Before agents, cost tradeoffs, or evals make sense, you need a mental model of what LLMs actually do — how they're trained, how they generate text, and where they break down.
TrainingInferenceModels
Latest
Foundations
New here? Begin with these three. Every other series builds on them.
Context Windows & Memory
3 parts
Every model has a fixed working memory, and half of product design flows from that constraint. Token budgets, why models "forget," and the patterns that fake persistence convincingly.
ContextMemoryRetrieval
Cost & Latency Tradeoffs
3 parts
Tokens aren't free — and output tokens cost five times what input does. How pricing actually works, why bigger isn't better, and the levers that turn an expensive prototype into a viable product.
TokensPricingOptimization
Building
Turning a model into a product: agents, grounding, and writing code with AI.
Agents & Tool Use
4 parts
A chatbot talks; an agent does. The think-act-observe loop, tool use, and the orchestration patterns that let LLMs act on the real world — plus what breaks when they do.
Agent LoopTool UseOrchestration
Grounding & Guardrails
4 parts
LLMs are confident about everything — including things they made up. RAG that grounds answers in your data, retrieval worth trusting, and the guardrails that keep your product's mouth in check.
RAGSafetyHallucinations
You don't need to become an engineer to ship software anymore. The tools, the craft, and an honest account of where vibe coding shines — and where it bites.
ReplitClaude CodePromptingDebugging
Shipping
Knowing when it's good enough, and designing for the times it isn't.
Evaluation & Testing
4 parts
You can't unit test creativity. The evaluation mindset, the testing playbook, and how to know your AI is good before your users find out it isn't.
EvalsMetricsTesting
95% accurate means failing one user in twenty. UX patterns for showing uncertainty honestly, recovering trust after failure, and keeping humans in the loop where it counts.
Uncertainty UXHITLReview Queues
To get updated whenever new explainers are added, follow @navsha on X.