AI Vacuum Cleaners
Are we still debating the same things?
24 months into the AI revolution and we are still having the same debates.
“I’m 100x more productive. I’m generating 50 PRs per day. Look at me go Daddy.”
“LLMs can’t reason. The code is buggy and littered with abstraction leaks. You’re confusing movement with velocity.”
“Skill issue. You’re holding it wrong.”
Sometimes it’s useful to compare agentic development to a robot vacuum cleaner.
Before I had one, I’d vacuum the house at most once a week. A daily vacuum would have been great but it costs too much in time and effort — I had better things to do.
And after I had one, I realised just how poorly they did their job. They’d get trapped under the table, miss dirt and dust, and remind me just how much better a human can do the job.
But here’s the rub … in some rooms, my Roomba was good enough. Big open spaces and flat surfaces make automated cleaning an attractive option. If I don’t have allergies or OCD, then the job it does is fine. But in other rooms, it was worse than useless. It would cost me time having to unblock and reset it every time it got stuck. And if I needed the job to be done well, I’d still have to do it myself.
Domain complexity and correctness standards can explain almost all of the AI debate.