The refrigerator door is wide open, cold air dumping straight onto the kitchen floor, and my kid is staring blankly at the middle shelf.
“Dad, do we have any milk?”
I don’t even have to look up from the counter. “Move the orange juice carton two inches to the left.”
There it is. A full gallon of whole milk, sitting right where it’s lived for six years. He didn’t look. He opened the door, gave it half a second of peripheral vision, and outsourced the work of basic physical observation to the nearest available adult.
That kitchen reflex is exactly how half the industry is using LLMs right now.
Burning megawatts to find the orange juice
Somewhere along the line, people convinced themselves that typing a twenty-word prompt into a chat window to solve an eight-second question is “working smart.” It isn’t. It’s heavy. It requires context switching, token round-trips, network latency, and a server farm boiling water in Iowa just so you don’t have to scan a config file with your own eyes.
I’ve worked in infrastructure for over twenty years. If there’s one law of systems engineering that never changes, it’s this: complexity is a killer of productivity.
Every time we reach for an LLM before checking grep, reading the error trace, or just moving the mental ketchup bottle, we aren’t creating leverage. We’re creating friction disguised as magic. When you fire up an agent cluster to do a task that takes four seconds of raw human focus, you aren’t an architect. You’re the kid standing in front of the open fridge.
Conductors vs. passive consumers
When I built a video game with my two kids last summer, the rule was simple: AI was the canvas, but they had to hold the brushes. They drew the sprites, decided the game physics, and tested the hitboxes. They were conductors of intelligence, not passive consumers waiting for the machine to spit out a finished product.
The danger of AI isn’t that models will take over our systems. The danger is that we’ll let our observational muscles atrophy until we can’t find anything without a prompt.
If you don’t practice looking, you lose the ability to see. When a server goes down at 2 AM or an integration drops payloads on the floor, the model can’t smell the smoke for you. You need engineers who know what the shelf is supposed to look like before the door opened.
Honest caveats
Honest caveats, because writing about AI discipline without them is just preaching: I use LLMs every single day. They generate boilerplate, scaffold schemas, and speed up discovery across complex stacks. When I’m working with a new API schema or translating legacy spaghetti code, having an LLM at my side is pure rocket fuel.
The distinction isn’t anti-AI vs. pro-AI. The distinction is knowing when you’re using a tool to build fence, and when you’re just standing there asking someone else to hold the hammer.
Close the fridge door. Move the juice. Look for yourself first.
