Drills: The Frontier.
Five questions at a time from the whole volume, until every frontier claim comes with its status.
This plate draws rounds of five from the questions and predictions of the ten plates before it, with the reasoning and source plate after each answer.
The volume’s thread is to read the frontier with the course’s methods: separate established from reported from contested, ask for the measurement behind each claim, and expect dates and numbers to move.
Below: a round of five. Answer each, read the explanation, and press “Next round” for another five.
- Tag claims by status.
- Read scaling and efficiency trends.
- Explain data and memory constraints.
- State open problems precisely.
What to carry forward
Test-time computation pays on checkable problems (plate XXX.1); RL scales along curves with ceilings (plate XXX.2); memory across sessions is external because continual learning is unsolved (plate XXX.3); hybrids trade memory for recall (plate XXX.4); synthetic data helps when verified and anchored (plate XXX.5).
Human text is finite while wanted tokens grow (plate XXX.6); capability gets cheap while the frontier gets dear (plate XXX.7); learned world models compound errors (plate XXX.8); safety problems are measurable in part and unsolved (plate XXX.9); and the biggest questions become tractable as measurements (plate XXX.10). Vol. XXXI looks back and asks you to build.
Terms
- Established
- Replicated or long-standing.
- Reported
- From a single source, not independently reproduced.
- Contested
- Actively disputed.
- Open
- Little evidence either way.