glossary
Recurring terms in the magazine.
Terms that surface across multiple issues. Edited by hand. Each issue can attach a per-issue glossary block via its glossary_terms frontmatter field.
api
prompt caching
aka cache_control · ephemeral cache
- An Anthropic API feature that lets repeated prefix tokens (typically a system prompt or long style guide) be billed and processed once across many calls, by marking that block with cache_control.
tool use
aka tool-use · structured outputs
- An Anthropic API mode in which the model is given JSON-schema-typed tools and emits structured calls instead of free text. The magazine's curate and write steps both use it for typed output.
architecture
RAG
aka Retrieval-Augmented Generation · retrieval-augmented
- A pattern where a model is given relevant documents — usually fetched from a vector store — as part of its prompt, instead of relying only on its training data.
benchmarks
SWE-bench
first seen 2026-05-11
- A benchmark of real GitHub issues used to evaluate coding agents end-to-end. The model is given a repository and an issue and asked to produce a patch that the test suite accepts.
dev-tools
MCP
aka Model Context Protocol
first seen 2026-05-11
- A protocol for letting language models call external tools through a standard interface, originally introduced by Anthropic in 2024 and since adopted across most major coding-agent platforms.
magazine-meta
signal strength
aka signal
- A 0-100 integer per issue, computed deterministically from the diversity and average editorial weight of the day's cited sources. Surfaced in the masthead and at the foot of every article.
policy
FLOP
aka FLOPs · floating-point operation
first seen 2026-05-12
- A floating-point operation. Training-compute thresholds in AI policy are usually expressed as 10^N FLOP — the total number of arithmetic operations used during model training.