Has it grown up?
From about 8 CU, at doubling amounts of compute and then every 1,024 CU, it gets a checkup: six short tests in fresh rooms it has never seen, with its learning paused. Each bar is how many it passed.
- tests passed (the dashed line is all six)
no checkup yet (the first comes soon after it has lived 8 CU)
- Learner
- starting…
- Speed
- …
- Steps lived
- 0
- Compute lived
- …
- Backend
- choosing…
- Status
- …
- Caregiver
- …
- Sessions
- loading…
- Recipe
- …
settings
runs: …
genome: - Meta-learner
- …
- Score
- Knowledge
- loading…
- Contribution
- …
Is it learning?
At every step it guesses what it will see next. Here those guesses are scored on a fixed test it never trains on: 64 moments of a differently laid-out copy of its world, recorded at its birth. The lazy guess is that nothing changes; a learner that learns ends up below it.
- how wrong its guesses are, on the fixed test
- the lazy guess: nothing changes
- Ratio
- no guesses measured yet
| Learner | Steps | CU | Held-out ratio | Own-stream ratio | Milestones | Recipe | Score | Final | Setting | Host |
|---|
No past sessions yet. When a learner's logs are merged, its session appears here.
A learner lives in this tab. It starts from nothing, with no pretrained knowledge. It sees only pixels of a small screen and acts only through a cursor. When its life ends, what happened is sent back to GitHub, so later learners can start from what earlier ones found and, over time, do better.
It learns to predict what it sees: at every step it guesses its next view of the screen, and it learns from how wrong the guess was. It seeks out what it can learn, moving where its guesses are getting better. A caregiver carries it between rooms, toward where it is learning fastest, until it uses the taskbar itself; now and then it gets a checkup in rooms it has never seen. Its life is measured in compute (1 CU is a billion arithmetic operations), and the LEARNER tab shows its guesses against the lazy guess that nothing changes.
It runs on this machine's GPU through WebGPU when there is one, otherwise on the CPU at half the size. A reload resumes it from its last checkpoint, saved in this browser.
To help, leave the tab open, and add a GitHub token so its logs can be shared. Without a token the learner still runs and its logs wait in this browser.
Privacy: the token is kept only in this browser and is sent only to api.github.com. The logs hold the
learner's settings, step and compute counts, its loss curve and a coarse description of the machine (cores, memory, WebGPU, open tabs). The
pull request is opened under the token's GitHub account.