knaxuslearn · build · ship

Roadmaps / AI / How the plan works

Read first

How the plan works

Time model. One study week is 8 hours: about 2 hours of reading, 5 of building and 1 of review. The core path is 16 weeks, 128 hours. Optional stretch work adds about 30 hours, so the full path is 158 hours. Fresher pacing: plan 20 weeks (160 hours), with one extra week in each phase for debugging and rerunning the checkpoint. At 5 hours a week, allow 26 weeks for the core path. Week 0 is an extra 10 hours and is optional. Practice projects are the building time of their weeks, not extra hours.

Models and spend. The core path works with either a hosted model through an API or an open-weight model run on your own machine. With a hosted model, set a hard spending limit in the provider's console before you start, and know that some limits only send an alert. Every exercise below has a version that costs a few dollars or nothing. Model names and prices change often, so look up the current ones when you start and record them in your ADRs.

Every phase has the same parts. A weekly table, up to three practice projects, up to five repositories to read, one milestone checkpoint and five questions to check yourself. The checkpoint is the point of the phase. The weekly deliverables are steps toward it.

Scoring. Score each checkpoint yourself, then ask one other engineer to score it. Use six dimensions: specification and budgets; trade-offs and decisions; correctness under failure; measurement; security; communication and influence. Use four levels: 1 Works on your machine; 2 Sound (specified, tested, repeatable: a stranger can rerun your measurements); 3 Proven (trade-offs measured, holds at the limits, a seeded fault is caught); 4 Leading (a method other teams reuse, with a measured outcome). With no second engineer, post your hand-in for review in a community forum and cap your self-score at 2.

CheckpointJudged mainly on
1 Measured modelMeasurement; specification
2 EvaluatedMeasurement; trade-offs
3 GroundedCorrectness under failure; measurement
4 Safe and shippedSecurity; correctness; communication

Three rules of the bench. Run every comparison on the same evaluation set and report the spread over several runs, because one run of a stochastic system proves little. Write your prediction before each measurement; the deliverables of Weeks 4, 8, 12 and 16 compare predicted and measured numbers. Keep every prompt, model name and setting in version control.