Old Man's War

In John Scalzi’s Old Man’s War, Earth recruits elderly people to fight wars in space.

The soldiers leave their worn-out bodies behind. Their consciousness gets transferred into new bodies built for combat. They are stronger, faster, and very hard to kill.

Then there are the Ghost Brigades. They are born directly into those bodies. They never learned to move through the world as fragile humans.

That difference changes how they fight.

One of them explains that he learned to fight knowing he could sacrifice an arm and grow it back. Everyone who was born as a human fights while trying to protect both arms. They aren’t used to a world where you can grow a new one.

Today, I realized most companies are fighting the AI war while protecting both arms.

They “fight” like intelligence is finite.

They protect things that have become cheap. They spend hours designing processes to conserve software, analysis, code, and machine attention. They are bringing instincts built in the world of atoms into a world of bits.

Take Slack channels.

The normal company assumes channels are expensive to create and painful to manage. This produces naming conventions, approval processes, channel audits, and a long argument about who owns the taxonomy.

An agent can create (and monitor) 150 channels in three minutes.

Let the agent make a channel for every ticket. Watch which ones people use. Shut down the dead ones (an agent can do this in 30 seconds).

Make 500 more and run another test.

That sounds insane if you think a Slack channel is a room in a building. Rooms cost money. They take time to construct. Somebody has to clean them.

A Slack channel is bits.

The same mistake shows up in software engineering - surprisingly.

Every bone in a good engineer’s body says to spend three hours setting up the right webhook. The webhook will be clean. It will fire at the right moment. It will fail gracefully.

Or you can run a cheap open source AI model every 15 minutes and ask whether anything needs attention.

Maybe the loop fails 10% of the time. For plenty of internal use cases, that is fine. The cost of running it is close to zero. The software engineer’s three hours are expensive.

I have an even better example for an agent we just built.

Say you have 100 documents and need to find the important ones. The old approach builds a keyword system, extracts every page with OCR, defines a search index, and maintains the pipeline.

The new approach sends all 100 documents through a cheap model with one question:

Is this interesting?

Yes or no. If the model is unsure, call it a yes.

Then send the ten interesting documents to a smarter model.

You do not need the cheap model to be brilliant. You need it to reduce the pile by 90% without dropping anything dangerous. The expensive intelligence only touches the small pile that deserves it.

This way of working feels wasteful because we spent our entire careers conserving the wrong things.

Software used to be expensive. Code took weeks. Analysis required a person. Customized workflows belonged to companies with large IT budgets. Every new system became another permanent object somebody had to protect.

Now you can create a machine that creates other machines.

The Ghost Brigades would never spend three hours protecting an arm they can regrow. We should stop spending three hours protecting a workflow an agent can rebuild before lunch.

That does not mean cost and failure disappear!

You still need a safe boundary. A payroll system cannot be wrong 10% of the time. A customer-facing agent cannot invent refund policies. Cheap intelligence aimed at the wrong target can produce an enormous amount of hot garbage.

The scarce work moved up a level.

You have to choose the target. Define what a tolerable failure looks like. Decide which outputs can stay cheap and which 10% need better intelligence or a human. Watch what happens in the real world. Kill the loops nobody uses.

Four questions help:

1. Which assumption in this process came from a time when intelligence or software was expensive?

2. What would we try if machine attention were free?

3. If I had a 1000 smart interns to point at this, what could I do?

4. Where can a cheap, imperfect loop replace a carefully engineered first version?

5. Which failures would hurt enough that a human or stronger model must catch them?

Build inside those boundaries and make more attempts.

The companies that learn to fight with regrowable arms will look reckless for a while.

Then everyone else will wonder how they ship six large agents in a week (we just did).

Yallah Habibi,

Jon