What would convince me that a model understands physics?
Interventions, hidden state, and the difference between predicting a view and predicting an experiment.
Notes.
What could “searching a mind” actually mean?
The inverse problem underneath my most speculative idea.
Can permission expire after it has been delegated?
A revocation problem that begins where the first agent stops.
A solvent insurer can still miss Tuesday’s payment
The currency-hedging question that forced me to separate a destination from its financing path.
Before the error signal becomes a bit
Could analog information help a quantum controller track a drifting machine?
Drag the curve. Keep the experiment.
An output can become an input, provided the system knows what it is allowed to change.
What should survive before the question arrives?
Learned forgetting, composed memories, and the information a later decision will need.
What a cold-storage model doesn’t know
When two explanations fit the meter, the next useful output may be a measurement.
A student fund has an unusually short memory
What Astir should retain when the people making its decisions keep graduating.
What if the important technology already exists?
A question about physical productivity, adoption costs, and the investments that get overlooked.
A contact skill should survive a new hand
What could transfer between robot bodies when actuator commands cannot?
The demonstration leaves out the rule
Learning when an automation should apply, including when it should leave the files alone.
Ask for the spectral quantity the decision needs
An inverse problem can remain ambiguous after the practical question is already settled.
The value is right. Is the derivative?
A numerical repair has more than one contract to satisfy.
The boundary correction made it worse
What broke when I transferred a neural PDE solution into a barycentric interpolant.
The best next judge was often no next judge
What survived after I tested an adaptive verification policy against simpler alternatives.
The changed page was not where the forgetting lived
Why my locality-aware replay policy lost its main argument.
A token budget is not an energy budget
The controller stayed fixed. The machine changed what its actions cost.
A memory system has to decide when to interrupt
Gardener made capture easier. The harder problem starts after the note is saved.
Can ethics wait until the product works?
A question about experiment design, irreversible effects, and who pays for learning.
After the row has been restored
Recovery has to preserve other people’s work and account for effects outside the database.
Knowing the threshold did not enforce it
A small legal-reasoning failure with a larger systems lesson.