Dr Strangelove or: How I learned to stop worrying and love AI

2026-03-02

The War Room scene from Dr. Strangelove — **Figure 1.** The War Room in *Dr. Strangelove*, a visual metaphor for systemic risk and escalation. Source: Wikimedia Commons (public domain).

When I joined a startup in 2021, I expected to find lots of machine learning. It was the flavour of the month. Every company was “AI-powered.” I assumed I would walk into a codebase dense with models, training pipelines, feature engineering, and statistical nuance.

Instead, I found something much simpler. There were small learned models in specific places. But most of the system was composed of rules, transformations, heuristics, and well-structured pipelines. It wasn’t artificial general intelligence hiding in production.

It was mostly engineering. That surprised me.

Not because it was underwhelming — but because it was clarifying. The real leverage wasn’t in exotic models. It was in how components were composed, how data flowed, and how clearly responsibilities were defined.

That was my first shift in perspective.

The first stage: acceleration

A few years later, the next wave arrived: AI-assisted coding. Tools like Copilot began generating entire functions, tests, even small modules. The effect was immediate. This is the first stage of AI adoption:

Use it to code faster.

And it works. Boilerplate disappears. Endpoints materialise quickly. Tickets close sooner. It feels like velocity has doubled.

But acceleration has a shape. When AI is used primarily as an autocomplete engine, it optimises for local correctness. It solves the immediate prompt. It does not optimise for global coherence across a system. Files grow. Patterns repeat. Similar abstractions diverge slightly. The system functions — but its internal logic becomes harder to see.

The cost isn’t writing code. The cost is evolving it.

The second stage: friction

After a few cycles, a different pattern emerges. Time to write a feature decreases. Time to integrate a feature increases. Builds grow slower. More files are touched per change. Code reviews take longer because reviewers must reconstruct intent from generated structure. Seemingly local changes trigger failures in distant parts of the system. The surface area expands faster than understanding.

This is where the real cost appears.

AI-generated code often optimises for solving the immediate prompt. It does not account for long-term architectural coherence. It does not consolidate abstractions unless asked. It does not remove duplication unless instructed. It does not feel the cumulative weight of a growing system. So complexity accumulates. Not in dramatic failures — but in friction:

Longer CI times.
Broader dependency graphs.
Brittler tests.
Bugs that are not confined to the feature location, but systemic.
Reviews that focus on deciphering structure rather than evaluating behaviour.

The third stage: amplification

The real shift happens when AI stops being a typing assistant and becomes a thinking tool. Instead of asking:

“Write a controller that does X.”

You start asking:

“What architectural options isolate this responsibility?”

“Where might coupling emerge in this design?”

“If this grows 10x, what breaks first?”

This is the second stage of adoption — not acceleration, but amplification. AI becomes a way to explore design space. To simulate trade-offs. To pressure-test structure before committing to it. It does not replace engineering judgment. It sharpens it.

What changed for me

I stopped worrying about AI replacing engineers. Because typing speed was never the constraint. The constraint was always:

clarity
structure
evolvability
cost of change

AI makes code cheaper to produce. That makes architecture more valuable, not less. When generation becomes abundant, coherence becomes scarce. And coherence is a human responsibility.

The strangelove moment

In Dr Strangelove, the danger wasn’t the machine itself. It was the system built around it.

AI is similar. Used naively, it increases entropy faster than we can contain it. Used deliberately, it compresses feedback loops and expands the space of ideas we can explore. I no longer worry about AI.

Dr. Strangelove close-up (Peter Sellers) — **Figure 2.** Dr. Strangelove (Peter Sellers): “It’s simple to understand. And completely credible and convincing.” Source: Wikimedia Commons (public domain). Quote reference: Wikiquote.

I pay attention to how we choose to use it. Because the future of engineering will not be determined by how fast we can generate code. It will be determined by how intentionally we shape systems.