Skip to main content
SubZeroDev

Professional uncertainty since 2026.

Well… Why not?


Discount Store Marcus Aurelius on AI

· 5 min read

I own books by Marcus Aurelius, Epictetus, and assorted other dead people who apparently figured this shit out a while ago.

I haven't fucking read them yet.

So if you came here for a scholarly treatment of Stoicism, congratulations. You have entered the wrong shop.

This is the clearance aisle.

Today we're discussing AI.

“Do not spend today's tokens on tomorrow's hypothetical catastrophe.”

— Lucius Contextus Maximus

People keep asking whether AI will take my job.

Maybe.

What do you want me to do with that possibility this afternoon?

Learn something? Try the tools? Notice which parts of my work are becoming easier? Excellent. Those are activities.

Sit here experiencing an imaginary redundancy meeting until reality gets around to scheduling one?

That sounds like a fucking terrible use of a Wednesday.

Reality Gets a Vote​

I'm not claiming nothing bad can happen. Jobs can disappear. Transitions can hurt. A useful tool does not automatically produce a kind employer or a sensible economy.

But preparation and advance suffering are different activities.

Preparation changes what I can do. Advance suffering mostly changes whether I enjoy lunch.

I can pay attention. Build skills. Keep options open. Try things while the cost of trying them is manageable. If the situation changes, respond to the situation that actually exists.

I don't need to emotionally inhabit every possible future before selecting one.

The fucking multiverse can manage its own staffing problems.

Please Do Not Taxidermy My Opinion​

This is what I think now.

Based on the tools I've used, the work I've done, and the evidence available to me.

If any of those change, my opinion is allowed to change with them.

Apparently this needs saying because people treat an old opinion like a signed blood covenant. You said something six months ago. Now you have learned something. Explain yourself before the Tribunal of Screenshots.

Yes. I updated the model in my head.

That was the fucking point of acquiring new information.

My enthusiasm for AI isn't a promise to defend every product, every output, or every idiot with a slide deck. A tool can become worse. A workflow can stop paying for itself. A capability I dismissed can become useful.

I want to notice.

Being right once is pleasant. Remaining able to become less wrong is considerably more useful.

The Tools Are Not Family​

“The architecture was correct when the constraints were different.”

— Decimus Refactorus

I can like a language without making it my ancestral homeland.

I can use a model without marrying it.

I can build an architecture, be proud of it, and later throw part of it away because the assumptions that justified it have changed.

The same applies to my job title. I don't need to preserve every task underneath it as a protected historical monument.

If something used to require three days and now takes an hour, the missing two days and seven hours are not sacred remains. We don't need a procession.

We can do something else.

Even my preferred workflow is provisional. The point is to produce useful, reliable results under the constraints I actually have. It isn't to demonstrate lifelong loyalty to the workaround I invented last February.

The Oracle Has to Run the Tests​

“Confidence is not an exit code.”

— Tiberius Validatus

AI is a tool. It is not an authority.

I can find it useful without treating every sentence it produces as a tiny burning bush.

If it claims the code works, run it. If it claims a source says something, read the source. If it changes behavior, check that behavior against the requirement.

And check the requirement too. A model can write a test that enthusiastically agrees with the same mistake it made in the implementation.

Two wrong things nodding at each other is not verification. It's a committee.

I want a process where confidence matters less because evidence is available: observable results, explicit contracts, meaningful tests, review, and a way to recover when something fails.

That doesn't remove judgment. It gives judgment something better to work with than the machine's excellent bedside manner.

Even the Clever Model Has a Bill​

“Send not the entire empire to investigate a missing semicolon.”

— Gaius Bottleneckus

More context isn't automatically more useful context. A long chat can contain the decision, three abandoned decisions, and an extremely articulate explanation of why the second abandoned decision was definitely the final one.

Wonderful archive. Questionable briefing.

I want durable project state outside the conversation, and enough relevant context inside it to do the current job. I want capability appropriate to the task. Sometimes that means a stronger model. Sometimes it means a smaller request. Sometimes it means ordinary code should do the mechanical work and the model should stop narrating.

The economics can change. Prices, speed, quality, context handling, the cost of checking the result. When they change, I can reconsider how I divide the work.

No funeral required for the previous arrangement.

What Becomes Possible​

I’m much more interested in what becomes possible than in protecting what used to be necessary.

That doesn't tell me everything AI will do. It tells me where to put my attention.

Can I attempt something I would previously have postponed? Explore another design? Make a useful little thing for an audience too small to interest anybody with a quarterly earnings call?

Then let's find out.

If it fails, I'll have learned something specific. If it works, I'll have something that didn't exist before.

Either seems more useful than defending the continued existence of a chore.

I may eventually read the books.

Until then, reality retains veto power.

Discount Store Marcus Aurelius out.