How building works

One command builds every model a creator needs, starting from a list of their YouTube videos. It takes one to two days, mostly unattended, and stops only when a person has to decide something.

./alive down                        # building needs the whole card
./alive build huberman --dry-run    # what would run next, in every stage
./alive build huberman              # do it; run the same command again after any stop
./alive build huberman --stage voice   # one stage only

Stages, in order: videos, answers, voice, clips, motion, render. The live app is not a build stage; it is a settings file (stage 6). It runs on the system python3 and picks each step's environment itself.

What needs what

Six stages. Each makes something the later ones read, which is why they run in this order.

What one build step looks like

For each stage, the build asks that recipe's status command what is next and runs it, until the recipe says it is done. When the next step is a decision only a person can make, it looks for a recorded answer.

What a stop looks like, and how to clear it

STOPPED at a decision only a person can make:
  voice: choose the step to release from the sweep table (RECIPE.md), and record it
  as voice.release_step in build.json
Record it, then run the same command again; finished stages are skipped.

Every stop names the file to change. You edit it by hand; there is no separate tool. The stops, and the edit each one wants:

The creator's build file

creators/<creator>/build.json says which videos and which settings, and records the decisions that are one value. Larger decisions sit beside it in creators/<creator>/build/.

{
 "videos":  {"folder": "huberman", "ids": ["QmOF0crdyRU", "XeN6eGO6FVQ", ...],   ← YouTube ids
             "download": ["--full", "--no-video", "--a2f", "--subs"]},   ← whole video kept as a file; frames are
                                                                   cut later, only for the tracked one
 "answers": {"creator": "huberman"},
 "voice":   {"profile": "huberman", "release_step": 8000},                    ← which trained step to keep
 "clips":   {"profile": "huberman", "takes": ["vXTK0Ac9i1Q"]},                ← which of the ids above to track
 "motion":  {"profile": "huberman", "run": "huberman_S_seed0", "wrong": "..."}, ← which training run to keep
 "render":  {"profile": "huberman", "deployed": true, "wrong": "..."}
}

The stages at a glance