Learning paths: recap
The paths in one view
| Path | Starts with | Spends the most time on | Ends with |
|---|---|---|---|
| Senior PM | A technical footing | The decisions only the PM owns | Optional depth where the roadmap needs it |
| AI Product Lead | Model and data foundations | Context, agents, and trust | Running the product |
| AI Engineer | Machine learning and LLM basics | Agents and the harness | Production |
| AI Engineering Lead | Model behaviour and cost | Quality, safety and access | Working with product |
Which path reads how much of each track
"Full" means every lesson. A number means that many lessons out of the track's total. A dash means the path skips the track.
| Track | Senior PM | AI Product Lead | AI Engineer | AI Engineering Lead |
|---|---|---|---|---|
| Product sense | 1 of 7 | 1 of 7 | — | 1 of 7 |
| Technical product sense | 5 of 9 | 1 of 9 | — | 2 of 9 |
| Technical product management | 1 of 9 | 5 of 9 | — | 7 of 9 |
| Generative AI | Full | 3 of 6 | — | — |
| LLMs | 6 of 7 | 6 of 7 | 6 of 7 | 5 of 7 |
| Prompt engineering | 4 of 10 | 1 of 10 | 7 of 10 | — |
| Evaluation & observability | Full | Full | — | Full |
| Cost optimization | Full | Full | — | Full |
| AI security & guardrails | 2 of 3 | Full | 2 of 3 | Full |
| Agentic AI | 2 of 8 | 2 of 8 | — | — |
| RAG & vector databases | 4 of 10 | 6 of 10 | 9 of 10 | 5 of 10 |
| Memory & context | 2 of 5 | Full | 2 of 5 | 2 of 5 |
| Context engineering | — | 6 of 7 | 2 of 7 | 3 of 7 |
| Tool calling | 1 of 3 | Full | 2 of 3 | 1 of 3 |
| AI agents | — | Full | 3 of 5 | 3 of 5 |
| Agentic workflows | — | Full | Full | Full |
| Knowledge graphs | — | 3 of 8 | — | 2 of 8 |
| Access control | — | 2 of 8 | 3 of 8 | 4 of 8 |
| System design | — | — | 3 of 16 | 4 of 16 |
| APIs & integrations | — | — | Full | — |
| Machine learning | 4 of 8 | 5 of 8 | Full | 7 of 8 |
| Tensors | — | — | Full | 3 of 5 |
| CNNs | — | — | Full | — |
| Transformers | — | — | Full | 4 of 6 |
| Fine-tuning and post-training | — | — | Full | 3 of 6 |
| AI-assisted development | — | — | 3 of 6 | 5 of 6 |
| Building AI-native products | — | 6 of 11 | — | — |
| Visual explainers | 7 of 12 | 6 of 12 | — | — |
| AI engineering stack (content/) | — | — | 15 of 23 | 12 of 23 |
| Harness engineering | — | — | 41 of 41 lessons | 18 of 41 lessons (part skim) |
What the paths share
Three of the four paths read the evaluation, security and cost lessons. The PM reads them to make the call. The product lead and the engineering lead read them to set the standard. The AI engineer meets the same ideas by building them, in the harness track and the AI engineering stack.
Test yourself
-
You are a senior PM with a launch in two weeks. Which stage do you do first, and why?
Answer
Stage 3, the decisions. It covers the eval plan, cost and risk, which a launch needs. Read stage 1 and 2 later. Each stage's checkpoint shows what you missed. (Path page) -
What is the difference between the AI Product Lead path and the Senior PM path?
Answer
The lead path is wider and deeper on the layers between a model and a product: grounding, context, memory, agents, workflows, trust and cost. It assumes the PM path or equivalent. (Path page) -
Why does the AI Engineer path put the harness lessons in the middle and not at the start?
Answer
You need models, APIs and grounding first. The harness lessons build an agent from those parts, so they need them. (Path page) -
An engineering lead skims the harness phases. What do they lose, and what do they keep?
Answer
They lose the failures you only meet by typing the code. They keep the shape of each part: the loop, tools, permissions, budgets and failure handling, which is what a design review needs. (Path page) -
What is a checkpoint, and what do you do if you cannot meet it?
Answer
A checkpoint is something you can do or explain at the end of a stage. If you cannot meet it, read the stage again before you move on. (Path page)
Sources
Lesson counts come from the track build configs in this repository. The time estimates are a rule of thumb, not measured.