blog
Posts derived from first principles: one invariant, every objection answered where it arises, the mechanism built from scratch, and the trade-offs priced. Browse by category above, or filter across all of them by tag.
- AI
How the models actually run: what the serving layer does with your request, what it costs, and which behaviour you can predict rather than measure.
1 topics · 1 post
Filter by tag — they cut across categories.
ai · llm ·
Prompt caching, from the causal mask outwardPrompt caching explained from the one fact it follows from: attention is causal, so the cacheable unit is a prefix, never a block.
- prompt-caching
- kv-cache
- inference
- claude-api
- cost
- latency
- agents