# The AI Field Guide, 2026 Edition > Fifteen Tufte-coded cards on how AI actually works - written for the people who approve the strategies, sign the contracts, and answer for the outcomes. A four-week Beyond Reason series by Eric Porres, Chief AI Officer at Logitech, collected as a free 22-page PDF (CC BY 4.0) and a permanent home on the web at porres.com/aifieldguide. The guide is fifteen single-topic cards in three clusters plus a synthesis. The cards are visual infographics; each card page below carries the full-size card image and a PNG download. Each card stands alone, but the strongest reading is sequential - the synthesis on Card 15 depends on the layers the earlier cards make legible. If a colleague has time for exactly one card, send Card 04 (Context Assembly); it is the card most likely to change a buying decision. A plain-text rendering of the entire guide - the full text of all fifteen cards plus the six wrapping pages and the receipts - is at https://porres.com/aifieldguide/llms-full.txt. ## The guide - [The AI Field Guide (home)](https://porres.com/aifieldguide/): Landing page - the four essays, the PDF, and all fifteen cards. - [Full guide as plain text](https://porres.com/aifieldguide/llms-full.txt): The entire guide as text, including all fifteen card transcriptions and every wrapping page. - [Download the full PDF](https://porres.com/aifieldguide/assets/ai-field-guide-2026-beyond-reason-eric-porres.pdf): 22 pages - fifteen cards, six wrapping pages, full receipts. 8.5 x 14 in. Free. CC BY 4.0. ## The fifteen cards Part 1 - What happens when you hit send (the runtime of a single request): - [Card 01 - An LLM Request, End to End](https://porres.com/aifieldguide/cards/01-request-end-to-end.html): The eight-step path of a single live request, and what "training data OFF vs. ON" actually changes (nothing, in the moment). - [Card 02 - Caching, Cost, and Latency](https://porres.com/aifieldguide/cards/02-caching-cost-latency.html): Why the same request can be fast, slow, cheap, or expensive - and where caching helps or hurts. - [Card 03 - How Tools Fit Into an LLM Request](https://porres.com/aifieldguide/cards/03-tools.html): Where external tools enter the path - at context assembly and again at result integration. - [Card 04 - Context Assembly, End to End](https://porres.com/aifieldguide/cards/04-context-assembly.html): How a request gathers, filters, ranks, and packs what the model actually sees. Start here. - [Card 05 - The Context Window Is Not Memory](https://porres.com/aifieldguide/cards/05-context-vs-memory.html): What persists vs. what is temporary, and why the context window is working memory, not memory. Part 2 - Where the answers come from (training, retrieval, infrastructure): - [Card 06 - An LLM Training Run, End to End](https://porres.com/aifieldguide/cards/06-training-run.html): How a model is trained end to end, and what a knowledge cutoff really means. - [Card 07 - Inside an LLM Training Run](https://porres.com/aifieldguide/cards/07-inside-training.html): Tokenization, embeddings, weights, and the predict-measure error-update loop. - [Card 08 - RAG and Retrieval](https://porres.com/aifieldguide/cards/08-rag-retrieval.html): How external knowledge is found, ranked, and injected into context before the model answers. - [Card 09 - Where Your LLM Prompt Goes](https://porres.com/aifieldguide/cards/09-data-centers.html): The physical path of a prompt - network edges, regional routing, and model-serving clusters. - [Card 10 - A Prompt Footprint, in Human Terms](https://porres.com/aifieldguide/cards/10-prompt-footprint.html): The energy, water, and carbon of a prompt, with disclosed 2025-2026 numbers and everyday comparisons. Part 3 - Why you can trust some AI systems (the trust layer): - [Card 11 - Agents](https://porres.com/aifieldguide/cards/11-agents.html): From one answer to multi-step work - the agent loop, control points, and failure modes. - [Card 12 - Permissions and Data Boundaries](https://porres.com/aifieldguide/cards/12-permissions.html): Same prompt, different identity, different safe answer. Access defines what AI can say. - [Card 13 - Routing and Model Selection](https://porres.com/aifieldguide/cards/13-routing.html): Choosing the right model, tool, or agent for each request automatically - model choice as a job, not a habit. - [Card 14 - Evaluation: From Vibes to Evidence](https://porres.com/aifieldguide/cards/14-evaluation.html): Moving from demos to measured behavior at scale - eval sets, scorecards, and failure modes. - [Card 15 - The Enterprise AI Control Plane](https://porres.com/aifieldguide/cards/15-control-plane.html): The synthesis - the governed layer between people, models, and company data. ## The essays (Beyond Reason) - [How AI Actually Works: A Field Guide for People Who Make Decisions](https://promptedbyeric.substack.com/p/how-ai-actually-works-a-field-guide): June 9, 2026 - the manifesto. The bird metaphor, the literacy gap, and the four-week arc. - [What Happens When You Hit Send](https://promptedbyeric.substack.com/p/what-happens-when-you-hit-send): June 16, 2026 - Part 1, Cards 01-05. The runtime of a single AI request. - [Where the Answers Come From](https://promptedbyeric.substack.com/p/where-the-answers-come-from): June 23, 2026 - Part 2, Cards 06-10. Training, retrieval, infrastructure. - [Why You Can Trust Some AI Systems and Not Others](https://promptedbyeric.substack.com/p/why-you-can-trust-some-ai-systems): June 30, 2026 - Part 3, Cards 11-15. The trust layer. ## About - [Eric Porres](https://porres.com): Chief AI Officer at Logitech; author of Beyond Reason. Contact: eric@porres.com. - [Beyond Reason](https://promptedbyeric.substack.com): Eric's Substack on AI and the economy for practitioners. - License: Creative Commons Attribution 4.0 (CC BY 4.0) - share, print, modify, attribute.