April 23, 2024 — I wrapped my fingers around the white ceramic mug in the cold air. I felt the warmth on my hands. The caramel colored surface released snakes of steam. I brought the cup to my lips and took a slow sip of the coffee bean flavored water inside.

Happiness is a hot cup of coffee in a ceramic mug on a cold day.

But, like happiness, heat is fleeting.

How can I keep the coffee hot?

I am using a free ceramic mug I got at a work conference.

That's the issue.

I pour my coffee into a ceramic mug I bought at HomeGoods.

Expensive mugs hold heat better.

It gets cold anyway.

My hands can feel the heat rushing from the ceramic. It must be the material.

That's the issue.

I buy an insulated metal mug.

A better material than ceramic is the most important thing for holding heat.

It gets cold anyway.

I combine my theories. I buy an expensive insulated metal mug.

It gets cold anyway.

Wrong again.

Let me try a new approach.

I go back to HomeGoods and buy a thermometer and stopwatch.

I line up the 3 coffee mugs. Pour hot coffee in each.

Every five minutes, I write down the temperature in each mug.

I plot the data in a line chart.

There are some differences, but they are minor.

"What are you doing?" my cousin asks.

I explain the situation.

"Try my mugs, they have covers", he says.

I repeat the experiment and add the new data to my dataset.

The coffee in my cousin's covered metal mugs stays hot.

Materials and covers are the key.

I still can't explain why.

What is heat?

I know I can build a much better model.

But I have successfully discarded false models.

And now I have a dataset to test all future models against.

• Datasets are automated tests for world models.
• Building datasets can be done first, orthogonal to models. What Test Driven Development is to programming, Dataset Driven Development is to science.
• You can only hold so many concepts in your head at once, so automated testing from datasets helps more the more complex the system.
• Beware stories lacking datasets!
• Also beware stories that start with true datasets but then discreetly (but swiftly) switch to false ones.
• Love data sharers, beware data misers.
• An ounce of good data is worth a pound of bad models.
• An additional concept, measure, or measurement are all capable of triggering a huge leap toward a more accurate model.
• Datasets are a useful, simple, and timeless tool to discover better models.
• Datasets are particularly good at automating the identification of bad models.
• Tabular UIs have superior ergonomics for humans because they allow more efficient eye movement and require fewer tokens to manage.
• All that being said, remember that datasets are shadows, ultimately the models, not the datasets, are the important thing!

---

I took some narrative liberties above^[1], but for accuracy and fun^[2], I really did this experiment.

Index	Name	Diameter	Height	Covered	5	10	15	20	25	30	35	40	45	50	55
1	Thermos	3	7.25	TRUE	149.9	147	144.6	141.2	138.2	137.1	134	132.8	131.1	129.3	127.2
2	Yeti Covered	4	3.8	TRUE	146.6	141	136.7	132.5	129	125.6	123.4	121.6	118.5	117.6	115.7
3	Yeti	4	3.8	FALSE	135.5	118.9	107.9	100.5	94.6	90.8	87.6	85.4	82.9	81.3	79.7
4	SF	4	4.5	FALSE	128.8	113.9	102.5	95.9	90.5	86.3	83.6	81.6	79.8	78.2	76.6
5	Aruba	3	3.75	FALSE	130.8	114.6	104.9	98.9	92.3	87.6	85.4	82.9	81.1	80.2	78

---