LLMs & Generative AI
The fastest-moving domain on the map. Almost everything here is Layer 3 or worse, so consume it, do not study it — with the two exceptions marked Layer 2, which are where the durable reasoning actually lives.
Clusters
Section titled “Clusters”Frontier providers
The message/role/tool-call shape is now near-universal, so provider switching is cheap. Context limits, caching semantics, tool-call formats and…
Core techniques
Layer 2 and the durable half of the domain. Retrieval, chunking, evaluation and the context-window trade are reasoning that survives every…
Application frameworks
The chain/retriever/tool vocabulary transfers loosely. Abstractions here churn faster than the models under them, and the framework is usually the…
Agents & orchestration
The loop — plan, call a tool, observe, repeat — plus memory and hand-off patterns transfer. MCP is the outlier and the most durable member: it is a…
Inference & local serving
KV cache, continuous batching, paged attention and the throughput/latency trade transfer across all of them and are the real content. Quantisation…
Evaluation & observability
Golden datasets, LLM-as-judge and its known biases, and tracing a multi-step run transfer across every tool. This is the most under-invested cluster…
Multimodal generation
Conditioning, guidance and sampling schedules transfer across image models. Speech and image share almost nothing beyond the word “generation”.
Safety & guardrails
Layer 2: the threat model — untrusted input reaching a privileged tool call — is ordinary security reasoning and does not decay with the models.