YG

Distill

Insights and patterns uncovered through building AI systems. Each piece captures a principle that emerged from practice, not theory.

01Decision vs. Sampling
Ask the same question twice.
You may get the same answer
but not the same reasoning.
That’s not a decision system.
That’s a sampling system.
If the reasoning shifts each time,
you don’t have a system you can rely on.
You have something that sounds right.
That's the difference.
The model encodes statistical patterns,
not a consistent decision process.
A decision system isn’t defined by what it knows.
It’s defined by:
how it evaluates inputs
what it prioritizes
how it reaches a conclusion
Without structure, knowledge is sampled.
With structure, knowledge is applied.
From building an accounting knowledge Q&A system, I learned that the same question, asked twice, can produce different reasoning paths. The answer might match, but the logic behind it shifts. That inconsistency undermines reliability. The fix was giving the model the right structure to reason within.
02When the path is unknown
Don’t define the solution
when the path is still unknown.
Ask.
What are you assuming?
What breaks this?
What did you ignore?
What would another perspective say?
Let it answer.
Then question again.
Not to lead,
but to apply pressure.
Surface assumptions.
Introduce counterexamples.
Shift perspectives.
Follow what doesn’t resolve.
Set only what must be true.
Use separate passes.
Let them evaluate independently.
Remove what doesn’t hold.
Until the answer
can stand on its own.
A method I developed for guiding coding agents through ambiguous problems, where the solution can’t be prescribed because the territory hasn’t been mapped yet.