Reliable engineeringrequires correct decisions.

The whole point of software is to get things right repeatedly.

You cannot outsource decisions to an LLM and assume it will choose as you would.

Productivity

We decided while we did the work.

Reading the code meant checking its behaviour. The design decisions were part of the work we called coding.

AI writes
the code.

The code is faster. You still have to make the same decisions.

AI chooses
the result.

The larger time saving comes when you let the model make those decisions too.

We decided while we did the work.

Diamonds are decisions. A small amount of human attention remains after delegation.

We should be careful when handing decisions to an LLM

The implementation depends on which data must be erased and which records the business must retain.

What happens if you get the decision wrong?

What should delete do?
deleteUser(id)
User account ActiveExisting data Past orders
Remove the recordHard delete
Required records lost

Can destroy records the business is legally required to retain.

Keep it, mark inactiveSoft delete
Personal data retained

Can leave personal data stored when GDPR requires erasure.

AI productivity is oversold.

The AI industry's valuation includes productivity gains from removing decisions from humans. Where humans maintain decision control, the productivity gains do not materialize.

AI cannot make decisions for you the way you want them made. The best we can build is AI systems that never require a decision to be repeated.

What do you do with this?

One level up the abstraction hierarchy is where the gap exists to create value in the attention economy.

Decide what decisions you can afford to not make.