A few AI-related thoughts I had lately and can't be bothered to break into pieces “for the algorithm”. Apologies if there's no thread connecting them.
- Ctrl+C is now part of my workflow. ⌨️ When I see the model drifting, I stop it. The cheapest fix is the one you catch early. Watching the AI work and making a judgment call in real time is part of the job now.
- There's a real gap between vibe coding a side project for yourself and building software for hundreds of thousands of people. Small audience, low stakes? Vibe away. The feedback loop is fast and the blast radius is small. Complex enterprise product? Don't let AI run amok. The clients don't care how good your harness is, or how good your prompt was. 🤖
- Multi-agent pipelines running in the background sound powerful. But with costs going up, letting AI run unsupervised for 40 steps before reviewing it gets costly fast. The best ROI on AI spend today is probably a human watching in real time. Local models might change this eventually. For now, active supervision is both the smart financial move and a big part of the actual work. 💰
- My favorite thing I don't hear enough of: you can use AI to build things that don't use AI. Everyone assumes that if you used AI to build something, the product itself needs to contain AI. But just because a crane helped build it, doesn't mean it has to be a crane rental. 🏗️ You can ship a perfectly deterministic script, pipeline, or application with zero model calls at runtime. That is often the better choice.
That's it folks. Remember: keep the Ctrl+C handy. And feel free to disagree in the comments!
More notes on engineering, AI and the systems around the work.
Back to all writing