Primer series · mental models first
Demystify AI
A primer series for technical generalists: IT, security, ops, project managers, and anyone using tools daily who was never handed a working mental model for what happens under the hood.
Each piece starts with an almost-correct metaphor, refines it just enough to stay useful, then opens the next layer. Companion to Learn, with the same spine and a gentler register.
Start anywhere, but not nowhere
Primer catalog
Shorter pieces for building the vocabulary before deeper architecture decisions start to matter.
Start here
AI vs ML vs LLM vs agents — sorting out the words people keep mixing up
Four different words often collapse into one marketing pitch. A nested mental model makes the buying, building, and risk questions sharper.
LLM construction stages, from pretraining to LoRA
A language model moves through stages: pretraining, supervised tuning, preference tuning, evaluation, serving, retrieval, and adapter training. LoRA enters as a compact adaptation layer after the expensive base model exists.
What is MCP? The USB-C port for AI context
MCP is a standard way for an AI agent to ask another system for context or tools. Think less magic brain, more well-labeled port.
Tokens, context windows, attention — model mechanics without math
A working mental model for the path from prompt to returned text: tokens, context windows, and attention without a single equation.
Why LLMs hallucinate — same mechanism as the looseness, different consequence
Hallucination comes from the same retrieval looseness behind useful LLM answers, with a different consequence.
LLMs work like word-query databases, but looser
A practical mental model for LLMs: word-based queries over learned patterns, refined with the looseness behind iteration, useful surprises, and confident wrongness.