TL;DR: Four pages that take you from "AI is magic" to fluent: what the job actually is, how ML works, how LLMs work, and the vocabulary to discuss both. Everything in
02-technical-depthassumes these.
| Page | What you'll learn | Time to read | Why it matters |
|---|---|---|---|
| what-ai-pms-do.md | How AI PM differs from traditional PM · the 4 archetypes (platform, AI-native, feature, ML systems) · a week in the life · a 10-skill self-assessment · myths debunked | ~12 min | Tells you which AI PM job you're actually applying for. The same title covers genuinely different roles |
| ml-fundamentals.md | Rules vs. examples · the ML family tree · training vs. inference · precision/recall · overfitting · drift · labels · the "should this be ML?" decision tree | ~15 min | Ranking, fraud, and recommendations predate the LLM wave and are still the most common AI PM jobs |
| llm-fundamentals.md | Next-token prediction · tokens & context windows · the pre-train → SFT → preference-tuning pipeline · inference knobs · what LLMs are bad at · explain-it-to-an-exec cheat sheet | ~15 min | Every generative product decision (cost, latency, hallucination, prompting) falls out of how the model works |
| glossary.md | 100+ terms by theme: core ML, evaluation, LLMs, architectures, production economics, responsible AI. ~30 bolded as interview-fluent must-haves | Skim, then reference | Misusing a term is a louder signal in an interview than not knowing it |
Straight down the table. what-ai-pms-do frames the job so the rest has somewhere to land; ml-fundamentals before llm-fundamentals because LLMs are a branch of the tree, not a replacement for it; glossary last, and then keep it open in a tab forever.
If you already run ML products, skim 1-2 and read llm-fundamentals properly. If you came for chatbots, do not skip ml-fundamentals. See 00-start-here for path selection.
flowchart LR
A["what-ai-pms-do<br/>the job, the archetypes"] --> B["ml-fundamentals<br/>examples beat rules · error tradeoffs"]
B --> C["llm-fundamentals<br/>next-token prediction · tokens · adaptation"]
C --> G["glossary<br/>the shared vocabulary"]
C --> P["prompt-engineering"]
C --> R["RAG<br/>grounding vs. hallucination"]
C --> F["fine-tuning<br/>style & skill, not facts"]
C --> AG["ai-agents<br/>tool use in a loop"]
B --> E["evals<br/>quality as a distribution"]
C --> E
C --> M["model-selection-and-cost<br/>tokens → unit economics"]
B --> M
Concretely: the adaptation ladder previewed in llm-fundamentals §6 becomes prompt-engineering → rag → fine-tuning → ai-agents. The precision/recall operating point from ml-fundamentals §3 becomes evals. Tokens become model-selection-and-cost.
Five self-tests, all answerable from the pages above. No partial credit; say them out loud.
- Explain hallucination to an exec in one breath: that it's the failure mode of plausibility-based next-token prediction, not a bug awaiting a patch, and name the four-part product response: grounding, calibration UX, evals, stakes-based design.
- Name precision vs. recall as a product decision, not an ML one, and defend a specific operating point for a specific product (99% precision / 60% recall on fraud blocks almost no good users but misses 40% of fraud; who eats that?).
- Place RAG vs. fine-tuning on the adaptation ladder, and answer "just train the model on our docs so it stops hallucinating" without flinching (fine-tuning teaches style and skill; RAG grounds facts; prompting comes first, always).
- Say which of the four archetypes you're targeting (platform, AI-native, feature, or ML systems), and what technical depth that one demands.
- Use ten bolded glossary terms correctly in a sentence each, to an engineer and to a CEO. Try: inference, baseline, drift, threshold, context window, temperature, system prompt, eval, groundedness, prompt injection.
Can't clear one? The gap names the page to reread. Score yourself on the 10-skill inventory too; the low scores are your study plan.
➡️ Next: 02-technical-depth, where these concepts become architecture decisions you'll defend in interviews and design reviews. Start with prompt-engineering.