We Deployed the Most Powerful Technology in Human History Before Anyone Knew How to Use It — Including the People Who Built It
If you are currently responsible for “implementing AI” at your company, there is a good chance you feel like you’re failing.
You’ve tried different prompts.
You’ve built workflows.
You’ve experimented with agents, chains, automations, knowledge bases, tools that promise to orchestrate tools that orchestrate other tools.
Sometimes it works.
Sometimes it works beautifully.
And then, without warning, it doesn’t.
The same workflow that produced a brilliant result yesterday suddenly returns something that feels… wrong. Shallow. Incomplete. Slightly off in a way that’s hard to explain.
So you tweak the prompt.
Then you tweak it again.
Maybe you add more structure. More instructions. More steps.
Sometimes that helps.
Sometimes it makes things worse.
At some point you probably start wondering a quiet, slightly uncomfortable question:
Am I just bad at this?
The good news is that you’re not.
The slightly more unsettling news is that no one else is good at it either.
Not the consultants selling AI transformation roadmaps.
Not the executives demanding AI strategy updates.
Not the venture capitalists predicting trillion-dollar productivity gains.
And, if we’re being completely honest, not even the people building the models themselves.
Right now, across the entire technology industry, something very strange is happening.
We are collectively trying to operationalize a technology that we do not yet fully understand.
And we’re doing it at scale.
The Situation, Explained From Orbit
Let’s zoom out for a moment.
Humanity has recently built a new kind of machine.
These machines are capable of producing language, analysis, and reasoning patterns that often appear strikingly intelligent.
They can summarize documents.
Write software.
Design experiments.
Analyze contracts.
Explain quantum mechanics.
Draft marketing campaigns.
And occasionally, they can also confidently tell you that Abraham Lincoln invented Wi-Fi.
The strange part is not that they make mistakes.
The strange part is that we don’t fully understand why they work when they work.
Modern language models are not traditional software systems.
They are large statistical structures trained to predict the next token in enormous sequences of human text.
From that simple training objective, something remarkable emerges:
Reasoning-like behavior.
Planning.
Problem decomposition.
Pattern recognition.
Even something that looks suspiciously like creativity.
But the internal mechanisms that produce those behaviors are still being actively studied.
Which means the current state of the art in AI looks something like this:
We have built a system that can reason.
We know how to make it bigger.
We know how to train it.
We know how to use it.
But we do not yet have a complete theory of how its reasoning actually works.
And this is where the situation becomes slightly ridiculous.
Because the moment these systems became powerful enough to be useful…
…the world immediately deployed them everywhere.
The Enterprise Deployment Phase
If you want to see where the absurdity begins, look at what happened next.
Companies everywhere began asking the same question:
“How do we integrate AI into our workflows?”
Which is a perfectly reasonable question.
Except for one small complication.
In most fields of engineering, the timeline looks something like this:
First, scientists discover a phenomenon.
Then engineers figure out how it behaves.
Then systems engineers develop reliability theory.
Then operators build procedures.
Then industries deploy the technology widely.
With AI, the timeline looks more like this:
Discovery → global deployment → figure everything else out later.
Which means thousands of operators around the world are currently being asked to do something like this:
“Please design reliable operational procedures for a technology whose governing theory is still under construction.”
This is the technological equivalent of being asked to write aviation safety manuals before anyone understands why the airplane managed to get off the ground.
The plane clearly flies.
But the aerodynamics textbook is still being written.
The Middle Management Event Horizon
Now imagine you’re the person sitting in the middle of this.
Your company has decided AI is important.
Your leadership has announced an AI initiative.
Your team has been told to “figure out how to use it.”
You are now responsible for operationalizing systems whose internal reasoning processes are largely opaque.
There is no established discipline for doing this.
No widely accepted theory.
No standard reliability model.
But there are expectations.
Deadlines.
Roadmaps.
Productivity targets.
And an entire industry telling you this should be easy.
Which leads to the current operational strategy most organizations are using:
Step 1: deploy AI models
Step 2: write prompts
Step 3: hope
When this works, it looks miraculous.
When it doesn’t, it looks like chaos.
Which explains why so many operators currently feel like they’re doing something wrong.
They assume everyone else must have figured this out.
But the truth is much simpler.
Everyone else is improvising too.
The Universal AI Idiot Condition
There is a quiet secret circulating through the AI ecosystem right now.
No one says it out loud.
But everyone who works closely with these systems eventually notices it.
At some level, we are all AI idiots right now.
The junior analyst experimenting with prompts.
The middle manager building automation workflows.
The startup founder pitching AI-powered products.
The researchers publishing interpretability papers.
The executives announcing AI strategy.
And yes — even the people running the companies building the models.
This is not an insult.
It is simply the natural state of humanity encountering a new class of system before the surrounding discipline exists.
We are doing what humans always do when we discover powerful technology.
We experiment.
We improvise.
We build procedures that sometimes work.
And we slowly develop the theories needed to make those procedures reliable.
Except this time, we skipped the “slowly” part.
The Compressed Revolution
Historically, new technological disciplines emerge over decades.
Electric power systems required half a century of experimentation before safety and reliability engineering matured.
Aviation took decades to develop modern flight control systems.
Software engineering spent thirty years evolving practices that make large systems manageable.
AI skipped that entire timeline.
The transformer architecture was introduced in 2017.
Large language models began transforming industry around 2023.
In historical terms, we are currently somewhere around 1905 in aviation.
Except instead of a few experimental aircraft flying over sand dunes…
…the entire global economy has decided to board the plane.
The Cosmic Perspective
At this point it helps to step back.
Very far back.
Far enough that the whole situation becomes visible.
Humanity has built a probabilistic reasoning engine.
We do not fully understand how its reasoning emerges.
We have deployed it into legal systems, hospitals, financial institutions, supply chains, research labs, and customer service pipelines.
And we are attempting to control its behavior primarily by asking it politely to follow instructions written in natural language.
If you zoom out far enough, this begins to look less like a well-planned technological rollout and more like something from a very dry science fiction novel.
Which is where a certain piece of advice becomes relevant.
Don’t panic.
Not because everything is under control.
But because this level of chaos is exactly what technological revolutions look like while the theory is still forming.
The Part That Actually Matters
Once you see the situation clearly, two important realizations follow.
First:
If AI workflows feel unreliable, it’s not because you’re incompetent.
It’s because you are trying to operate systems whose control model has not been fully invented yet.
Second:
The people building these systems are discovering that control model at the same time you are.
Which means the awkward improvisation phase is not a failure.
It’s the early stage of a new engineering discipline forming in real time.
Eventually there will be textbooks.
Standard architectures.
Reliability theory.
Formal procedures.
Right now, there are mostly experiments.
Grab Your Towel
So if you are currently responsible for implementing AI somewhere in the organizational middle layer of modern industry, here is the most honest status update anyone can give you:
Yes, this situation is difficult.
Yes, the expectations are unrealistic.
Yes, the technology behaves in ways that often feel unpredictable.
And yes, everyone, from the newest prompt engineering interns to the CEOs of the largest AI companies, is still figuring this out.
Which means the correct response is not panic.
(Though a brief session of breathing into a paper bag is understandable.)
Instead, it’s curiosity, experimentation, and a mountain of something we’ve largely ignored until now:
caution.
In the meantime, occasionally stepping far enough back to appreciate the absurdity of what we are all collectively attempting can at least make the circumstances marginally load-bearing.
Humanity has always invented the theory after discovering the phenomenon.
We just usually give ourselves a few decades before deploying it everywhere.
This time we gave ourselves about six months.
Which, in retrospect, may have been slightly ambitious.
Still.
The plane did get off the ground.
And all of humanity boarded it.
Now we just have to figure out how to fly it.
And in the meantime, it helps to keep a towel handy.