external optimization pipeline
Offline training
Valid claim: Fits parameters before evaluation using the declared data and compute.
Bounded substrate comparison
Offline training and online adaptation
Which costs and capabilities belong to training before deployment versus changes during operation?
Evidence status
Cites 2 sources, none of which has been read
The sources are named, but none has been retrieved and read as part of building this page. Nothing here has been matched to a passage, so the citations show where a reader might look rather than what was checked.
Rely on this page for
Orientation: how the topic is organised, which terms matter, and where to start reading.
Do not rely on it for
A claim you intend to act on or repeat. Follow the cited material yourself first.
operating system or living substrate
Online adaptation
Valid claim: Changes state during operation under the declared feedback.
Comparable axes
- Information available
- Update budget
- Held-out future performance
Non-equivalences
- Pretraining cost cannot disappear from lifecycle accounting.
- State drift is not necessarily learning.
Comparison procedure
- 1.Separate phases.
- 2.Freeze feedback.
- 3.Test retention and transfer.
Prohibited inference
Do not call any online state change learning unless it improves a preregistered held-out outcome beyond drift, damage, repeated exposure, and controller-only controls.
Connected concepts and sources
Direct answer
- Which costs and capabilities belong to training before deployment versus changes during operation?
Mechanism and method
- Separate phases.
- Freeze feedback.
- Test retention and transfer.
What is measured
- Information available
- Update budget
- Held-out future performance
Comparison: Offline training and online adaptation
- Pretraining cost cannot disappear from lifecycle accounting.
- State drift is not necessarily learning.
- Must not be read as: Do not call any online state change learning unless it improves a preregistered held-out outcome beyond drift, damage, repeated exposure, and controller-only controls.
Limitations
- Pretraining cost cannot disappear from lifecycle accounting.
- State drift is not necessarily learning.
What this does not establish
- Do not call any online state change learning unless it improves a preregistered held-out outcome beyond drift, damage, repeated exposure, and controller-only controls.