# Generated diagram assets

This folder contains course teaching aids. The raster illustrations were generated with the built-in ChatGPT image generation tool; the labeled SVG diagrams were authored as deterministic vector graphics so their wording, arrows, and numerical relationships can be checked exactly.

These are teaching aids, not primary sources. The Markdown captions keep exact explanations nearby because generated diagrams can simplify or slightly distort details.

## Review status

Reviewed visually on 2026-07-04 against the surrounding article text.

- `scale-ai-compute-units.png` - factually acceptable as a magnitude ladder; caption notes that visual object sizes are illustrative, not measured scale.
- `matrix-multiply-tensor-cores.png` - conceptually correct row/column dot product and tiled parallelism; caption notes simplification versus real tensor-core fragments/layouts/precision.
- `prompt-to-first-token.png` - correct high-level prefill/KV/decode flow; caption notes that the KV cache is drawn separately but normally lives in accelerator memory, usually HBM.
- `ai-compute-industry-pipeline.png` - useful but visually too linear; caption notes that HBM is a parallel input into advanced packaging, and the adjacent Mermaid diagram is more precise.
- `demand-flywheel-ai-compute.png` - acceptable as a feedback-loop scenario model; caption states it is not deterministic.
- `supply-steps-vs-demand.png` - acceptable qualitative supply/demand model; caption clarifies that "Relief" means temporary easing, not necessarily durable surplus.
- `text-becomes-tokens.png` - acceptable concept flow; caption notes token IDs and next-token probabilities are illustrative and tokenizer/model dependent.
- `agent-task-many-inferences.png` - acceptable simplified timeline; caption notes real tasks may have fewer or many more iterations.
- `ai-number-format-bit-layouts.svg` - deterministic SVG diagram, not model-generated; checked for FP32/BF16/FP16/FP8/FP4/INT4 bit-budget labels and 70B model storage arithmetic.
- `transformer-next-token-journey.svg` - deterministic SVG diagram, not model-generated; checked against the transformer flow in the original Transformer paper and the linked 3Blue1Brown lesson.
- `attention-qkv-one-layer.svg` - deterministic SVG diagram, not model-generated; checked against scaled dot-product attention, with all match strengths labeled as illustrative.
- `kv-cache-prefill-decode.svg` - deterministic SVG diagram, not model-generated; checked against the documented per-layer KV-cache update sequence.
- A rejected earlier supply/demand chart with a label artifact was reviewed, excluded from the course, and removed from the project assets. The original remains only in the default `.codex/generated_images` generation folder.

## Files

- `scale-ai-compute-units.png` - unit scale ladders for bytes, FLOPs, tokens, and watts.
- `matrix-multiply-tensor-cores.png` - matrix multiplication mapped to tensor-core tiles.
- `prompt-to-first-token.png` - prompt tokens, prefill, KV cache, HBM, decode, streamed answer.
- `ai-compute-industry-pipeline.png` - industry pipeline from tools/foundry/HBM to AI products.
- `demand-flywheel-ai-compute.png` - capability, revenue, commitments, supply buildout flywheel.
- `supply-steps-vs-demand.png` - staircase supply versus smooth demand.
- `text-becomes-tokens.png` - text to tokens, IDs, embeddings, transformer, next-token probabilities.
- `agent-task-many-inferences.png` - one agent task as repeated model/tool/context loops.
- `ai-number-format-bit-layouts.svg` - exact bit-budget and memory-impact diagram for FP32/BF16/FP16/FP8/FP4/INT4.
- `transformer-next-token-journey.svg` - next-token path from token embeddings through repeated attention and feed-forward layers.
- `attention-qkv-one-layer.svg` - one attention head's query-key comparison and weighted use of values.
- `kv-cache-prefill-decode.svg` - per-layer cache creation during prefill and one-position growth during decode.
- `generated-diagrams-contact-sheet.png` - contact sheet of selected generated diagrams.

## Prompt set

### AI compute industry pipeline

Use case: scientific-educational. Asset type: wide classroom infographic for a technical web course, 16:9 aspect ratio. Create a polished visual diagram titled "AI Compute Industry Pipeline" showing the flow from upstream industrial inputs to AI products. Use a clean modern textbook style, not a marketing slide. Show nine large labeled stations connected by thick arrows: Tools, Foundry, HBM Memory, Advanced Packaging, Accelerator Vendor, Server & Rack, Datacenter Power, Cloud / Neocloud, AI Products. Add bottleneck valve icons at Foundry, HBM Memory, Advanced Packaging, and Datacenter Power. Include visual hints: EUV machine, wafer, stacked HBM, interposer package, GPU, liquid-cooled rack, power substation, cloud dashboard, chatbot/agent. No logos, no brand names, no watermark.

### Demand flywheel for AI compute

Use case: scientific-educational. Asset type: wide classroom infographic, 16:9 aspect ratio. Create a polished visual diagram titled "Demand Flywheel for AI Compute" with six large nodes around a loop: Better Models, Better Agents, More Revenue, Bigger Compute Commitments, More Datacenters + HBM, Larger Training & Inference Fleets. Add a side box labeled "Forward-looking demand" showing preorder contracts, power reservations, and take-or-pay commitments as simple icons. No logos, no watermark.

### Supply steps versus demand

Use case: scientific-educational. Asset type: wide classroom infographic, 16:9 aspect ratio. Create a clean visual diagram titled "Supply Steps vs Demand". Make a graph with y-axis Capacity and x-axis Time. Draw a red rising Demand curve and a blue staircase Supply line. Add one red shaded gap labeled Scarcity and one green shaded area after a supply step labeled Relief. Put simple unlabeled memory stack, chip package, datacenter hall, and power substation icons on supply steps. Use only the words: Supply Steps vs Demand, Capacity, Time, Demand, Supply, Scarcity, Relief.

### From prompt to first token

Use case: scientific-educational. Asset type: wide classroom explainer, 16:9 aspect ratio. Create a polished visual diagram titled "From Prompt to First Token" showing prompt tokens entering a GPU server, a Prefill block, a KV Cache memory shelf, HBM near the GPU, a Decode loop, First token, and Streamed answer. Use exactly these labels: Prompt tokens, Prefill, KV Cache, Decode, HBM, First token, Streamed answer. No logos, no brand names, no watermark.

### Matrix multiply on tensor cores

Use case: scientific-educational. Asset type: wide classroom explainer, 16:9 aspect ratio. Create a polished visual diagram titled "Matrix Multiply on Tensor Cores". Show Matrix A and Matrix B, a highlighted row and column, a Dot product tile, Output C, and many Tensor cores working as Many tiles in parallel. No logos, no watermark.

### Text becomes tokens

Use case: scientific-educational. Asset type: wide classroom explainer, 16:9 aspect ratio. Create a polished visual diagram titled "Text Becomes Tokens". Show the exact sentence "The quick brown fox" breaking into Tokens, Token IDs, Embeddings, entering a Transformer, and producing Next token probabilities. Include a bottom analogy strip showing words chopped into puzzle pieces and translated into numbers. No logos, no watermark.

### One agent task equals many inferences

Use case: scientific-educational. Asset type: wide classroom explainer, 16:9 aspect ratio. Create a polished visual diagram titled "One Agent Task = Many Inferences". Show a timeline: User prompt -> Model call -> Search files -> Model call -> Edit file -> Run tests -> Model call -> Browser tool -> Model call -> Final work product. Add tiny GPU/HBM icons under model calls, shell/file/browser icons under tools, and a Context summary notebook updated between steps. No logos, no watermark.

### Scale of AI compute units

Use case: scientific-educational. Asset type: wide classroom explainer, 16:9 aspect ratio. Create a polished visual diagram titled "Scale of AI Compute Units". Show four vertical ladders: Bytes, FLOPs, Tokens, Watts. Bytes ladder: bit, byte, GB, TB. FLOPs ladder: FMA, TFLOP, PFLOP, EFLOP. Tokens ladder: word, sentence, corpus. Watts ladder: GPU, rack, datacenter. Use increasing-size visual blocks to make magnitude feel concrete. No logos, no watermark.