Curriculum

Foundations

Technical fluency and ML tradeoffs is the single strongest theme (90 percent of JDs, 77 companies), so the course opens here. The report is explicit that this must be decision-grade, not architecture-grade: give the PM enough of the model s

  1. 01.01

    The AI PM job now

    The AI product manager spends less time writing specs and handing off, and more time choosing among model outputs, judging quality, and owning the sign-off that a result is good enough to ship.
  2. 01.02

    Make the API call yourself

    You send a raw model request, set the roles and parameters that control it, and read the usage block that reports exactly what it cost.
  3. 01.03

    One token at a time

    Models generate one token at a time by predicting the next one, and how text, images, and audio are turned into tokens is what decides the bill.
  4. 01.04

    Context windows and attention

    Attention and the KV cache explain why an advertised million-token window is not a million-token memory, and why where you place instructions and how long a session runs both decay quality.
  5. 01.05

    Pretraining vs post-training

    Pretraining installs what the model knows and post-training installs how it behaves; reasoning models let you buy quality with extra thinking, and none of it makes the output deterministic.
  6. 01.06

    The AI-or-not decision

    Before reaching for an LLM, decide when SQL, a classical classifier, or a plain rule is the better product, and name the classifier's real price in labeled data and a chosen threshold.
  7. 01.07

    Read a benchmark honestly

    Read a public benchmark for contamination, saturation, and harness variance, then triangulate it against a style-controlled arena rating and a fixed probe set you built from your own tasks.
  8. 01.08

    Attribute every failure to a layer

    Place every AI failure in the model, the context, the harness, or the surface, then fix the cheapest layer that owns it, the diagnostic reflex the whole course reuses.
  9. 01.09

    INTEGRATION CASE

    With your product, you can write five testable failure predictions, tag each to one layer, and disprove one.