This page answers short questions about the parts every solution had to add in this book. It is for a reader who has read the chapter, or who wants a quick answer first. Each answer links to the section that explains it in full.
Contents#
- Why walk eight stands when one look covers the whole zone?
- How does a policy that only draws paths say it has finished?
- Why did so many glasses break while the arm was only travelling?
- Why does a copying model learn to look and not to push?
- How can the push a model scores best be its worst mistake?
- Why is a safety check that refuses the teacher a broken check?
1. Why walk eight stands when one look covers the whole zone?#
One look sees the whole floor of the zone, but not every glass on it. Glasses hide each other, and a hidden glass is simply missing from the list. From eight sides, no glass stays hidden. The ring of eight stands gives the measurement.
2. How does a policy that only draws paths say it has finished?#
It cannot say "done", so the bench reads it from what the policy does. Sixteen answers in a row that could not touch a glass count as "done". The number sits above the teacher's longest quiet spells. How a solution says it is done explains the rule.
3. Why did so many glasses break while the arm was only travelling?#
A waypoint names the middle of the hand, and the fingers hang about 70 mm lower. A path that looked high dragged the fingers across the rims of the glasses. The hand hangs below the waypoint and hands that swept the rims explain it.
4. Why does a copying model learn to look and not to push?#
Four of the teacher's answers in five are moves to look. Only one in 39 is a push at the task. A model trained to be close to the teacher on average gives the common answer, which is a look. Four answers in five are looks gives the counts.
5. How can the push a model scores best be its worst mistake?#
A search takes the best of thousands of predicted pushes. A guess that is too hopeful looks like a good push, so the search picks the largest errors. The winner's curse shows the measurement.
6. Why is a safety check that refuses the teacher a broken check?#
The teacher is known to be safe, so a rule that refuses its answers is refusing safe answers. Running the teacher through each new rule is a cheap test. It once caught a rule that refused every push in the project. The test against the teacher tells the story.
← What fixed them, and what did not · One fixed nudge: what is different →