually does.
Text · temperature
Temperature
Low temperature favors the top token; raise it and less-likely continuations enter the sample. Watch the bars flatten in the Tokenizer.
Image · default
Default
When a prompt is vague, the system invents bodies, settings, and roles. Compare those choices in the Default Test.
Image · diffusion
Diffusion
The image is refined step by step from noise toward a prompt-guided result. Pause each stage in the Diffusion Viewer.
Video · drift
Drift
The subject keeps changing when it should stay stable. Temporal tools make it visible frame by frame — start with Temporal Telephone.
Video · coherence
Coherence
Frames belong together: identity, objects, camera, and physics stay consistent. Inspect it with the Frame-by-Frame Viewer.
Sounds confident: 94%
"According to Rivera (2021)…"
Evidence: source not found
"According to Rivera (2021)…"
Evidence: source not found
Cross · hallucination
Hallucination
A response can sound fluent and specific while lacking grounding. Treat it as a claim to verify in the Confidence Is Not Truth Explorer.
predictgenerate or inspectverify evidencerevise / refuse
Cross · human-in-the-loop
Human-in-the-loop
Human judgment belongs before, during, and after generation — setting purpose, checking evidence, naming limits. Document it with the Model Card.