One prompt, end to end
What is the capital of Ontario?
Toronto.
Between those two lines a language model does a specific, finite amount of work. This film follows all of it — seven tokens in, one token out — and draws each stage at the point where it actually happens.
Composed entirely in code. Every token ID on screen was measured, not estimated.
o200k_base · 200,019 entries · the dot marks a leading space
The film
Contents
Seven stages
The order is the pipeline's own. Select a stage to jump the film to it.
Accuracy
What was measured
Verified
The prompt is seven tokens under o200k_base, checked
with tiktoken rather than assumed.
Toronto is a single token, id 151837. With a leading
space it is 27812. The period is 13. Its bytes
are 54 6f 72 6f 6e 74 6f.
Toronto is in fact the capital of Ontario.
Deliberately not measured
The probability bars carry no numeric labels. Those values were never measured, so the panel shows relative height only and says so on screen.
The embedding strip is labelled schematic.
Architecture is given as a range — 32 to 80 blocks — so no single vendor's configuration is implied.
Corrections
Three things the outline had wrong
The brief this was built from described the pipeline in a way that does not match how current models run. The film says so on screen.
Speculative decoding is not a stage
It is not a step between softmax and de-tokenizing. It wraps the whole forward-pass loop, which is why scene seven is drawn as a bracket around stages 02 to 04 rather than a box after them.
Position does not enter at embedding
GPT-2 added learned positional embeddings at that layer. Llama, Qwen and Mistral apply RoPE to Q and K inside each attention layer. The embedding scene leaves that slot visibly empty and the next one fills it.
The block is stacked, not singular
There is no one transformer block. There are dozens, identical, each doing the same thing to a slightly different signal — drawn with the residual spine that runs through all of them.
Colophon
- Composition
- HTML and GSAP, rendered with HyperFrames
- Narration
- Kokoro, voice bf_emma, ten segments mixed to a single timeline
- Captions
- Word-synced with Whisper, four mishearings corrected by hand
- Sound
- Deterministic FFmpeg drone
- Type
- League Gothic and JetBrains Mono
- Master
- 1920 × 1080, 30 fps, 144.000 s, H.264