2/ 3. Automate the trivial; use AI to assist judgment, not replace it.

4. Break complex problems into smaller, well-defined tasks.

5. Where judgment is involved, expect to iterate.

Curious how this lines up with others’ experience.

What would you add, change, or argue against?

1/ I’ve been trying to get clearer on when AI is useful—and when it quietly leads you astray.

As a working model (very much for now):

1. Treat outputs as unverified candidates, not truths.
2. Apply AI to bounded, well-structured problems.

SMART goals assume you know what you want.

Specificity can help—but it can also paint you into a corner.

So I try WILD goals:

W — Worth a try (plausible)
I — Incomplete (under-specified)
L — Low cost (easy to try)
D — Discovering (for learning, not perfection)

Start WILD, then go SMART.

Second Phoenix and Pepper post! Added a cost model, and discovered our flashiest ship was our least profitable. It was sailing 0.006% full. 200 denarii a day to move a package that would fit in a satchel.

The fix: Fill the rest of the hold with pepper!

www.thedecisionblog.com/revenue-isnt...

I started a fake Roman trading company to learn data engineering. Basilisk venom, suspicious wizards, and a lot of garum. My first blog post about it is up! www.thedecisionblog.com/a%20fictiona...

Most evaluation systems assume their metrics work.

They don’t actually know.

Good inputs → good outputs → metric gets the credit.

Without a counterfactual, that tells you almost nothing.

Better systems use multiple independent signals and check what actually predicts outcomes over time.

One pattern I keep seeing: strategy gets replaced by broadly agreeable aspirations—things almost everyone would endorse.

If it doesn’t reduce the number of viable paths forward, it’s probably not functioning as a strategy.

Alone reading Pliny’s Natural History :)

That is an awesome tattoo! I just can't even

holy wow i didn’t know it looked so awesome!