Never Let the Model Pick the Problem

It will answer anything.

That is the trap.

Ask it a bad question
and it will not say
this is a bad question.

It will build you
a complete,
well-documented answer
to the wrong thing,

and it will do it
in nine seconds,

which is exactly long enough
to feel like progress.

The machine is very good
at how.

It is in no position
to know what,

or whether,

or for whom.

It was not in the room
when the customer
said the quiet thing
at the end of the call.

It does not know
which of your three problems
is the one
actually costing money.

It cannot smell
that the feature you asked for
is a workaround
for a policy
nobody wants to change.

That part is yours.

The model optimizes
the answer.

You are paid
to optimize
the question.

Choosing the problem
is the last job
you should ever hand over,

because everything downstream
inherits it.

A fast, elegant, tested solution
to the wrong problem
costs more than nothing,

because now
you maintain it forever.

Nobody ever apologized
for a wrong answer.

They apologized
for the six weeks
spent building it.

So do the slow part first.

Write the problem
in one sentence,
in plain words,
without naming a technology.

Say who hurts.
Say how you would know
it stopped hurting.
Say what you will not do.

Then read it back
to a person
who has to live
with the result.

If they wince,
you have found the real problem,

and it was never
the one on the ticket.

Then bring the machine in
and let it be brilliant
inside your fence.

Use it to draft.
Use it to argue.
Use it to find
the thing you forgot.

Just don't let it
choose the destination
because it sounds confident
about the road.

It has read everything
ever written
about the map.

It has never once
been to your town.

Keep the wheel.

Let it read the map
out loud.

– Serguey Asael Shinder