Distinguish parameters, memory, and context
Use the declaration manuscript’s bookkeeping state to identify what actually changed.
Course lessons · 3 of 5
Classify state changes and avoid confusing a context reset with untraining.
Useful background: Trace a learning and query path
A model instance has several kinds of state
The declaration manuscript writes S_t = (θ_t, m_t, c_t, o_t). Here θ denotes learned parameters, m persistent memory, c runtime context or activity, and o observations. The tuple is a bookkeeping convention, not a claim that every component has been measured.
A prompt can change context and output while parameters remain fixed. A training update changes parameters through a defined update operation. A persistent-memory write can survive a context reset without changing the model’s learned weights.
Persistence and reversal need a defined boundary
To test persistence, state what is reset, what remains stored, and how the retained change is observed. A changed text output is not enough to identify which internal component changed. Save the relevant before-and-after artifacts when available.
Gate5000’s XOR example is locally reversible under a repeated fixed mask. That property does not reconstruct unrelated discarded history or undo arbitrary training. Reversibility of one operation must be kept distinct from rollback of a complete system.
Trace a stipulated context reset
1Specify the starting record
In this teaching scenario, parameters are P1, persistent memory is M1, context is C1, and the observation log is O1.
2Define the reset operation
The operation replaces only context with an empty context C0. The scenario explicitly keeps P1 and M1 unchanged.
3Name the actual change
The operation changed c, not θ. Calling it untraining would describe a different operation than the one specified.
Make the idea respond
The lab below is a scalar illustration, s_(t+1) = a·s_t + b·u_t. It retains one number between updates; it does not implement the four-part bookkeeping tuple or the XOR operation in this lesson. With its defaults s₀ = 0, a = 0.8, b = 0.2, and constant input u = 1, the first three updated states are 0.2, 0.36, and 0.488. Use these to inspect retention and change in a deliberately small model.
Change one input at a time. Predict what should happen, then compare the computed result with your prediction.
Teaching model · computed in your browser
Watch a state remember
Apply a sequence of inputs to one persistent number. Explore forgetting, oscillation, and unstable feedback.
s is the persistent scalar state; a controls the retained prior state; b scales the input u. The same stored state is updated at every step. For this linear recurrence, differences in initial conditions shrink when |a| < 1. This one-number demonstration is not a neural network, a trained model, or an operating Pantheon runtime.
Try this: choose a single pulse and compare a = 0.8, −0.8, and 1.2. Which settings forget an earlier state, and which amplify it?
Playback starts only when you choose Play. It pauses when this tab is hidden or the lab leaves view. Reduced-motion preferences slow the updates; Step always gives you direct control. Runs stop after 40 updates or at the display limit.
The default worked example is shown below.
Stable memory: differences in initial state shrink because |a| < 1. Press Step to apply the first input to the initial state. The constant-input fixed point is b / (1 − a) = 1; convergence to it requires |a| < 1.
| Update | Input u | Previous state | New state |
|---|---|---|---|
| 0 | Not yet applied | Initial condition | 0 |
Your turn.
Try an answer before opening a hint. Feedback will explain the result, and you can practice again.
Interactive practice is loading. The lesson, worked example, and printable questions remain available below.
View printable practice and solutions
Use these questions when studying without interactive controls. Opening a written solution does not earn progress.
In S = (θ, m, c, o), which component denotes learned parameters?
- m
- c
- o
- θ
Read the solution
θ denotes learned parameters in this bookkeeping convention.
A stipulated operation empties context while keeping weights and persistent memory unchanged. What is it?
- A new independent model training run.
- A context reset.
- A parameter untraining operation.
- Proof that all persistent memory is erased.
Read the solution
It is a context reset under the stated boundary.
A binary state bit starts at 1. Apply XOR with mask bit 1 twice. What is the final bit?
Read the solution
1 XOR 1 = 0, then 0 XOR 1 = 1.
What does applying the same XOR mask twice establish?
- Reversal of that local fixed-mask operation.
- Reconstruction of every discarded training example.
- Reversal of all unrelated memory writes.
- A verified current Pantheon deployment.
Read the solution
The original bits are restored for this operation and mask; broader rollback needs additional mechanisms.
Follow the original record
Use the definitions you just practiced to read the project’s own notation, assumptions, and evidence.
The ideas to keep
- Name the state component that changes.
- Test persistence across a defined reset.
- Local reversibility does not imply complete system rollback.
Reading completion is separate from your practice evidence.
Progress is kept in this browser. How learning works