Histology: Preparing Fish Health Data for AI

Computer scientists have been arguing about versions of this problem for decades.

One of the first things you learn when studying computer science is how to sort a list.

Imagine I hand you a shuffled deck of numbered cards and ask you to put them in order.

There are a lot of ways to do it.

You could start with one card, take the next card and put it in the right position, and take the third card and insert that in the right position.

Keep repeating that process until you've worked through the entire deck.

That's basically insertion sort.

Or you could break the deck into smaller groups, sort those groups separately, and then efficiently combine them back together.

That's the basic idea behind merge sort.

There are bubble sorts, quick sorts, heap sorts, an entire family of algorithms built around what sounds like an incredibly simple problem.

Put these things in order.

The reason computer science students learn this isn't because the world needs more people who know how to sort playing cards.

It's because it teaches you something fundamental about computation.

There can be many ways to solve exactly the same problem, but those approaches don't necessarily perform the same way.

With 10 numbers, it might not matter.

With 10 million numbers, it matters a lot.

And with 10 billion, a bad approach can become practically unusable.

That lesson feels relevant again today because we're entering a world where computing power can hide a lot of bad habits.

Give ChatGPT, Claude, Copilot, or any other AI system a messy problem, and increasingly, it will find some way through it.

It can search through information, rewrite things, generate code, transform files, and keep trying different approaches until something works.

But there is a cost to that, and I've mentioned it multiple times on this podcast.

Give a model messy data, inconsistent terminology, unclear instructions, and you're asking it to spend computational effort figuring out what you meant before you can even start solving the actual problem.

give it clean data, consistent structure, and clear definitions, and suddenly, that same problem becomes much easier.

It may sound obvious, but I think it is getting lost in some of the excitement around what's possible with AI.

We've spent the last few years talking about bigger models, more compute, larger context windows, agents, reasoning models, and a ton of what these systems will be eventually able to do.