20 passes through the data
Training loss
—final loss
Observed worldCraftax Classic
- Experience
- transitions
- Possible moves
- actions
- What the model sees
- values
recorded actions
This is the world the model learned from: a random-policy environment rollout, not the model dreaming yet.
Recorded observation
Seed
Frame 00 / 51
Before the first action
Awake
Step00/
- Decision
- Before the first action
- Reward
- 0.00
- Return so far
- 0.00
Initial observation saved.
- Total return
- Ending
- Discoveries
- Achievements
Prediction task
Not a picture. This first model tried to predict the hidden numbers describing what changes next.
Current observation
state valuesWhat happened
possible actionsSmall neural model
hidden featuresPrediction
- Next state
- Reward
- Alive or dead
It practiced on recorded transitions, then faced transitions from episodes it had never seen.
A transition is one moment: what the world looked like, the action taken, and what happened next.
Prediction error
That made “predict nothing changes” a surprisingly strong opponent.
—
stayed still
—
changed
Metric
Across the whole state, copying won.
Lower error is better ↓
Learned model
Simple baseline
Training and rollout error
Two curves tell the story: practice made the loss fall, but looking several steps ahead made the model's mistakes compound.
Forecast horizon
Rollout error
—model at step 8
Learned modelCopy-state baseline
The model remains worse than simply copying the starting state across all eight imagined steps.
Baseline comparison
Every learned number sits beside the simplest baseline that could expose a misleading success.
Run files
Saved on this website
Portable evidence
Kept on the Spark
Heavy working state
Open the reproduction receiptSeeds, versions, hashes, and raw files
- Experiment ID
- Lab source
- SWM source
↗- Dataset
- Checkpoint