On 2026‑10‑14 I posted the “AI engineering master tree for 2026” on LinkedIn, laying out a thirteen‑layer taxonomy that describes the full stack of production AI systems. I offered a 380‑page PDF for free and asked the community for feedback on additions, removals, or renames.
The AI engineering master tree for 2026!
It covers thirteen layers of production AI systems.
- Foundations – tokens, embeddings, transformers, attention, positional encoding, context windows, mixture of experts, scaling laws
- Model behavior – pretraining, post‑training, sampling, temperature, reasoning models, test‑time compute, multimodality, hallucinations
- Prompt engineering – system prompts, few‑shot, chain‑of‑thought, structured outputs, self‑consistency, meta‑prompting, prompt caching, optimization
- Retrieval – document parsing, chunking, vector databases, hybrid search, query rewriting, reranking, GraphRAG, agentic RAG
- Agents – function calling, ReAct, planning, reflection, multi‑agent systems, computer use, human‑in‑the‑loop, sandboxed execution
- Context engineering – compaction, memory, MCP, agent skills, subagents, just‑in‑time retrieval, structured notes
- Fine‑tuning – SFT, LoRA, QLoRA, RLHF, DPO, GRPO, RLVR, distillation, synthetic data
- Inference optimization – quantization, KV caching, continuous batching, speculative decoding, FlashAttention, PagedAttention, parallelism, prefill‑decode splitting
- Deployment – serving engines, model routing, semantic caching, streaming, load balancing, autoscaling, fallbacks, on‑device inference
- Evaluation – benchmarks, LLM‑as‑judge, golden datasets, RAG metrics, hallucination detection, trajectory evaluation, regression tests, online A/B tests
- LLMOps – observability, tracing, cost tracking, prompt versioning, feedback loops, drift monitoring, model registries, continual learning
- Safety and security – guardrails, prompt injection, jailbreaks, PII redaction, red teaming, content moderation, tool permissions, audit logs
- Frontier patterns – decision models, harness engineering, loop engineering, recursive and contrastive LMs, verifier engineering, WebMCP, forkable agent runs